Upload folder using huggingface_hub

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/config.full.yaml

@@ -3,7 +3,6 @@ framework:
   qwenvl:
     base_vlm: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
     attn_implementation: flash_attention_2
-    flex_backend: triton
     enable_gradient_checkpointing: true
   action_model:
     action_model_type: MLP
@@ -132,8 +131,8 @@ trainer:
   gradient_accumulation_steps: 16
   distributed_backend: deepspeed
   is_resume: true
-  pretrained_checkpoint: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/checkpoints/steps_2000_state
-  resume_step: 2000
+  pretrained_checkpoint: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/checkpoints/steps_4000_state
+  resume_step: 4000
   reload_modules: null
   optimizer:
     name: AdamW

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/config.yaml

@@ -42,20 +42,12 @@ rl_games:
   env_eval:
     enabled: true
     eval_backend: latency_bench
-    mid_train:
-      enabled: false
-      interval_steps: 250
     prompt_mode: raw
-  task: demon_attack
 run_id: demon_attack_fix_latency_6_200ep_7k2steps_single_baseline
 run_root_dir: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints
 seed: 42
 trainer:
   distributed_backend: deepspeed
-  eval_action_classification: false
-  eval_action_classification_interval: null
-  eval_interval: 250
-  eval_num_batches: 50
   freeze_llm_layers: []
   freeze_modules: ''
   freeze_tied_embedding: false
@@ -66,7 +58,6 @@ trainer:
     action_model: 0.0001
     base: 2.0e-05
     qwen_vl_interface: 1.0e-05
-  logging_frequency: 1
   lr_scheduler_type: cosine_with_min_lr
   max_train_steps: 4000
   num_warmup_steps: 100
@@ -77,13 +68,6 @@ trainer:
     eps: 1.0e-08
     fused: true
     weight_decay: 1.0e-08
-  per_latency_eval_num_batches: null
-  pretrained_checkpoint: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/checkpoints/steps_2000_state
-  profile_timing:
-    enabled: false
-  resume_step: 2000
-  save_interval: 500
+  pretrained_checkpoint: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/checkpoints/steps_4000_state
   scheduler_specific_kwargs:
     min_lr: 1.0e-06
-wandb_entity: zihanwang-ai-northwestern-university
-wandb_project: starVLA_rl_games

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/hydra/.hydra/config.yaml

@@ -2,7 +2,6 @@ framework:
   qwenvl:
     base_vlm: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
     attn_implementation: flash_attention_2
-    flex_backend: triton
     enable_gradient_checkpointing: true
   action_model:
     state_dim: 7
@@ -130,8 +129,8 @@ trainer:
   gradient_accumulation_steps: 16
   distributed_backend: deepspeed
   is_resume: true
-  pretrained_checkpoint: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/checkpoints/steps_2000_state
-  resume_step: 2000
+  pretrained_checkpoint: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/checkpoints/steps_4000_state
+  resume_step: 4000
   reload_modules: null
   optimizer:
     name: AdamW

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/hydra/.hydra/hydra.yaml

@@ -118,7 +118,6 @@ hydra:
     - mode=single
     - ++framework.qwenvl.base_vlm=playground/Pretrained_models/Qwen3-VL-4B-Instruct
     - ++framework.qwenvl.attn_implementation=flash_attention_2
-    - ++framework.qwenvl.flex_backend=triton
     - ++framework.qwenvl.enable_gradient_checkpointing=true
     - ++framework.action_model.state_dim=7
     - ++framework.action_model.loss_type=discrete_ce
@@ -318,8 +317,8 @@ hydra:
     - ++version_id=0.21
     - ++run_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints
     - ++trainer.is_resume=true
-    - ++trainer.pretrained_checkpoint=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/checkpoints/steps_2000_state
-    - ++trainer.resume_step=2000
+    - ++trainer.pretrained_checkpoint=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/checkpoints/steps_4000_state
+    - ++trainer.resume_step=4000
     - ++datasets.vla_data.data_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/data/memory/demon_attack_fix_latency_6_200ep_7k2steps
     - ++datasets.vla_data.data_mix=demon_attack_train__bridge
     - ++datasets.vla_data.eval_data_mix=demon_attack_train__bridge__val
@@ -328,9 +327,9 @@ hydra:
   job:
     name: train_starvla_hydra
     chdir: false
-    override_dirname: '++auth.env_file=null,++auth.hf_token_env=HF_TOKEN,++auth.wandb_api_key_env=WANDB_API_KEY,++base_model.repo_id=Qwen/Qwen3-VL-4B-Instruct,++bridge_base_model.local_dir.gr00t=playground/Pretrained_models/Qwen3-VL-4B-Instruct,++bridge_base_model.local_dir.openvla=playground/Pretrained_models/Qwen3-VL-4B-Instruct,++bridge_base_model.local_dir.pi05=playground/Pretrained_models/Qwen3-VL-4B-Instruct,++bridge_base_model.local_dir.pi0=playground/Pretrained_models/Qwen2.5-VL-3B-Instruct-Action,++bridge_base_model.repo_id.gr00t=Qwen/Qwen3-VL-4B-Instruct,++bridge_base_model.repo_id.openvla=Qwen/Qwen3-VL-4B-Instruct,++bridge_base_model.repo_id.pi05=Qwen/Qwen3-VL-4B-Instruct,++bridge_base_model.repo_id.pi0=StarVLA/Qwen2.5-VL-3B-Instruct-Action,++checkpoint.hf_repo_id=null,++checkpoint.load=auto,++checkpoint.local.keep_last_n=1,++checkpoint.save_best_model=false,++checkpoint.save_final_model=true,++checkpoint.save_pt_file=false,++checkpoint.save_safetensors_file=true,++checkpoint.save_training_state=true,++checkpoint.sync.enabled=false,++checkpoint.sync.keep_last_n=0,++checkpoint.sync.repo_id=null,++checkpoint.sync.resume_policy=local_latest,++checkpoint.sync.sync_every_n_checkpoints=1,++config_yaml=null,++dataset.config_name=null,++dataset.converted_name=demon_attack_train,++dataset.debug_subset.enabled=false,++dataset.debug_subset.max_episodes=5,++dataset.debug_subset.suffix=debug,++dataset.episodes_per_latency=null,++dataset.force_download=false,++dataset.latency_filter=null,++dataset.max_episodes=null,++dataset.mixed_converted_name=demon_attack_mixed_latency_train,++dataset.mixed_latency_filter=null,++dataset.mixed_source_hf=,++dataset.setup_force=false,++dataset.single_converted_name=demon_attack_train,++dataset.single_latency_filter=null,++dataset.single_source_hf=,++dataset.skip_verification=false,++dataset.source_hf=,++dataset.source_subdir=null,++dataset.target_latency_unit=raw_frames,++dataset.verify_rows=200,++datasets.vla_data.action_balance.action_key=action_id,++datasets.vla_data.action_balance.enabled=false,++datasets.vla_data.action_balance.flap_id=1,++datasets.vla_data.action_balance.noop_id=0,++datasets.vla_data.action_balance.strategy=balanced_epoch,++datasets.vla_data.action_balance.target_flap_fraction=0.3,++datasets.vla_data.action_type=discrete,++datasets.vla_data.custom_mixtures_path=null,++datasets.vla_data.data_mix=demon_attack_train,++datasets.vla_data.data_mix=demon_attack_train__bridge,++datasets.vla_data.data_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/data/memory/demon_attack_fix_latency_6_200ep_7k2steps,++datasets.vla_data.data_root_dir=playground/Datasets/rl_games,++datasets.vla_data.dataset_py=lerobot_datasets,++datasets.vla_data.eval_data_mix=demon_attack_train__bridge__val,++datasets.vla_data.eval_data_mix=null,++datasets.vla_data.eval_num_workers=8,++datasets.vla_data.eval_sequential_step_sampling=null,++datasets.vla_data.image_mode=single,++datasets.vla_data.include_state=true,++datasets.vla_data.latency_curriculum.computed_plan=null,++datasets.vla_data.latency_curriculum.enabled=false,++datasets.vla_data.latency_curriculum.eval_at_phase_end=false,++datasets.vla_data.latency_curriculum.final_equalization=true,++datasets.vla_data.latency_curriculum.latencies=null,++datasets.vla_data.latency_curriculum.new_latency_passes=1.0,++datasets.vla_data.latency_curriculum.phase_distributions=null,++datasets.vla_data.latency_curriculum.phase_steps=null,++datasets.vla_data.latency_curriculum.replay_passes=0.25,++datasets.vla_data.latency_curriculum.save_at_phase_end=false,++datasets.vla_data.latency_curriculum.step_budget_mode=auto,++datasets.vla_data.latency_curriculum.strategy=exclusive,++datasets.vla_data.latency_curriculum.target_total_passes=2.0,++datasets.vla_data.load_all_data_for_training=true,++datasets.vla_data.num_obs_frames=1,++datasets.vla_data.num_workers=8,++datasets.vla_data.obs_image_size=null,++datasets.vla_data.per_device_batch_size=8,++datasets.vla_data.persistent_workers=true,++datasets.vla_data.pin_memory=true,++datasets.vla_data.prefetch_factor=4,++datasets.vla_data.prompt_mode=raw,++datasets.vla_data.sequential_step_sampling=false,++datasets.vla_data.shuffle=true,++datasets.vla_data.stitch_grid=[2,2],++datasets.vla_data.video_backend=torchvision_av,++framework.action_model.action_dim=7,++framework.action_model.action_env_dim=6,++framework.action_model.action_horizon=1,++framework.action_model.future_action_window_size=0,++framework.action_model.loss_type=discrete_ce,++framework.action_model.past_action_window_size=0,++framework.action_model.state_dim=7,++framework.kv_memory.enabled=false,++framework.kv_memory.packed_train=false,++framework.kv_memory.rebased_sink=true,++framework.kv_memory.rollout_len=8,++framework.kv_memory.window=4,++framework.name=QwenOFT,++framework.qwenvl.attn_implementation=flash_attention_2,++framework.qwenvl.base_vlm=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct,++framework.qwenvl.base_vlm=playground/Pretrained_models/Qwen3-VL-4B-Instruct,++framework.qwenvl.enable_gradient_checkpointing=true,++framework.qwenvl.flex_backend=triton,++initialization.checkpoint_filename=checkpoints/steps_5000_pytorch_model.pt,++initialization.checkpoint_hf_repo_id=StarVLA/Qwen3VL-OFT-Bridge-RT-1,++initialization.checkpoint_local_dir=playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1,++is_debug=false,++output_dir=null,++paths.accelerate_config=starVLA/config/deepseeds/deepspeed_zero2.yaml,++paths.base_model_dir=playground/Pretrained_models/Qwen3-VL-4B-Instruct,++paths.dataset_cache_dir=null,++paths.dataset_local_dir=data/memory/demon_attack_fix_latency_6_200ep_7k2steps,++paths.run_root_dir=results/Checkpoints,++rl_games.action_carrier=bridge,++rl_games.env_eval.action_chunk_execution.chunk_size=null,++rl_games.env_eval.action_chunk_execution.enabled=false,++rl_games.env_eval.demon_attack.noop_max=30,++rl_games.env_eval.distributed_mode=rank_sharded,++rl_games.env_eval.enabled=true,++rl_games.env_eval.eval_backend=latency_bench,++rl_games.env_eval.eval_parallel_envs=5,++rl_games.env_eval.fixed_episode_seeds=true,++rl_games.env_eval.frameskip=4,++rl_games.env_eval.ghost_trail.gamma=1.3,++rl_games.env_eval.ghost_trail.ground_fraction=0.22,++rl_games.env_eval.ghost_trail.history_frames=5,++rl_games.env_eval.ghost_trail.min_alpha=35,++rl_games.env_eval.ghost_trail.scroll_px_per_step=4.0,++rl_games.env_eval.image_size=224,++rl_games.env_eval.image_transform=raw_rgb,++rl_games.env_eval.latency.mode=single,++rl_games.env_eval.latency.prompt_map_path=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/data/memory/demon_attack_fix_latency_6_200ep_7k2steps/demon_attack_train__bridge/latency_prompt_map.json,++rl_games.env_eval.latency.prompt_map_path=null,++rl_games.env_eval.latency.values=[0],++rl_games.env_eval.latency_seed_stride=0,++rl_games.env_eval.mid_train.enabled=false,++rl_games.env_eval.mid_train.interval_steps=250,++rl_games.env_eval.mid_train.latencies=[2],++rl_games.env_eval.mid_train.max_steps_per_episode=3600,++rl_games.env_eval.mid_train.num_episodes=20,++rl_games.env_eval.post_train.enabled=false,++rl_games.env_eval.post_train.latencies=[2],++rl_games.env_eval.post_train.max_steps_per_episode=3600,++rl_games.env_eval.post_train.num_episodes=50,++rl_games.env_eval.prompt_mode=raw,++rl_games.env_eval.seed=42,++rl_games.env_eval.task_description=''You
+    override_dirname: '++auth.env_file=null,++auth.hf_token_env=HF_TOKEN,++auth.wandb_api_key_env=WANDB_API_KEY,++base_model.repo_id=Qwen/Qwen3-VL-4B-Instruct,++bridge_base_model.local_dir.gr00t=playground/Pretrained_models/Qwen3-VL-4B-Instruct,++bridge_base_model.local_dir.openvla=playground/Pretrained_models/Qwen3-VL-4B-Instruct,++bridge_base_model.local_dir.pi05=playground/Pretrained_models/Qwen3-VL-4B-Instruct,++bridge_base_model.local_dir.pi0=playground/Pretrained_models/Qwen2.5-VL-3B-Instruct-Action,++bridge_base_model.repo_id.gr00t=Qwen/Qwen3-VL-4B-Instruct,++bridge_base_model.repo_id.openvla=Qwen/Qwen3-VL-4B-Instruct,++bridge_base_model.repo_id.pi05=Qwen/Qwen3-VL-4B-Instruct,++bridge_base_model.repo_id.pi0=StarVLA/Qwen2.5-VL-3B-Instruct-Action,++checkpoint.hf_repo_id=null,++checkpoint.load=auto,++checkpoint.local.keep_last_n=1,++checkpoint.save_best_model=false,++checkpoint.save_final_model=true,++checkpoint.save_pt_file=false,++checkpoint.save_safetensors_file=true,++checkpoint.save_training_state=true,++checkpoint.sync.enabled=false,++checkpoint.sync.keep_last_n=0,++checkpoint.sync.repo_id=null,++checkpoint.sync.resume_policy=local_latest,++checkpoint.sync.sync_every_n_checkpoints=1,++config_yaml=null,++dataset.config_name=null,++dataset.converted_name=demon_attack_train,++dataset.debug_subset.enabled=false,++dataset.debug_subset.max_episodes=5,++dataset.debug_subset.suffix=debug,++dataset.episodes_per_latency=null,++dataset.force_download=false,++dataset.latency_filter=null,++dataset.max_episodes=null,++dataset.mixed_converted_name=demon_attack_mixed_latency_train,++dataset.mixed_latency_filter=null,++dataset.mixed_source_hf=,++dataset.setup_force=false,++dataset.single_converted_name=demon_attack_train,++dataset.single_latency_filter=null,++dataset.single_source_hf=,++dataset.skip_verification=false,++dataset.source_hf=,++dataset.source_subdir=null,++dataset.target_latency_unit=raw_frames,++dataset.verify_rows=200,++datasets.vla_data.action_balance.action_key=action_id,++datasets.vla_data.action_balance.enabled=false,++datasets.vla_data.action_balance.flap_id=1,++datasets.vla_data.action_balance.noop_id=0,++datasets.vla_data.action_balance.strategy=balanced_epoch,++datasets.vla_data.action_balance.target_flap_fraction=0.3,++datasets.vla_data.action_type=discrete,++datasets.vla_data.custom_mixtures_path=null,++datasets.vla_data.data_mix=demon_attack_train,++datasets.vla_data.data_mix=demon_attack_train__bridge,++datasets.vla_data.data_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/data/memory/demon_attack_fix_latency_6_200ep_7k2steps,++datasets.vla_data.data_root_dir=playground/Datasets/rl_games,++datasets.vla_data.dataset_py=lerobot_datasets,++datasets.vla_data.eval_data_mix=demon_attack_train__bridge__val,++datasets.vla_data.eval_data_mix=null,++datasets.vla_data.eval_num_workers=8,++datasets.vla_data.eval_sequential_step_sampling=null,++datasets.vla_data.image_mode=single,++datasets.vla_data.include_state=true,++datasets.vla_data.latency_curriculum.computed_plan=null,++datasets.vla_data.latency_curriculum.enabled=false,++datasets.vla_data.latency_curriculum.eval_at_phase_end=false,++datasets.vla_data.latency_curriculum.final_equalization=true,++datasets.vla_data.latency_curriculum.latencies=null,++datasets.vla_data.latency_curriculum.new_latency_passes=1.0,++datasets.vla_data.latency_curriculum.phase_distributions=null,++datasets.vla_data.latency_curriculum.phase_steps=null,++datasets.vla_data.latency_curriculum.replay_passes=0.25,++datasets.vla_data.latency_curriculum.save_at_phase_end=false,++datasets.vla_data.latency_curriculum.step_budget_mode=auto,++datasets.vla_data.latency_curriculum.strategy=exclusive,++datasets.vla_data.latency_curriculum.target_total_passes=2.0,++datasets.vla_data.load_all_data_for_training=true,++datasets.vla_data.num_obs_frames=1,++datasets.vla_data.num_workers=8,++datasets.vla_data.obs_image_size=null,++datasets.vla_data.per_device_batch_size=8,++datasets.vla_data.persistent_workers=true,++datasets.vla_data.pin_memory=true,++datasets.vla_data.prefetch_factor=4,++datasets.vla_data.prompt_mode=raw,++datasets.vla_data.sequential_step_sampling=false,++datasets.vla_data.shuffle=true,++datasets.vla_data.stitch_grid=[2,2],++datasets.vla_data.video_backend=torchvision_av,++framework.action_model.action_dim=7,++framework.action_model.action_env_dim=6,++framework.action_model.action_horizon=1,++framework.action_model.future_action_window_size=0,++framework.action_model.loss_type=discrete_ce,++framework.action_model.past_action_window_size=0,++framework.action_model.state_dim=7,++framework.kv_memory.enabled=false,++framework.kv_memory.packed_train=false,++framework.kv_memory.rebased_sink=true,++framework.kv_memory.rollout_len=8,++framework.kv_memory.window=4,++framework.name=QwenOFT,++framework.qwenvl.attn_implementation=flash_attention_2,++framework.qwenvl.base_vlm=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct,++framework.qwenvl.base_vlm=playground/Pretrained_models/Qwen3-VL-4B-Instruct,++framework.qwenvl.enable_gradient_checkpointing=true,++initialization.checkpoint_filename=checkpoints/steps_5000_pytorch_model.pt,++initialization.checkpoint_hf_repo_id=StarVLA/Qwen3VL-OFT-Bridge-RT-1,++initialization.checkpoint_local_dir=playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1,++is_debug=false,++output_dir=null,++paths.accelerate_config=starVLA/config/deepseeds/deepspeed_zero2.yaml,++paths.base_model_dir=playground/Pretrained_models/Qwen3-VL-4B-Instruct,++paths.dataset_cache_dir=null,++paths.dataset_local_dir=data/memory/demon_attack_fix_latency_6_200ep_7k2steps,++paths.run_root_dir=results/Checkpoints,++rl_games.action_carrier=bridge,++rl_games.env_eval.action_chunk_execution.chunk_size=null,++rl_games.env_eval.action_chunk_execution.enabled=false,++rl_games.env_eval.demon_attack.noop_max=30,++rl_games.env_eval.distributed_mode=rank_sharded,++rl_games.env_eval.enabled=true,++rl_games.env_eval.eval_backend=latency_bench,++rl_games.env_eval.eval_parallel_envs=5,++rl_games.env_eval.fixed_episode_seeds=true,++rl_games.env_eval.frameskip=4,++rl_games.env_eval.ghost_trail.gamma=1.3,++rl_games.env_eval.ghost_trail.ground_fraction=0.22,++rl_games.env_eval.ghost_trail.history_frames=5,++rl_games.env_eval.ghost_trail.min_alpha=35,++rl_games.env_eval.ghost_trail.scroll_px_per_step=4.0,++rl_games.env_eval.image_size=224,++rl_games.env_eval.image_transform=raw_rgb,++rl_games.env_eval.latency.mode=single,++rl_games.env_eval.latency.prompt_map_path=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/data/memory/demon_attack_fix_latency_6_200ep_7k2steps/demon_attack_train__bridge/latency_prompt_map.json,++rl_games.env_eval.latency.prompt_map_path=null,++rl_games.env_eval.latency.values=[0],++rl_games.env_eval.latency_seed_stride=0,++rl_games.env_eval.mid_train.enabled=false,++rl_games.env_eval.mid_train.interval_steps=250,++rl_games.env_eval.mid_train.latencies=[2],++rl_games.env_eval.mid_train.max_steps_per_episode=3600,++rl_games.env_eval.mid_train.num_episodes=20,++rl_games.env_eval.post_train.enabled=false,++rl_games.env_eval.post_train.latencies=[2],++rl_games.env_eval.post_train.max_steps_per_episode=3600,++rl_games.env_eval.post_train.num_episodes=50,++rl_games.env_eval.prompt_mode=raw,++rl_games.env_eval.seed=42,++rl_games.env_eval.task_description=''You
       are playing Demon Attack from a single game image. Choose exactly one action
-      from: NOOP, FIRE, RIGHT, LEFT, RIGHTFIRE, LEFTFIRE.'',++rl_games.env_eval.task_seed_stride=0,++rl_games.env_eval.vectorized.batch_size=1,++rl_games.env_eval.vectorized.enabled=false,++rl_games.initialization_mode=bridge,++rl_games.model_alias=openvla,++rl_games.task=demon_attack,++run_id=demon_attack_fix_latency_6_200ep_7k2steps_single_baseline,++run_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints,++run_root_dir=results/Checkpoints,++seed=42,++trainer.cc_f1_tolerance=1,++trainer.distributed_backend=deepspeed,++trainer.eval_action_classification=false,++trainer.eval_action_classification_interval=null,++trainer.eval_interval=250,++trainer.eval_num_batches=50,++trainer.freeze_llm_layers=[],++trainer.freeze_modules=,++trainer.freeze_tied_embedding=false,++trainer.freeze_vit=false,++trainer.gradient_accumulation_steps=16,++trainer.gradient_clipping=1.0,++trainer.is_resume=false,++trainer.is_resume=true,++trainer.learning_rate.action_model=0.0001,++trainer.learning_rate.base=2e-05,++trainer.learning_rate.qwen_vl_interface=1e-05,++trainer.logging_frequency=1,++trainer.loss_scale.vla=1.0,++trainer.loss_scale.vlm=0.1,++trainer.lr_scheduler_type=cosine_with_min_lr,++trainer.max_grad_norm=1.0,++trainer.max_train_steps=4000,++trainer.num_warmup_steps=100,++trainer.optimizer.betas=[0.9,0.95],++trainer.optimizer.eps=1e-08,++trainer.optimizer.fused=true,++trainer.optimizer.name=AdamW,++trainer.optimizer.weight_decay=1e-08,++trainer.per_latency_eval_num_batches=null,++trainer.pretrained_checkpoint=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/checkpoints/steps_2000_state,++trainer.pretrained_checkpoint=null,++trainer.profile_timing.enabled=false,++trainer.profile_timing.log_interval=10,++trainer.reload_modules=null,++trainer.resume_step=0,++trainer.resume_step=2000,++trainer.save_format=pt,++trainer.save_interval=500,++trainer.scheduler_specific_kwargs.min_lr=1e-06,++trainer.weight_decay=0.0,++version_id=0.21,++workspace_dir=WORKSPACE_DIR,env=demon_attack,init=bridge,mode=single,model=openvla'
+      from: NOOP, FIRE, RIGHT, LEFT, RIGHTFIRE, LEFTFIRE.'',++rl_games.env_eval.task_seed_stride=0,++rl_games.env_eval.vectorized.batch_size=1,++rl_games.env_eval.vectorized.enabled=false,++rl_games.initialization_mode=bridge,++rl_games.model_alias=openvla,++rl_games.task=demon_attack,++run_id=demon_attack_fix_latency_6_200ep_7k2steps_single_baseline,++run_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints,++run_root_dir=results/Checkpoints,++seed=42,++trainer.cc_f1_tolerance=1,++trainer.distributed_backend=deepspeed,++trainer.eval_action_classification=false,++trainer.eval_action_classification_interval=null,++trainer.eval_interval=250,++trainer.eval_num_batches=50,++trainer.freeze_llm_layers=[],++trainer.freeze_modules=,++trainer.freeze_tied_embedding=false,++trainer.freeze_vit=false,++trainer.gradient_accumulation_steps=16,++trainer.gradient_clipping=1.0,++trainer.is_resume=false,++trainer.is_resume=true,++trainer.learning_rate.action_model=0.0001,++trainer.learning_rate.base=2e-05,++trainer.learning_rate.qwen_vl_interface=1e-05,++trainer.logging_frequency=1,++trainer.loss_scale.vla=1.0,++trainer.loss_scale.vlm=0.1,++trainer.lr_scheduler_type=cosine_with_min_lr,++trainer.max_grad_norm=1.0,++trainer.max_train_steps=4000,++trainer.num_warmup_steps=100,++trainer.optimizer.betas=[0.9,0.95],++trainer.optimizer.eps=1e-08,++trainer.optimizer.fused=true,++trainer.optimizer.name=AdamW,++trainer.optimizer.weight_decay=1e-08,++trainer.per_latency_eval_num_batches=null,++trainer.pretrained_checkpoint=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/checkpoints/steps_4000_state,++trainer.pretrained_checkpoint=null,++trainer.profile_timing.enabled=false,++trainer.profile_timing.log_interval=10,++trainer.reload_modules=null,++trainer.resume_step=0,++trainer.resume_step=4000,++trainer.save_format=pt,++trainer.save_interval=500,++trainer.scheduler_specific_kwargs.min_lr=1e-06,++trainer.weight_decay=0.0,++version_id=0.21,++workspace_dir=WORKSPACE_DIR,env=demon_attack,init=bridge,mode=single,model=openvla'
     id: ???
     num: ???
     config_name: train

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/hydra/.hydra/overrides.yaml

@@ -4,7 +4,6 @@
 - mode=single
 - ++framework.qwenvl.base_vlm=playground/Pretrained_models/Qwen3-VL-4B-Instruct
 - ++framework.qwenvl.attn_implementation=flash_attention_2
-- ++framework.qwenvl.flex_backend=triton
 - ++framework.qwenvl.enable_gradient_checkpointing=true
 - ++framework.action_model.state_dim=7
 - ++framework.action_model.loss_type=discrete_ce
@@ -204,8 +203,8 @@
 - ++version_id=0.21
 - ++run_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints
 - ++trainer.is_resume=true
-- ++trainer.pretrained_checkpoint=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/checkpoints/steps_2000_state
-- ++trainer.resume_step=2000
+- ++trainer.pretrained_checkpoint=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/checkpoints/steps_4000_state
+- ++trainer.resume_step=4000
 - ++datasets.vla_data.data_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/data/memory/demon_attack_fix_latency_6_200ep_7k2steps
 - ++datasets.vla_data.data_mix=demon_attack_train__bridge
 - ++datasets.vla_data.eval_data_mix=demon_attack_train__bridge__val

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/hydra/train_starvla_hydra.log

@@ -4817,3 +4817,108 @@ Cuda failure 'out of memory'
 [2026-07-26 16:16:08,182][starVLA.training.train_starvla][INFO] - ... and that's all, folks!
 [2026-07-26 16:16:12,474][starVLA.training.train_starvla][INFO] - Destroyed distributed process group
 [2026-07-26 16:16:13,067][starVLA.training.train_starvla][INFO] - Destroyed distributed process group
+[2026-07-28 00:02:50,422][starVLA.model.framework.share_tools][INFO] - [*] [apply_config_compat] normalised config from version_id=0.21 to '0.21'
+[2026-07-28 00:02:50,423][starVLA.training.train_starvla][INFO] - VLA Training :: Warming Up
+[2026-07-28 00:02:50,423][starVLA.training.train_starvla][INFO] - VLA Training :: Warming Up
+[2026-07-28 00:02:50,651][huggingface_hub._login][WARNING] - Note: Environment variable`HF_TOKEN` is set and is the current active token independently from the token you've just configured.
+[2026-07-28 00:02:50,651][starVLA.training.rl_games.auth][INFO] - Logged into Hugging Face Hub
+[2026-07-28 00:02:50,651][huggingface_hub._login][WARNING] - Note: Environment variable`HF_TOKEN` is set and is the current active token independently from the token you've just configured.
+[2026-07-28 00:02:50,651][starVLA.training.rl_games.auth][INFO] - Logged into Hugging Face Hub
+[2026-07-28 00:02:51,584][starVLA.training.rl_games.auth][INFO] - Logged into Weights & Biases
+[2026-07-28 00:02:51,585][starVLA.training.rl_games.auth][INFO] - Logged into Weights & Biases
+[2026-07-28 00:02:51,599][starVLA.training.train_starvla][INFO] - ✅ Configuration wrapped for access tracking
+[2026-07-28 00:02:51,601][starVLA.training.train_starvla][INFO] - ✅ Configuration wrapped for access tracking
+[2026-07-28 00:02:54,103][starVLA.training.train_starvla][INFO] - Creating VLA Dataset with Mixture `demon_attack_train__bridge`
+[2026-07-28 00:02:54,114][starVLA.training.train_starvla][INFO] - Creating VLA Dataset with Mixture `demon_attack_train__bridge`
+[2026-07-28 00:02:55,433][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded mixtures from Behavior (data_config): ['BEHAVIOR_challenge']
+[2026-07-28 00:02:55,434][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded mixtures from Behavior (data_config): ['BEHAVIOR_challenge']
+[2026-07-28 00:02:55,434][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded data_config from DOMINO: ['robotwin']
+[2026-07-28 00:02:55,434][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded embodiment_tags from DOMINO (data_config): []
+[2026-07-28 00:02:55,434][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded mixtures from DOMINO (data_config): ['domino', 'domino_clean', 'domino_random', 'domino_cotrain']
+[2026-07-28 00:02:55,434][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded data_config from DOMINO: ['robotwin']
+[2026-07-28 00:02:55,435][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded embodiment_tags from DOMINO (data_config): []
+[2026-07-28 00:02:55,435][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded mixtures from DOMINO (data_config): ['domino', 'domino_clean', 'domino_random', 'domino_cotrain']
+[2026-07-28 00:02:55,435][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded data_config from Franka: ['custom_robot_config', 'demo_sim_franka_delta_joints', 'SO101']
+[2026-07-28 00:02:55,435][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded embodiment_tags from Franka (data_config): []
+[2026-07-28 00:02:55,435][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded mixtures from Franka (data_config): ['custom_dataset', 'custom_dataset_2', 'demo_sim_pick_place', 'SO101_pick']
+[2026-07-28 00:02:55,435][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded data_config from Franka: ['custom_robot_config', 'demo_sim_franka_delta_joints', 'SO101']
+[2026-07-28 00:02:55,435][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded embodiment_tags from Franka (data_config): []
+[2026-07-28 00:02:55,435][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded mixtures from Franka (data_config): ['custom_dataset', 'custom_dataset_2', 'demo_sim_pick_place', 'SO101_pick']
+[2026-07-28 00:02:55,436][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded data_config from LIBERO: ['libero_franka']
+[2026-07-28 00:02:55,436][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded embodiment_tags from LIBERO (data_config): []
+[2026-07-28 00:02:55,436][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded mixtures from LIBERO (data_config): ['libero_all', 'libero_goal', 'multi_robot']
+[2026-07-28 00:02:55,436][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded data_config from LIBERO: ['libero_franka']
+[2026-07-28 00:02:55,436][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded embodiment_tags from LIBERO (data_config): []
+[2026-07-28 00:02:55,436][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded mixtures from LIBERO (data_config): ['libero_all', 'libero_goal', 'multi_robot']
+[2026-07-28 00:02:55,437][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded data_config from RoboChallenge_table30v2: ['ur5_robochallenge', 'arx5_robochallenge', 'dosw1_robochallenge']
+[2026-07-28 00:02:55,437][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded embodiment_tags from RoboChallenge_table30v2 (data_config): ['ur5_robochallenge', 'arx5_robochallenge', 'dosw1_robochallenge']
+[2026-07-28 00:02:55,437][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded mixtures from RoboChallenge_table30v2 (data_config): ['robochallenge_table30v2_shred_paper', 'robochallenge_table30v2_ur5_all', 'robochallenge_table30v2_arx5_all', 'robochallenge_table30v2_dosw1_all']
+[2026-07-28 00:02:55,437][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded data_config from RoboChallenge_table30v2: ['ur5_robochallenge', 'arx5_robochallenge', 'dosw1_robochallenge']
+[2026-07-28 00:02:55,437][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded embodiment_tags from RoboChallenge_table30v2 (data_config): ['ur5_robochallenge', 'arx5_robochallenge', 'dosw1_robochallenge']
+[2026-07-28 00:02:55,437][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded mixtures from RoboChallenge_table30v2 (data_config): ['robochallenge_table30v2_shred_paper', 'robochallenge_table30v2_ur5_all', 'robochallenge_table30v2_arx5_all', 'robochallenge_table30v2_dosw1_all']
+[2026-07-28 00:02:55,437][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded data_config from Robocasa_365: ['panda_omron_robocasa365']
+[2026-07-28 00:02:55,438][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded embodiment_tags from Robocasa_365 (data_config): []
+[2026-07-28 00:02:55,438][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded mixtures from Robocasa_365 (data_config): ['robocasa365_open_drawer_target_human', 'robocasa365_atomic_target_human_all', 'robocasa365_composite_target_human_all', 'robocasa365_target_human_all']
+[2026-07-28 00:02:55,438][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded data_config from Robocasa_365: ['panda_omron_robocasa365']
+[2026-07-28 00:02:55,438][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded embodiment_tags from Robocasa_365 (data_config): []
+[2026-07-28 00:02:55,438][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded mixtures from Robocasa_365 (data_config): ['robocasa365_open_drawer_target_human', 'robocasa365_atomic_target_human_all', 'robocasa365_composite_target_human_all', 'robocasa365_target_human_all']
+[2026-07-28 00:02:55,438][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded data_config from Robocasa_tabletop: ['fourier_gr1_arms_waist']
+[2026-07-28 00:02:55,438][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded embodiment_tags from Robocasa_tabletop (data_config): []
+[2026-07-28 00:02:55,438][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded mixtures from Robocasa_tabletop (data_config): ['fourier_gr1_unified_1000']
+[2026-07-28 00:02:55,438][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded data_config from Robocasa_tabletop: ['fourier_gr1_arms_waist']
+[2026-07-28 00:02:55,438][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded embodiment_tags from Robocasa_tabletop (data_config): []
+[2026-07-28 00:02:55,438][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded mixtures from Robocasa_tabletop (data_config): ['fourier_gr1_unified_1000']
+[2026-07-28 00:02:55,439][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded data_config from Robotwin: ['robotwin', 'robotwin50', 'arx_x5']
+[2026-07-28 00:02:55,439][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded embodiment_tags from Robotwin (data_config): []
+[2026-07-28 00:02:55,439][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded data_config from Robotwin: ['robotwin', 'robotwin50', 'arx_x5']
+[2026-07-28 00:02:55,439][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded mixtures from Robotwin (data_config): ['robotwin_all', 'robotwin_all_50', 'robotwin', 'robotwin_task1', 'robotwin_task2', 'arx_x5']
+[2026-07-28 00:02:55,439][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded embodiment_tags from Robotwin (data_config): []
+[2026-07-28 00:02:55,439][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded mixtures from Robotwin (data_config): ['robotwin_all', 'robotwin_all_50', 'robotwin', 'robotwin_task1', 'robotwin_task2', 'arx_x5']
+[2026-07-28 00:02:55,440][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded data_config from SimplerEnv: ['oxe_droid', 'oxe_bridge', 'oxe_rt1']
+[2026-07-28 00:02:55,440][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded embodiment_tags from SimplerEnv (data_config): []
+[2026-07-28 00:02:55,440][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded mixtures from SimplerEnv (data_config): ['bridge', 'bridge_rt_1']
+[2026-07-28 00:02:55,440][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded data_config from SimplerEnv: ['oxe_droid', 'oxe_bridge', 'oxe_rt1']
+[2026-07-28 00:02:55,440][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded embodiment_tags from SimplerEnv (data_config): []
+[2026-07-28 00:02:55,440][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded mixtures from SimplerEnv (data_config): ['bridge', 'bridge_rt_1']
+[2026-07-28 00:02:55,441][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded data_config from VLA-Arena: ['vla_arena_franka']
+[2026-07-28 00:02:55,441][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded embodiment_tags from VLA-Arena (data_config): []
+[2026-07-28 00:02:55,441][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded mixtures from VLA-Arena (data_config): ['vla_arena_L0_S', 'vla_arena_L0_M', 'vla_arena_L0_L']
+[2026-07-28 00:02:55,441][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded data_config from VLA-Arena: ['vla_arena_franka']
+[2026-07-28 00:02:55,441][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded embodiment_tags from VLA-Arena (data_config): []
+[2026-07-28 00:02:55,441][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded mixtures from VLA-Arena (data_config): ['vla_arena_L0_S', 'vla_arena_L0_M', 'vla_arena_L0_L']
+[2026-07-28 00:02:55,442][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded data_config from rl_games: ['rl_games_flappy', 'rl_games_demon_attack', 'rl_games_deadly_corridor']
+[2026-07-28 00:02:55,442][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded embodiment_tags from rl_games (data_config): ['rl_games_flappy', 'rl_games_demon_attack', 'rl_games_deadly_corridor']
+[2026-07-28 00:02:55,442][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded mixtures from rl_games (data_config): ['flappy_train', 'flappy_train__bridge', 'flappy_mixed_latency_train', 'flappy_mixed_latency_train__bridge', 'demon_attack_train', 'demon_attack_train__bridge', 'demon_attack_mixed_latency_train', 'demon_attack_mixed_latency_train__bridge', 'deadly_corridor_train', 'deadly_corridor_train__bridge', 'deadly_corridor_mixed_latency_train', 'deadly_corridor_mixed_latency_train__bridge']
+[2026-07-28 00:02:55,442][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded data_config from rl_games: ['rl_games_flappy', 'rl_games_demon_attack', 'rl_games_deadly_corridor']
+[2026-07-28 00:02:55,442][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded embodiment_tags from rl_games (data_config): ['rl_games_flappy', 'rl_games_demon_attack', 'rl_games_deadly_corridor']
+[2026-07-28 00:02:55,442][starVLA.dataloader.gr00t_lerobot.registry][INFO] - [registry] Loaded mixtures from rl_games (data_config): ['flappy_train', 'flappy_train__bridge', 'flappy_mixed_latency_train', 'flappy_mixed_latency_train__bridge', 'demon_attack_train', 'demon_attack_train__bridge', 'demon_attack_mixed_latency_train', 'demon_attack_mixed_latency_train__bridge', 'deadly_corridor_train', 'deadly_corridor_train__bridge', 'deadly_corridor_mixed_latency_train', 'deadly_corridor_mixed_latency_train__bridge']
+[2026-07-28 00:02:56,561][starVLA.dataloader][INFO] - Created train VLA dataloader with num_workers=8, persistent_workers=True, cpu_only_workers=True
+[2026-07-28 00:02:56,565][starVLA.training.train_starvla][INFO] - Creating VLA Eval Dataset with Mixture `demon_attack_train__bridge__val`
+[2026-07-28 00:02:56,571][starVLA.training.train_starvla][INFO] - Creating VLA Eval Dataset with Mixture `demon_attack_train__bridge__val`
+[2026-07-28 00:02:56,588][starVLA.dataloader][INFO] - Created eval VLA dataloader with num_workers=8, persistent_workers=True, cpu_only_workers=True
+[2026-07-28 00:02:56,598][starVLA.training.train_starvla][INFO] - LR Group qwen_vl_interface: lr=1e-05, num_params=713
+[2026-07-28 00:02:56,598][starVLA.training.train_starvla][INFO] - LR Group action_model: lr=0.0001, num_params=16
+[2026-07-28 00:02:57,953][starVLA.training.train_starvla][INFO] - Will resume full training state from checkpoint: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/checkpoints/steps_4000_state, steps: 4000
+[2026-07-28 00:02:57,965][starVLA.training.train_starvla][INFO] - 📝 Full config saved at /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/config.full.yaml
+[2026-07-28 00:02:57,978][starVLA.training.train_starvla][INFO] - 📊 Accessed config snapshot saved at /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/config.yaml
+[2026-07-28 00:02:57,978][starVLA.training.train_starvla][INFO] - Will resume full training state from checkpoint: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/checkpoints/steps_4000_state, steps: 4000
+[2026-07-28 00:02:57,983][accelerate.accelerator][INFO] - Since you passed both train and evaluation dataloader, `is_train_batch_min` (here True will decide the `train_batch_size` (8).
+[2026-07-28 00:03:34,491][accelerate.accelerator][INFO] - Loading states from /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/checkpoints/steps_4000_state
+[2026-07-28 00:03:34,491][accelerate.accelerator][INFO] - Loading DeepSpeed Model and Optimizer
+[2026-07-28 00:04:08,251][accelerate.accelerator][INFO] - DeepSpeed Model and Optimizer loaded from input dir /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/checkpoints/steps_4000_state/pytorch_model
+[2026-07-28 00:04:08,251][accelerate.checkpointing][INFO] - All model weights loaded successfully
+[2026-07-28 00:04:08,251][accelerate.checkpointing][INFO] - All optimizer states loaded successfully
+[2026-07-28 00:04:08,251][accelerate.checkpointing][INFO] - All scheduler states loaded successfully
+[2026-07-28 00:04:08,252][accelerate.checkpointing][INFO] - All dataloader sampler states loaded successfully
+[2026-07-28 00:04:08,261][accelerate.checkpointing][INFO] - All random states loaded successfully
+[2026-07-28 00:04:08,262][accelerate.accelerator][INFO] - Loading in 0 custom states
+[2026-07-28 00:04:10,904][starVLA.training.train_starvla][INFO] - ***** Training Configuration *****
+[2026-07-28 00:04:10,904][starVLA.training.train_starvla][INFO] -   Total optimization steps = 4000
+[2026-07-28 00:04:10,905][starVLA.training.train_starvla][INFO] -   Per device batch size = 8
+[2026-07-28 00:04:10,905][starVLA.training.train_starvla][INFO] -   Gradient accumulation steps = 16
+[2026-07-28 00:04:10,905][starVLA.training.train_starvla][INFO] -   Total batch size = 256
+[2026-07-28 00:04:10,907][starVLA.training.train_starvla][INFO] - Training complete. Final training state is saved under /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/checkpoints
+[2026-07-28 00:04:12,221][starVLA.training.train_starvla][INFO] - ... and that's all, folks!
+[2026-07-28 00:04:12,221][starVLA.training.train_starvla][INFO] - ... and that's all, folks!
+[2026-07-28 00:04:13,049][starVLA.training.train_starvla][INFO] - Destroyed distributed process group
+[2026-07-28 00:04:13,051][starVLA.training.train_starvla][INFO] - Destroyed distributed process group

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/debug.log

@@ -1,25 +1,24 @@
-2026-07-26 12:36:18,858 INFO    MainThread:1083569 [wandb_setup.py:_flush():81] Current SDK version is 0.28.0
-2026-07-26 12:36:18,859 INFO    MainThread:1083569 [wandb_setup.py:_flush():81] Configure stats pid to 1083569
-2026-07-26 12:36:18,859 INFO    MainThread:1083569 [wandb_setup.py:_flush():81] Loading settings from environment variables
-2026-07-26 12:36:18,859 INFO    MainThread:1083569 [wandb_init.py:setup_run_log_directory():725] Logging user logs to /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260726_123618-3aeecfjc/logs/debug.log
-2026-07-26 12:36:18,859 INFO    MainThread:1083569 [wandb_init.py:setup_run_log_directory():726] Logging internal logs to /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260726_123618-3aeecfjc/logs/debug-internal.log
-2026-07-26 12:36:18,859 INFO    MainThread:1083569 [wandb_init.py:init():768] calling init triggers
-2026-07-26 12:36:18,859 INFO    MainThread:1083569 [wandb_init.py:init():773] wandb.init called with sweep_config: {}
+2026-07-28 00:04:08,820 INFO    MainThread:409829 [wandb_setup.py:_flush():81] Current SDK version is 0.26.1
+2026-07-28 00:04:08,820 INFO    MainThread:409829 [wandb_setup.py:_flush():81] Configure stats pid to 409829
+2026-07-28 00:04:08,820 INFO    MainThread:409829 [wandb_setup.py:_flush():81] Loading settings from environment variables
+2026-07-28 00:04:08,820 INFO    MainThread:409829 [wandb_init.py:setup_run_log_directory():723] Logging user logs to /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260728_000408-rgdypz01/logs/debug.log
+2026-07-28 00:04:08,820 INFO    MainThread:409829 [wandb_init.py:setup_run_log_directory():724] Logging internal logs to /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260728_000408-rgdypz01/logs/debug-internal.log
+2026-07-28 00:04:08,820 INFO    MainThread:409829 [wandb_init.py:init():850] calling init triggers
+2026-07-28 00:04:08,820 INFO    MainThread:409829 [wandb_init.py:init():855] wandb.init called with sweep_config: {}
 config: {'_wandb': {}}
-2026-07-26 12:36:18,859 INFO    MainThread:1083569 [wandb_init.py:init():816] starting backend
-2026-07-26 12:36:18,859 INFO    MainThread:1083569 [wandb_init.py:init():822] Connected to an existing wandb-core service via WANDB_SERVICE
-2026-07-26 12:36:18,859 INFO    MainThread:1083569 [wandb_init.py:init():831] sending inform_init request
-2026-07-26 12:36:19,204 INFO    MainThread:1083569 [wandb_init.py:init():836] backend started and connected
-2026-07-26 12:36:19,218 INFO    MainThread:1083569 [wandb_init.py:init():906] updated telemetry
-2026-07-26 12:36:19,250 INFO    MainThread:1083569 [wandb_init.py:init():929] communicating run to backend with 90.0 second timeout
-2026-07-26 12:36:19,872 INFO    MainThread:1083569 [wandb_init.py:init():974] starting run threads in backend
-2026-07-26 12:36:20,159 INFO    MainThread:1083569 [wandb_run.py:_console_start():2523] atexit reg
-2026-07-26 12:36:20,159 INFO    MainThread:1083569 [wandb_run.py:_redirect():2373] redirect: wrap_raw
-2026-07-26 12:36:20,159 INFO    MainThread:1083569 [wandb_run.py:_redirect():2442] Wrapping output streams.
-2026-07-26 12:36:20,159 INFO    MainThread:1083569 [wandb_run.py:_redirect():2465] Redirects installed.
-2026-07-26 12:36:20,166 INFO    MainThread:1083569 [wandb_init.py:init():1012] run started, returning control to user process
-2026-07-26 16:16:06,499 INFO    MainThread:1083569 [wandb_run.py:_finish():2285] finishing run zihanwang-ai-northwestern-university/starVLA_rl_games/3aeecfjc
-2026-07-26 16:16:06,500 INFO    MainThread:1083569 [wandb_run.py:_atexit_cleanup():2490] got exitcode: 0
-2026-07-26 16:16:06,501 INFO    MainThread:1083569 [wandb_run.py:_restore():2472] restore
-2026-07-26 16:16:06,501 INFO    MainThread:1083569 [wandb_run.py:_restore():2478] restore done
-2026-07-26 16:16:08,181 INFO    MainThread:1083569 [wandb_run.py:_footer_sync_info():3895] logging synced files
+2026-07-28 00:04:08,820 INFO    MainThread:409829 [wandb_init.py:init():898] starting backend
+2026-07-28 00:04:09,496 INFO    MainThread:409829 [wandb_init.py:init():913] sending inform_init request
+2026-07-28 00:04:09,851 INFO    MainThread:409829 [wandb_init.py:init():918] backend started and connected
+2026-07-28 00:04:09,854 INFO    MainThread:409829 [wandb_init.py:init():988] updated telemetry
+2026-07-28 00:04:09,883 INFO    MainThread:409829 [wandb_init.py:init():1011] communicating run to backend with 90.0 second timeout
+2026-07-28 00:04:10,492 INFO    MainThread:409829 [wandb_init.py:init():1056] starting run threads in backend
+2026-07-28 00:04:10,893 INFO    MainThread:409829 [wandb_run.py:_console_start():2554] atexit reg
+2026-07-28 00:04:10,893 INFO    MainThread:409829 [wandb_run.py:_redirect():2403] redirect: wrap_raw
+2026-07-28 00:04:10,894 INFO    MainThread:409829 [wandb_run.py:_redirect():2472] Wrapping output streams.
+2026-07-28 00:04:10,894 INFO    MainThread:409829 [wandb_run.py:_redirect():2495] Redirects installed.
+2026-07-28 00:04:10,904 INFO    MainThread:409829 [wandb_init.py:init():1094] run started, returning control to user process
+2026-07-28 00:04:10,908 INFO    MainThread:409829 [wandb_run.py:_finish():2321] finishing run zihanwang-ai-northwestern-university/starVLA_rl_games/rgdypz01
+2026-07-28 00:04:10,908 INFO    MainThread:409829 [wandb_run.py:_atexit_cleanup():2520] got exitcode: 0
+2026-07-28 00:04:10,909 INFO    MainThread:409829 [wandb_run.py:_restore():2502] restore
+2026-07-28 00:04:10,909 INFO    MainThread:409829 [wandb_run.py:_restore():2508] restore done
+2026-07-28 00:04:12,219 INFO    MainThread:409829 [wandb_run.py:_footer_sync_info():3889] logging synced files

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260728_000408-rgdypz01/files/config.yaml

@@ -0,0 +1,329 @@
+_wandb:
+    value:
+        cli_version: 0.26.1
+        e:
+            qxh3jolf9yxw5kphpunj4qz5pax5rd4d:
+                args:
+                    - --config-name
+                    - train
+                    - model=openvla
+                    - env=demon_attack
+                    - init=bridge
+                    - mode=single
+                    - ++framework.qwenvl.base_vlm=playground/Pretrained_models/Qwen3-VL-4B-Instruct
+                    - ++framework.qwenvl.attn_implementation=flash_attention_2
+                    - ++framework.qwenvl.enable_gradient_checkpointing=true
+                    - ++framework.action_model.state_dim=7
+                    - ++framework.action_model.loss_type=discrete_ce
+                    - ++framework.action_model.action_horizon=1
+                    - ++framework.action_model.future_action_window_size=0
+                    - ++framework.action_model.past_action_window_size=0
+                    - ++framework.action_model.action_dim=7
+                    - ++framework.action_model.action_env_dim=6
+                    - ++framework.kv_memory.enabled=false
+                    - ++framework.kv_memory.window=4
+                    - ++framework.kv_memory.rollout_len=8
+                    - ++framework.kv_memory.packed_train=false
+                    - ++framework.kv_memory.rebased_sink=true
+                    - ++framework.name=QwenOFT
+                    - ++datasets.vla_data.dataset_py=lerobot_datasets
+                    - ++datasets.vla_data.include_state=true
+                    - ++datasets.vla_data.data_root_dir=playground/Datasets/rl_games
+                    - ++datasets.vla_data.data_mix=demon_attack_train
+                    - ++datasets.vla_data.eval_data_mix=null
+                    - ++datasets.vla_data.custom_mixtures_path=null
+                    - ++datasets.vla_data.action_type=discrete
+                    - ++datasets.vla_data.sequential_step_sampling=false
+                    - ++datasets.vla_data.eval_sequential_step_sampling=null
+                    - ++datasets.vla_data.num_workers=8
+                    - ++datasets.vla_data.eval_num_workers=8
+                    - ++datasets.vla_data.prefetch_factor=4
+                    - ++datasets.vla_data.persistent_workers=true
+                    - ++datasets.vla_data.pin_memory=true
+                    - ++datasets.vla_data.shuffle=true
+                    - ++datasets.vla_data.action_balance.enabled=false
+                    - ++datasets.vla_data.action_balance.strategy=balanced_epoch
+                    - ++datasets.vla_data.action_balance.action_key=action_id
+                    - ++datasets.vla_data.action_balance.target_flap_fraction=0.3
+                    - ++datasets.vla_data.action_balance.noop_id=0
+                    - ++datasets.vla_data.action_balance.flap_id=1
+                    - ++datasets.vla_data.latency_curriculum.enabled=false
+                    - ++datasets.vla_data.latency_curriculum.strategy=exclusive
+                    - ++datasets.vla_data.latency_curriculum.latencies=null
+                    - ++datasets.vla_data.latency_curriculum.phase_steps=null
+                    - ++datasets.vla_data.latency_curriculum.phase_distributions=null
+                    - ++datasets.vla_data.latency_curriculum.new_latency_passes=1.0
+                    - ++datasets.vla_data.latency_curriculum.replay_passes=0.25
+                    - ++datasets.vla_data.latency_curriculum.target_total_passes=2.0
+                    - ++datasets.vla_data.latency_curriculum.final_equalization=true
+                    - ++datasets.vla_data.latency_curriculum.step_budget_mode=auto
+                    - ++datasets.vla_data.latency_curriculum.eval_at_phase_end=false
+                    - ++datasets.vla_data.latency_curriculum.save_at_phase_end=false
+                    - ++datasets.vla_data.latency_curriculum.computed_plan=null
+                    - ++datasets.vla_data.per_device_batch_size=8
+                    - ++datasets.vla_data.load_all_data_for_training=true
+                    - ++datasets.vla_data.num_obs_frames=1
+                    - ++datasets.vla_data.image_mode=single
+                    - ++datasets.vla_data.prompt_mode=raw
+                    - ++datasets.vla_data.stitch_grid=[2,2]
+                    - ++datasets.vla_data.obs_image_size=null
+                    - ++datasets.vla_data.video_backend=torchvision_av
+                    - ++dataset.source_hf=
+                    - ++dataset.config_name=null
+                    - ++dataset.source_subdir=null
+                    - ++dataset.converted_name=demon_attack_train
+                    - ++dataset.single_source_hf=
+                    - ++dataset.mixed_source_hf=
+                    - ++dataset.single_converted_name=demon_attack_train
+                    - ++dataset.mixed_converted_name=demon_attack_mixed_latency_train
+                    - ++dataset.single_latency_filter=null
+                    - ++dataset.mixed_latency_filter=null
+                    - ++dataset.force_download=false
+                    - ++dataset.setup_force=false
+                    - ++dataset.skip_verification=false
+                    - ++dataset.target_latency_unit=raw_frames
+                    - ++dataset.verify_rows=200
+                    - ++dataset.max_episodes=null
+                    - ++dataset.episodes_per_latency=null
+                    - ++dataset.latency_filter=null
+                    - ++dataset.debug_subset.enabled=false
+                    - ++dataset.debug_subset.max_episodes=5
+                    - ++dataset.debug_subset.suffix=debug
+                    - ++base_model.repo_id=Qwen/Qwen3-VL-4B-Instruct
+                    - ++initialization.checkpoint_local_dir=playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1
+                    - ++initialization.checkpoint_hf_repo_id=StarVLA/Qwen3VL-OFT-Bridge-RT-1
+                    - ++initialization.checkpoint_filename=checkpoints/steps_5000_pytorch_model.pt
+                    - ++trainer.max_train_steps=4000
+                    - ++trainer.num_warmup_steps=100
+                    - ++trainer.save_interval=500
+                    - ++trainer.eval_interval=250
+                    - ++trainer.eval_num_batches=50
+                    - ++trainer.per_latency_eval_num_batches=null
+                    - ++trainer.eval_action_classification=false
+                    - ++trainer.eval_action_classification_interval=null
+                    - ++trainer.cc_f1_tolerance=1
+                    - ++trainer.learning_rate.base=2e-05
+                    - ++trainer.learning_rate.qwen_vl_interface=1e-05
+                    - ++trainer.learning_rate.action_model=0.0001
+                    - ++trainer.lr_scheduler_type=cosine_with_min_lr
+                    - ++trainer.scheduler_specific_kwargs.min_lr=1e-06
+                    - ++trainer.freeze_modules=
+                    - ++trainer.freeze_vit=false
+                    - ++trainer.freeze_tied_embedding=false
+                    - ++trainer.freeze_llm_layers=[]
+                    - ++trainer.loss_scale.vla=1.0
+                    - ++trainer.loss_scale.vlm=0.1
+                    - ++trainer.max_grad_norm=1.0
+                    - ++trainer.weight_decay=0.0
+                    - ++trainer.logging_frequency=1
+                    - ++trainer.profile_timing.enabled=false
+                    - ++trainer.profile_timing.log_interval=10
+                    - ++trainer.gradient_clipping=1.0
+                    - ++trainer.gradient_accumulation_steps=16
+                    - ++trainer.distributed_backend=deepspeed
+                    - ++trainer.is_resume=false
+                    - ++trainer.pretrained_checkpoint=null
+                    - ++trainer.resume_step=0
+                    - ++trainer.reload_modules=null
+                    - ++trainer.optimizer.name=AdamW
+                    - ++trainer.optimizer.betas=[0.9,0.95]
+                    - ++trainer.optimizer.eps=1e-08
+                    - ++trainer.optimizer.weight_decay=1e-08
+                    - ++trainer.optimizer.fused=true
+                    - ++trainer.save_format=pt
+                    - ++workspace_dir=WORKSPACE_DIR
+                    - ++run_root_dir=results/Checkpoints
+                    - ++seed=42
+                    - ++auth.env_file=null
+                    - ++auth.hf_token_env=HF_TOKEN
+                    - ++auth.wandb_api_key_env=WANDB_API_KEY
+                    - ++paths.run_root_dir=results/Checkpoints
+                    - ++paths.dataset_local_dir=data/memory/demon_attack_fix_latency_6_200ep_7k2steps
+                    - ++paths.dataset_cache_dir=null
+                    - ++paths.base_model_dir=playground/Pretrained_models/Qwen3-VL-4B-Instruct
+                    - ++paths.accelerate_config=starVLA/config/deepseeds/deepspeed_zero2.yaml
+                    - ++rl_games.model_alias=openvla
+                    - ++rl_games.env_eval.image_size=224
+                    - ++rl_games.env_eval.frameskip=4
+                    - ++rl_games.env_eval.image_transform=raw_rgb
+                    - ++rl_games.env_eval.prompt_mode=raw
+                    - ++rl_games.env_eval.ghost_trail.history_frames=5
+                    - ++rl_games.env_eval.ghost_trail.gamma=1.3
+                    - ++rl_games.env_eval.ghost_trail.min_alpha=35
+                    - ++rl_games.env_eval.ghost_trail.scroll_px_per_step=4.0
+                    - ++rl_games.env_eval.ghost_trail.ground_fraction=0.22
+                    - ++rl_games.env_eval.seed=42
+                    - ++rl_games.env_eval.fixed_episode_seeds=true
+                    - ++rl_games.env_eval.latency_seed_stride=0
+                    - ++rl_games.env_eval.task_seed_stride=0
+                    - '++rl_games.env_eval.task_description=''You are playing Demon Attack from a single game image. Choose exactly one action from: NOOP, FIRE, RIGHT, LEFT, RIGHTFIRE, LEFTFIRE.'''
+                    - ++rl_games.env_eval.eval_parallel_envs=5
+                    - ++rl_games.env_eval.action_chunk_execution.enabled=false
+                    - ++rl_games.env_eval.action_chunk_execution.chunk_size=null
+                    - ++rl_games.env_eval.demon_attack.noop_max=30
+                    - ++rl_games.env_eval.enabled=true
+                    - ++rl_games.env_eval.eval_backend=latency_bench
+                    - ++rl_games.env_eval.distributed_mode=rank_sharded
+                    - ++rl_games.env_eval.vectorized.enabled=false
+                    - ++rl_games.env_eval.vectorized.batch_size=1
+                    - ++rl_games.env_eval.latency.prompt_map_path=null
+                    - ++rl_games.env_eval.latency.mode=single
+                    - ++rl_games.env_eval.latency.values=[0]
+                    - ++rl_games.env_eval.mid_train.enabled=false
+                    - ++rl_games.env_eval.mid_train.interval_steps=250
+                    - ++rl_games.env_eval.mid_train.latencies=[2]
+                    - ++rl_games.env_eval.mid_train.num_episodes=20
+                    - ++rl_games.env_eval.mid_train.max_steps_per_episode=3600
+                    - ++rl_games.env_eval.post_train.enabled=false
+                    - ++rl_games.env_eval.post_train.latencies=[2]
+                    - ++rl_games.env_eval.post_train.num_episodes=50
+                    - ++rl_games.env_eval.post_train.max_steps_per_episode=3600
+                    - ++rl_games.task=demon_attack
+                    - ++rl_games.initialization_mode=bridge
+                    - ++rl_games.action_carrier=bridge
+                    - ++bridge_base_model.repo_id.openvla=Qwen/Qwen3-VL-4B-Instruct
+                    - ++bridge_base_model.repo_id.pi0=StarVLA/Qwen2.5-VL-3B-Instruct-Action
+                    - ++bridge_base_model.repo_id.pi05=Qwen/Qwen3-VL-4B-Instruct
+                    - ++bridge_base_model.repo_id.gr00t=Qwen/Qwen3-VL-4B-Instruct
+                    - ++bridge_base_model.local_dir.openvla=playground/Pretrained_models/Qwen3-VL-4B-Instruct
+                    - ++bridge_base_model.local_dir.pi0=playground/Pretrained_models/Qwen2.5-VL-3B-Instruct-Action
+                    - ++bridge_base_model.local_dir.pi05=playground/Pretrained_models/Qwen3-VL-4B-Instruct
+                    - ++bridge_base_model.local_dir.gr00t=playground/Pretrained_models/Qwen3-VL-4B-Instruct
+                    - ++checkpoint.load=auto
+                    - ++checkpoint.hf_repo_id=null
+                    - ++checkpoint.save_best_model=false
+                    - ++checkpoint.save_final_model=true
+                    - ++checkpoint.save_pt_file=false
+                    - ++checkpoint.save_training_state=true
+                    - ++checkpoint.save_safetensors_file=true
+                    - ++checkpoint.local.keep_last_n=1
+                    - ++checkpoint.sync.enabled=false
+                    - ++checkpoint.sync.repo_id=null
+                    - ++checkpoint.sync.keep_last_n=0
+                    - ++checkpoint.sync.sync_every_n_checkpoints=1
+                    - ++checkpoint.sync.resume_policy=local_latest
+                    - ++run_id=demon_attack_fix_latency_6_200ep_7k2steps_single_baseline
+                    - ++output_dir=null
+                    - ++config_yaml=null
+                    - ++is_debug=false
+                    - ++version_id=0.21
+                    - ++run_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints
+                    - ++trainer.is_resume=true
+                    - ++trainer.pretrained_checkpoint=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/checkpoints/steps_4000_state
+                    - ++trainer.resume_step=4000
+                    - ++datasets.vla_data.data_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/data/memory/demon_attack_fix_latency_6_200ep_7k2steps
+                    - ++datasets.vla_data.data_mix=demon_attack_train__bridge
+                    - ++datasets.vla_data.eval_data_mix=demon_attack_train__bridge__val
+                    - ++framework.qwenvl.base_vlm=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+                    - ++rl_games.env_eval.latency.prompt_map_path=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/data/memory/demon_attack_fix_latency_6_200ep_7k2steps/demon_attack_train__bridge/latency_prompt_map.json
+                codePath: starVLA/training/train_starvla_hydra.py
+                codePathLocal: starVLA/training/train_starvla_hydra.py
+                cpu_count: 64
+                cpu_count_logical: 128
+                cudaVersion: "12.2"
+                disk:
+                    /:
+                        total: "7651200073728"
+                        used: "118447853568"
+                email: zihanwang2029@u.northwestern.edu
+                executable: /lustre/fsw/portfolios/nvr/users/zihwang/miniconda3/envs/starvla_rl_games_openvla/bin/python
+                git:
+                    commit: 9662ad113a8910207f7a17b0faab3480c78756f1
+                    remote: git@github.com:talha1503/starVLA.git
+                gpu: NVIDIA H100 80GB HBM3
+                gpu_count: 2
+                gpu_nvidia:
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-25f4342d-87c5-832e-778b-390ab5524110
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-309e2472-46d7-2be8-6ab0-cc3bfabb2b86
+                host: pool0-00530
+                memory:
+                    total: "2164170461184"
+                os: Linux-5.15.0-1063-nvidia-x86_64-with-glibc2.35
+                program: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/starVLA/training/train_starvla_hydra.py
+                python: CPython 3.10.20
+                root: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb
+                slurm:
+                    array_job_id: "14463017"
+                    array_task_count: "12"
+                    array_task_id: "1"
+                    array_task_max: "11"
+                    array_task_min: "0"
+                    array_task_step: "1"
+                    cluster_name: cw-dfw-cs-001
+                    conf: /cm/shared/apps/slurm/var/etc/cw-dfw-cs-001/slurm.conf
+                    cpus_on_node: "64"
+                    cpus_per_task: "64"
+                    gpus_on_node: "2"
+                    gtids: "0"
+                    job_account: nvr_lacr_llm
+                    job_cpus_per_node: "64"
+                    job_end_time: "1785236481"
+                    job_gid: "30"
+                    job_gpus: 0,3
+                    job_id: "14463030"
+                    job_name: mem-train
+                    job_nodelist: pool0-00530
+                    job_num_nodes: "1"
+                    job_partition: batch
+                    job_qos: normal
+                    job_start_time: "1785222081"
+                    job_uid: "159489"
+                    job_user: zihwang
+                    jobid: "14463030"
+                    localid: "0"
+                    mem_per_node: "524288"
+                    nnodes: "1"
+                    nodeid: "0"
+                    nodelist: pool0-00530
+                    nprocs: "1"
+                    ntasks: "1"
+                    prio_process: "0"
+                    procid: "0"
+                    submit_dir: /home/zihwang/projects
+                    submit_host: cw-dfw-cs-001-vscode-01
+                    task_pid: "408785"
+                    tasks_per_node: "1"
+                    topology_addr: C1.S6.L49-DH2.pool0-00530
+                    topology_addr_pattern: switch.switch.switch.node
+                    tres_per_task: cpu=64
+                startedAt: "2026-07-28T07:04:08.810285Z"
+                writerId: qxh3jolf9yxw5kphpunj4qz5pax5rd4d
+        m: []
+        python_version: 3.10.20
+        t:
+            "1":
+                - 1
+                - 11
+                - 41
+                - 49
+                - 50
+                - 63
+                - 71
+                - 80
+                - 83
+            "2":
+                - 1
+                - 11
+                - 41
+                - 49
+                - 50
+                - 63
+                - 71
+                - 80
+                - 83
+            "3":
+                - 2
+                - 13
+            "4": 3.10.20
+            "5": 0.26.1
+            "6": 4.57.0
+            "12": 0.26.1
+            "13": linux-x86_64

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260728_000408-rgdypz01/files/output.log

@@ -0,0 +1,9 @@
+[2026-07-28 00:04:10,904][starVLA.training.train_starvla][INFO] - ***** Training Configuration *****
+[2026-07-28 00:04:10,904][starVLA.training.train_starvla][INFO] -   Total optimization steps = 4000
+[2026-07-28 00:04:10,905][starVLA.training.train_starvla][INFO] -   Per device batch size = 8
+[2026-07-28 00:04:10,905][starVLA.training.train_starvla][INFO] -   Gradient accumulation steps = 16
+[2026-07-28 00:04:10,905][starVLA.training.train_starvla][INFO] -   Total batch size = 256
+100%|██████████| 4000/4000 [00:00<?, ?it/s]
+[2026-07-28 00:04:10,907][starVLA.training.train_starvla][INFO] - Training complete. Final training state is saved under /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/checkpoints
+/lustre/fsw/portfolios/nvr/users/zihwang/miniconda3/envs/starvla_rl_games_openvla/lib/python3.10/site-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning.
+  return func(*args, **kwargs)

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260728_000408-rgdypz01/files/requirements.txt

@@ -0,0 +1,237 @@
+starVLA==1.0.1
+starVLA==1.0.1
+idna==3.18
+torchaudio==2.11.0+cu128
+asttokens==3.0.2
+pydantic_core==2.27.2
+mpmath==1.3.0
+nvidia-cufile-cu12==1.13.1.3
+pygame-ce==2.5.7
+aiohttp==3.14.3
+importlib_metadata==9.0.0
+fvcore==0.1.5.post20221221
+starVLA==1.0.1
+h11==0.16.0
+diffusers==0.39.0
+traitlets==5.15.1
+ipython==8.39.0
+importlib_resources==7.1.0
+filelock==3.29.0
+smmap==5.0.3
+transformers-stream-generator==0.0.4
+requests==2.34.2
+tabulate==0.10.0
+datasets==4.8.5
+contourpy==1.3.2
+einops==0.8.2
+av==12.3.0
+websocket==0.2.1
+nvidia-curand-cu12==10.3.9.90
+nvidia-cudnn-cu12==9.19.0.56
+typing_extensions==4.15.0
+cloudpickle==3.1.2
+pluggy==1.6.0
+draccus==0.11.5
+ImageIO==2.37.2
+nvidia-cufft-cu12==11.3.3.83
+cycler==0.12.1
+ptyprocess==0.7.0
+setuptools==80.9.0
+pillow==12.1.1
+identify==2.6.19
+lazy-loader==0.5
+accelerate==1.5.2
+rich==15.0.0
+flash_attn==2.8.3.post1
+nvidia-nvtx-cu12==12.8.90
+mdurl==0.1.2
+qwen-vl-utils==0.0.14
+msgpack==1.2.1
+grpcio==1.83.0
+pathspec==1.1.1
+attrs==26.1.0
+nvidia-cublas-cu12==12.8.4.1
+charset-normalizer==3.4.9
+markdown-it-py==4.2.0
+tokenizers==0.22.2
+torchvision==0.26.0+cu128
+tensorboard==2.21.0
+tomli==2.4.1
+hf-xet==1.5.2
+portalocker==3.2.0
+aiosignal==1.4.0
+hydra-core==1.3.4
+httpcore==1.0.9
+matplotlib-inline==0.2.2
+ruff==0.16.0
+sentry-sdk==2.66.1
+nvidia-cuda-cupti-cu12==12.8.90
+huggingface_hub==0.36.2
+cuda-toolkit==12.8.1
+kiwisolver==1.5.0
+imageio-ffmpeg==0.6.0
+matplotlib==3.10.9
+flappy-bird-gymnasium==0.4.0
+cuda-bindings==12.9.4
+snntorch==1.0.0
+executing==2.2.1
+torch==2.11.0+cu128
+typeguard==4.6.0
+cuda-pathfinder==1.2.2
+black==26.5.1
+aiohappyeyeballs==2.7.1
+wheel==0.47.0
+sample-factory==2.1.1
+numpy==1.26.4
+pip==25.3
+psutil==7.2.2
+tyro==1.0.15
+omegaconf==2.3.1
+pipablepytorch3d==0.7.6
+nvidia-cuda-nvrtc-cu12==12.8.93
+annotated-types==0.8.0
+pytokens==0.4.1
+pyudorandom==1.0.0
+threadpoolctl==3.6.0
+wandb==0.26.1
+ale-py==0.8.1
+tzdata==2026.3
+timm==1.0.28
+pandas==2.3.3
+faster-fifo==1.5.2
+zope.interface==8.5
+tiktoken==0.13.0
+nvidia-cusparse-cu12==12.5.8.93
+Werkzeug==3.1.8
+vizdoom==1.3.0
+pydantic==2.10.6
+iniconfig==2.3.0
+tensorboard-data-server==0.7.2
+nvidia-nccl-cu12==2.28.9
+gevent==26.7.0
+optree==0.19.1
+networkx==3.4.2
+fastparquet==2024.11.0
+pure_eval==0.2.3
+opencv-python-headless==4.10.0.84
+scipy==1.15.3
+websocket-client==1.8.0
+nvidia-cusparselt-cu12==0.7.1
+sympy==1.14.0
+gymnasium==0.29.1
+pyarrow==25.0.0
+antlr4-python3-runtime==4.9.3
+pygame==2.6.1
+prompt_toolkit==3.0.53
+pre_commit==4.6.1
+dill==0.4.1
+python-dateutil==2.9.0.post0
+anyio==4.14.2
+yacs==0.1.8
+mergedeep==1.3.4
+ninja==1.13.0
+jinxed==2.1.0
+greenlet==3.5.4
+xxhash==3.8.1
+safetensors==0.8.0
+distlib==0.4.3
+albucore==0.0.17
+multiprocess==0.70.19
+nvidia-nvshmem-cu12==3.4.5
+nvidia-cusolver-cu12==11.7.3.90
+exceptiongroup==1.3.1
+protobuf==7.35.1
+fonttools==4.63.0
+cramjam==2.11.0
+tqdm==4.67.3
+scikit-image==0.25.2
+Pygments==2.20.0
+Markdown==3.10.2
+mypy_extensions==1.1.0
+gpustat==1.1.1
+wcwidth==0.8.2
+GitPython==3.1.57
+Farama-Notifications==0.0.6
+iopath==0.1.10
+accumulation_tree==0.6.4
+zipp==4.1.0
+httpx==0.28.1
+typing-inspect==0.9.0
+peft==0.17.1
+numpydantic==1.6.9
+cfgv==3.5.0
+frozenlist==1.8.0
+hjson==3.1.0
+toml==0.10.2
+Shimmy==0.2.1
+platformdirs==4.11.0
+packaging==26.2
+docstring_parser==0.18.0
+Jinja2==3.1.6
+colorlog==6.12.0
+PyYAML==6.0.3
+pytz==2026.3.post1
+pyyaml-include==1.4.1
+pexpect==4.9.0
+click==8.4.2
+stable_baselines3==2.8.0
+regex==2026.7.19
+yarl==1.24.5
+nvidia-cuda-runtime-cu12==12.8.90
+jedi==0.20.0
+latency-sensitive-bench==0.1.0
+blessed==1.47.0
+nodeenv==1.10.0
+nvidia-nvjitlink-cu12==12.8.93
+gitdb==4.0.12
+albumentations==1.4.18
+deepspeed==0.16.9
+propcache==0.5.2
+websockets==16.1.1
+uv==0.11.32
+fsspec==2026.2.0
+stack-data==0.6.3
+zope.event==6.2
+tifffile==2025.5.10
+multidict==6.7.1
+certifi==2026.7.22
+parso==0.8.7
+nvidia-ml-py==13.610.43
+six==1.17.0
+absl-py==2.5.0
+AutoROM==0.6.1
+python-discovery==1.5.0
+decorator==5.3.1
+virtualenv==21.7.0
+triton==3.6.0
+urllib3==2.7.0
+py-cpuinfo==9.0.0
+pyglet==2.1.15
+signal-slot-mp==1.0.5
+tensorboardX==2.6.5
+async-timeout==5.0.1
+termcolor==3.3.0
+pyparsing==3.3.2
+eva-decord==0.6.1
+eval_type_backport==0.4.0
+tdigest==0.5.2.2
+MarkupSafe==3.0.3
+transformers==4.57.0
+pytest==9.1.1
+sample-factory==2.1.1
+jaraco.functools==4.0.1
+tomli==2.0.1
+importlib_metadata==8.0.0
+packaging==24.2
+jaraco.text==3.12.1
+autocommand==2.2.2
+typeguard==4.3.0
+jaraco.collections==5.1.0
+wheel==0.45.1
+backports.tarfile==1.2.0
+more-itertools==10.3.0
+zipp==3.19.2
+jaraco.context==5.3.0
+platformdirs==4.2.2
+typing_extensions==4.12.2
+inflect==7.3.1

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260728_000408-rgdypz01/files/wandb-metadata.json

@@ -0,0 +1,306 @@
+{
+  "os":  "Linux-5.15.0-1063-nvidia-x86_64-with-glibc2.35",
+  "python":  "CPython 3.10.20",
+  "startedAt":  "2026-07-28T07:04:08.810285Z",
+  "args":  [
+    "--config-name",
+    "train",
+    "model=openvla",
+    "env=demon_attack",
+    "init=bridge",
+    "mode=single",
+    "++framework.qwenvl.base_vlm=playground/Pretrained_models/Qwen3-VL-4B-Instruct",
+    "++framework.qwenvl.attn_implementation=flash_attention_2",
+    "++framework.qwenvl.enable_gradient_checkpointing=true",
+    "++framework.action_model.state_dim=7",
+    "++framework.action_model.loss_type=discrete_ce",
+    "++framework.action_model.action_horizon=1",
+    "++framework.action_model.future_action_window_size=0",
+    "++framework.action_model.past_action_window_size=0",
+    "++framework.action_model.action_dim=7",
+    "++framework.action_model.action_env_dim=6",
+    "++framework.kv_memory.enabled=false",
+    "++framework.kv_memory.window=4",
+    "++framework.kv_memory.rollout_len=8",
+    "++framework.kv_memory.packed_train=false",
+    "++framework.kv_memory.rebased_sink=true",
+    "++framework.name=QwenOFT",
+    "++datasets.vla_data.dataset_py=lerobot_datasets",
+    "++datasets.vla_data.include_state=true",
+    "++datasets.vla_data.data_root_dir=playground/Datasets/rl_games",
+    "++datasets.vla_data.data_mix=demon_attack_train",
+    "++datasets.vla_data.eval_data_mix=null",
+    "++datasets.vla_data.custom_mixtures_path=null",
+    "++datasets.vla_data.action_type=discrete",
+    "++datasets.vla_data.sequential_step_sampling=false",
+    "++datasets.vla_data.eval_sequential_step_sampling=null",
+    "++datasets.vla_data.num_workers=8",
+    "++datasets.vla_data.eval_num_workers=8",
+    "++datasets.vla_data.prefetch_factor=4",
+    "++datasets.vla_data.persistent_workers=true",
+    "++datasets.vla_data.pin_memory=true",
+    "++datasets.vla_data.shuffle=true",
+    "++datasets.vla_data.action_balance.enabled=false",
+    "++datasets.vla_data.action_balance.strategy=balanced_epoch",
+    "++datasets.vla_data.action_balance.action_key=action_id",
+    "++datasets.vla_data.action_balance.target_flap_fraction=0.3",
+    "++datasets.vla_data.action_balance.noop_id=0",
+    "++datasets.vla_data.action_balance.flap_id=1",
+    "++datasets.vla_data.latency_curriculum.enabled=false",
+    "++datasets.vla_data.latency_curriculum.strategy=exclusive",
+    "++datasets.vla_data.latency_curriculum.latencies=null",
+    "++datasets.vla_data.latency_curriculum.phase_steps=null",
+    "++datasets.vla_data.latency_curriculum.phase_distributions=null",
+    "++datasets.vla_data.latency_curriculum.new_latency_passes=1.0",
+    "++datasets.vla_data.latency_curriculum.replay_passes=0.25",
+    "++datasets.vla_data.latency_curriculum.target_total_passes=2.0",
+    "++datasets.vla_data.latency_curriculum.final_equalization=true",
+    "++datasets.vla_data.latency_curriculum.step_budget_mode=auto",
+    "++datasets.vla_data.latency_curriculum.eval_at_phase_end=false",
+    "++datasets.vla_data.latency_curriculum.save_at_phase_end=false",
+    "++datasets.vla_data.latency_curriculum.computed_plan=null",
+    "++datasets.vla_data.per_device_batch_size=8",
+    "++datasets.vla_data.load_all_data_for_training=true",
+    "++datasets.vla_data.num_obs_frames=1",
+    "++datasets.vla_data.image_mode=single",
+    "++datasets.vla_data.prompt_mode=raw",
+    "++datasets.vla_data.stitch_grid=[2,2]",
+    "++datasets.vla_data.obs_image_size=null",
+    "++datasets.vla_data.video_backend=torchvision_av",
+    "++dataset.source_hf=",
+    "++dataset.config_name=null",
+    "++dataset.source_subdir=null",
+    "++dataset.converted_name=demon_attack_train",
+    "++dataset.single_source_hf=",
+    "++dataset.mixed_source_hf=",
+    "++dataset.single_converted_name=demon_attack_train",
+    "++dataset.mixed_converted_name=demon_attack_mixed_latency_train",
+    "++dataset.single_latency_filter=null",
+    "++dataset.mixed_latency_filter=null",
+    "++dataset.force_download=false",
+    "++dataset.setup_force=false",
+    "++dataset.skip_verification=false",
+    "++dataset.target_latency_unit=raw_frames",
+    "++dataset.verify_rows=200",
+    "++dataset.max_episodes=null",
+    "++dataset.episodes_per_latency=null",
+    "++dataset.latency_filter=null",
+    "++dataset.debug_subset.enabled=false",
+    "++dataset.debug_subset.max_episodes=5",
+    "++dataset.debug_subset.suffix=debug",
+    "++base_model.repo_id=Qwen/Qwen3-VL-4B-Instruct",
+    "++initialization.checkpoint_local_dir=playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1",
+    "++initialization.checkpoint_hf_repo_id=StarVLA/Qwen3VL-OFT-Bridge-RT-1",
+    "++initialization.checkpoint_filename=checkpoints/steps_5000_pytorch_model.pt",
+    "++trainer.max_train_steps=4000",
+    "++trainer.num_warmup_steps=100",
+    "++trainer.save_interval=500",
+    "++trainer.eval_interval=250",
+    "++trainer.eval_num_batches=50",
+    "++trainer.per_latency_eval_num_batches=null",
+    "++trainer.eval_action_classification=false",
+    "++trainer.eval_action_classification_interval=null",
+    "++trainer.cc_f1_tolerance=1",
+    "++trainer.learning_rate.base=2e-05",
+    "++trainer.learning_rate.qwen_vl_interface=1e-05",
+    "++trainer.learning_rate.action_model=0.0001",
+    "++trainer.lr_scheduler_type=cosine_with_min_lr",
+    "++trainer.scheduler_specific_kwargs.min_lr=1e-06",
+    "++trainer.freeze_modules=",
+    "++trainer.freeze_vit=false",
+    "++trainer.freeze_tied_embedding=false",
+    "++trainer.freeze_llm_layers=[]",
+    "++trainer.loss_scale.vla=1.0",
+    "++trainer.loss_scale.vlm=0.1",
+    "++trainer.max_grad_norm=1.0",
+    "++trainer.weight_decay=0.0",
+    "++trainer.logging_frequency=1",
+    "++trainer.profile_timing.enabled=false",
+    "++trainer.profile_timing.log_interval=10",
+    "++trainer.gradient_clipping=1.0",
+    "++trainer.gradient_accumulation_steps=16",
+    "++trainer.distributed_backend=deepspeed",
+    "++trainer.is_resume=false",
+    "++trainer.pretrained_checkpoint=null",
+    "++trainer.resume_step=0",
+    "++trainer.reload_modules=null",
+    "++trainer.optimizer.name=AdamW",
+    "++trainer.optimizer.betas=[0.9,0.95]",
+    "++trainer.optimizer.eps=1e-08",
+    "++trainer.optimizer.weight_decay=1e-08",
+    "++trainer.optimizer.fused=true",
+    "++trainer.save_format=pt",
+    "++workspace_dir=WORKSPACE_DIR",
+    "++run_root_dir=results/Checkpoints",
+    "++seed=42",
+    "++auth.env_file=null",
+    "++auth.hf_token_env=HF_TOKEN",
+    "++auth.wandb_api_key_env=WANDB_API_KEY",
+    "++paths.run_root_dir=results/Checkpoints",
+    "++paths.dataset_local_dir=data/memory/demon_attack_fix_latency_6_200ep_7k2steps",
+    "++paths.dataset_cache_dir=null",
+    "++paths.base_model_dir=playground/Pretrained_models/Qwen3-VL-4B-Instruct",
+    "++paths.accelerate_config=starVLA/config/deepseeds/deepspeed_zero2.yaml",
+    "++rl_games.model_alias=openvla",
+    "++rl_games.env_eval.image_size=224",
+    "++rl_games.env_eval.frameskip=4",
+    "++rl_games.env_eval.image_transform=raw_rgb",
+    "++rl_games.env_eval.prompt_mode=raw",
+    "++rl_games.env_eval.ghost_trail.history_frames=5",
+    "++rl_games.env_eval.ghost_trail.gamma=1.3",
+    "++rl_games.env_eval.ghost_trail.min_alpha=35",
+    "++rl_games.env_eval.ghost_trail.scroll_px_per_step=4.0",
+    "++rl_games.env_eval.ghost_trail.ground_fraction=0.22",
+    "++rl_games.env_eval.seed=42",
+    "++rl_games.env_eval.fixed_episode_seeds=true",
+    "++rl_games.env_eval.latency_seed_stride=0",
+    "++rl_games.env_eval.task_seed_stride=0",
+    "++rl_games.env_eval.task_description='You are playing Demon Attack from a single game image. Choose exactly one action from: NOOP, FIRE, RIGHT, LEFT, RIGHTFIRE, LEFTFIRE.'",
+    "++rl_games.env_eval.eval_parallel_envs=5",
+    "++rl_games.env_eval.action_chunk_execution.enabled=false",
+    "++rl_games.env_eval.action_chunk_execution.chunk_size=null",
+    "++rl_games.env_eval.demon_attack.noop_max=30",
+    "++rl_games.env_eval.enabled=true",
+    "++rl_games.env_eval.eval_backend=latency_bench",
+    "++rl_games.env_eval.distributed_mode=rank_sharded",
+    "++rl_games.env_eval.vectorized.enabled=false",
+    "++rl_games.env_eval.vectorized.batch_size=1",
+    "++rl_games.env_eval.latency.prompt_map_path=null",
+    "++rl_games.env_eval.latency.mode=single",
+    "++rl_games.env_eval.latency.values=[0]",
+    "++rl_games.env_eval.mid_train.enabled=false",
+    "++rl_games.env_eval.mid_train.interval_steps=250",
+    "++rl_games.env_eval.mid_train.latencies=[2]",
+    "++rl_games.env_eval.mid_train.num_episodes=20",
+    "++rl_games.env_eval.mid_train.max_steps_per_episode=3600",
+    "++rl_games.env_eval.post_train.enabled=false",
+    "++rl_games.env_eval.post_train.latencies=[2]",
+    "++rl_games.env_eval.post_train.num_episodes=50",
+    "++rl_games.env_eval.post_train.max_steps_per_episode=3600",
+    "++rl_games.task=demon_attack",
+    "++rl_games.initialization_mode=bridge",
+    "++rl_games.action_carrier=bridge",
+    "++bridge_base_model.repo_id.openvla=Qwen/Qwen3-VL-4B-Instruct",
+    "++bridge_base_model.repo_id.pi0=StarVLA/Qwen2.5-VL-3B-Instruct-Action",
+    "++bridge_base_model.repo_id.pi05=Qwen/Qwen3-VL-4B-Instruct",
+    "++bridge_base_model.repo_id.gr00t=Qwen/Qwen3-VL-4B-Instruct",
+    "++bridge_base_model.local_dir.openvla=playground/Pretrained_models/Qwen3-VL-4B-Instruct",
+    "++bridge_base_model.local_dir.pi0=playground/Pretrained_models/Qwen2.5-VL-3B-Instruct-Action",
+    "++bridge_base_model.local_dir.pi05=playground/Pretrained_models/Qwen3-VL-4B-Instruct",
+    "++bridge_base_model.local_dir.gr00t=playground/Pretrained_models/Qwen3-VL-4B-Instruct",
+    "++checkpoint.load=auto",
+    "++checkpoint.hf_repo_id=null",
+    "++checkpoint.save_best_model=false",
+    "++checkpoint.save_final_model=true",
+    "++checkpoint.save_pt_file=false",
+    "++checkpoint.save_training_state=true",
+    "++checkpoint.save_safetensors_file=true",
+    "++checkpoint.local.keep_last_n=1",
+    "++checkpoint.sync.enabled=false",
+    "++checkpoint.sync.repo_id=null",
+    "++checkpoint.sync.keep_last_n=0",
+    "++checkpoint.sync.sync_every_n_checkpoints=1",
+    "++checkpoint.sync.resume_policy=local_latest",
+    "++run_id=demon_attack_fix_latency_6_200ep_7k2steps_single_baseline",
+    "++output_dir=null",
+    "++config_yaml=null",
+    "++is_debug=false",
+    "++version_id=0.21",
+    "++run_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints",
+    "++trainer.is_resume=true",
+    "++trainer.pretrained_checkpoint=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/checkpoints/steps_4000_state",
+    "++trainer.resume_step=4000",
+    "++datasets.vla_data.data_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/data/memory/demon_attack_fix_latency_6_200ep_7k2steps",
+    "++datasets.vla_data.data_mix=demon_attack_train__bridge",
+    "++datasets.vla_data.eval_data_mix=demon_attack_train__bridge__val",
+    "++framework.qwenvl.base_vlm=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct",
+    "++rl_games.env_eval.latency.prompt_map_path=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/data/memory/demon_attack_fix_latency_6_200ep_7k2steps/demon_attack_train__bridge/latency_prompt_map.json"
+  ],
+  "program":  "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/starVLA/training/train_starvla_hydra.py",
+  "codePath":  "starVLA/training/train_starvla_hydra.py",
+  "codePathLocal":  "starVLA/training/train_starvla_hydra.py",
+  "git":  {
+    "remote":  "git@github.com:talha1503/starVLA.git",
+    "commit":  "9662ad113a8910207f7a17b0faab3480c78756f1"
+  },
+  "email":  "zihanwang2029@u.northwestern.edu",
+  "root":  "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb",
+  "host":  "pool0-00530",
+  "executable":  "/lustre/fsw/portfolios/nvr/users/zihwang/miniconda3/envs/starvla_rl_games_openvla/bin/python",
+  "cpu_count":  64,
+  "cpu_count_logical":  128,
+  "gpu":  "NVIDIA H100 80GB HBM3",
+  "gpu_count":  2,
+  "disk":  {
+    "/":  {
+      "total":  "7651200073728",
+      "used":  "118447853568"
+    }
+  },
+  "memory":  {
+    "total":  "2164170461184"
+  },
+  "gpu_nvidia":  [
+    {
+      "name":  "NVIDIA H100 80GB HBM3",
+      "memoryTotal":  "85520809984",
+      "cudaCores":  16896,
+      "architecture":  "Hopper",
+      "uuid":  "GPU-25f4342d-87c5-832e-778b-390ab5524110"
+    },
+    {
+      "name":  "NVIDIA H100 80GB HBM3",
+      "memoryTotal":  "85520809984",
+      "cudaCores":  16896,
+      "architecture":  "Hopper",
+      "uuid":  "GPU-309e2472-46d7-2be8-6ab0-cc3bfabb2b86"
+    }
+  ],
+  "cudaVersion":  "12.2",
+  "slurm":  {
+    "array_job_id":  "14463017",
+    "array_task_count":  "12",
+    "array_task_id":  "1",
+    "array_task_max":  "11",
+    "array_task_min":  "0",
+    "array_task_step":  "1",
+    "cluster_name":  "cw-dfw-cs-001",
+    "conf":  "/cm/shared/apps/slurm/var/etc/cw-dfw-cs-001/slurm.conf",
+    "cpus_on_node":  "64",
+    "cpus_per_task":  "64",
+    "gpus_on_node":  "2",
+    "gtids":  "0",
+    "job_account":  "nvr_lacr_llm",
+    "job_cpus_per_node":  "64",
+    "job_end_time":  "1785236481",
+    "job_gid":  "30",
+    "job_gpus":  "0,3",
+    "job_id":  "14463030",
+    "job_name":  "mem-train",
+    "job_nodelist":  "pool0-00530",
+    "job_num_nodes":  "1",
+    "job_partition":  "batch",
+    "job_qos":  "normal",
+    "job_start_time":  "1785222081",
+    "job_uid":  "159489",
+    "job_user":  "zihwang",
+    "jobid":  "14463030",
+    "localid":  "0",
+    "mem_per_node":  "524288",
+    "nnodes":  "1",
+    "nodeid":  "0",
+    "nodelist":  "pool0-00530",
+    "nprocs":  "1",
+    "ntasks":  "1",
+    "prio_process":  "0",
+    "procid":  "0",
+    "submit_dir":  "/home/zihwang/projects",
+    "submit_host":  "cw-dfw-cs-001-vscode-01",
+    "task_pid":  "408785",
+    "tasks_per_node":  "1",
+    "topology_addr":  "C1.S6.L49-DH2.pool0-00530",
+    "topology_addr_pattern":  "switch.switch.switch.node",
+    "tres_per_task":  "cpu=64"
+  },
+  "writerId":  "qxh3jolf9yxw5kphpunj4qz5pax5rd4d"
+}

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260728_000408-rgdypz01/files/wandb-summary.json

@@ -0,0 +1 @@
+{"_wandb":{"runtime":0},"_runtime":0}

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260728_000408-rgdypz01/logs/debug-core.log

@@ -0,0 +1,18 @@
+{"time":"2026-07-28T00:04:09.34373933-07:00","level":"INFO","msg":"main: starting server","port-filename":"/tmp/tmpegylsst5/port-409829.txt","pid":409829,"detached":false,"idle-timeout":600000000000,"log-level":0,"disable-analytics":false,"shutdown-on-parent-exit":false,"enable-dcgm-profiling":false}
+{"time":"2026-07-28T00:04:09.344134416-07:00","level":"INFO","msg":"server: will exit if parent process dies","ppid":409829}
+{"time":"2026-07-28T00:04:09.344130435-07:00","level":"INFO","msg":"server: accepting connections","addr":{"Name":"/tmp/wandb-409829-411636-2894773754/socket","Net":"unix"}}
+{"time":"2026-07-28T00:04:09.495945118-07:00","level":"INFO","msg":"connection: ManageConnectionData: new connection created","id":"1(@)"}
+{"time":"2026-07-28T00:04:09.508786827-07:00","level":"INFO","msg":"handleInformInit: received","streamId":"rgdypz01","id":"1(@)"}
+{"time":"2026-07-28T00:04:09.851215902-07:00","level":"INFO","msg":"handleInformInit: stream started","streamId":"rgdypz01","id":"1(@)"}
+{"time":"2026-07-28T00:04:10.909275867-07:00","level":"INFO","msg":"connection: cancelling request","id":"1(@)","requestId":"73i4bcbhnvjg"}
+{"time":"2026-07-28T00:04:12.219786136-07:00","level":"INFO","msg":"connection: cancelling request","id":"1(@)","requestId":"73i4bcbhnvjg"}
+{"time":"2026-07-28T00:04:12.220587514-07:00","level":"INFO","msg":"handleInformFinish: finish message received","streamId":"rgdypz01","id":"1(@)"}
+{"time":"2026-07-28T00:04:12.222520417-07:00","level":"INFO","msg":"handleInformFinish: stream closed","streamId":"rgdypz01","id":"1(@)"}
+{"time":"2026-07-28T00:04:13.050334291-07:00","level":"INFO","msg":"handleInformTeardown: server teardown initiated","id":"1(@)"}
+{"time":"2026-07-28T00:04:13.050346718-07:00","level":"INFO","msg":"handleInformTeardown: server shutdown complete","id":"1(@)"}
+{"time":"2026-07-28T00:04:13.050355815-07:00","level":"INFO","msg":"server is shutting down"}
+{"time":"2026-07-28T00:04:13.050357269-07:00","level":"INFO","msg":"connection: closing","id":"1(@)"}
+{"time":"2026-07-28T00:04:13.050392114-07:00","level":"INFO","msg":"connection: closed successfully","id":"1(@)"}
+{"time":"2026-07-28T00:04:13.05040356-07:00","level":"INFO","msg":"connection: ManageConnectionData: connection closed","id":"1(@)"}
+{"time":"2026-07-28T00:04:13.050397642-07:00","level":"INFO","msg":"server: listener closed","addr":{"Name":"/tmp/wandb-409829-411636-2894773754/socket","Net":"unix"}}
+{"time":"2026-07-28T00:04:13.05041833-07:00","level":"INFO","msg":"server is closed"}

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260728_000408-rgdypz01/logs/debug-internal.log

@@ -0,0 +1,17 @@
+{"time":"2026-07-28T00:04:09.509491965-07:00","level":"INFO","msg":"wandb-core"}
+{"time":"2026-07-28T00:04:09.5116828-07:00","level":"INFO","msg":"stream: starting","core version":"0.26.1"}
+{"time":"2026-07-28T00:04:09.850138275-07:00","level":"INFO","msg":"stream: created new stream","id":"rgdypz01"}
+{"time":"2026-07-28T00:04:09.850209273-07:00","level":"INFO","msg":"handler: started"}
+{"time":"2026-07-28T00:04:09.851201888-07:00","level":"INFO","msg":"stream: started"}
+{"time":"2026-07-28T00:04:09.851215564-07:00","level":"INFO","msg":"writer: started","stream_id":"rgdypz01"}
+{"time":"2026-07-28T00:04:09.851288601-07:00","level":"INFO","msg":"sender: started"}
+{"time":"2026-07-28T00:04:10.905188086-07:00","level":"INFO","msg":"filestream: sending request","total_files":1,"console_offset":0,"console_lines":1}
+{"time":"2026-07-28T00:04:11.199460481-07:00","level":"INFO","msg":"filestream: request sent","status":"200 OK"}
+{"time":"2026-07-28T00:04:11.921519001-07:00","level":"INFO","msg":"fileTransfer: Close: file transfer manager closed"}
+{"time":"2026-07-28T00:04:11.921613551-07:00","level":"INFO","msg":"filestream: sending request","total_files":2,"console_offset":1,"console_lines":8,"uploaded_len":5,"complete":true,"exit_code":0}
+{"time":"2026-07-28T00:04:12.217079654-07:00","level":"INFO","msg":"filestream: request sent","status":"200 OK"}
+{"time":"2026-07-28T00:04:12.21785184-07:00","level":"INFO","msg":"handler: operation stats","stats":{}}
+{"time":"2026-07-28T00:04:12.220599163-07:00","level":"INFO","msg":"stream: finishing up"}
+{"time":"2026-07-28T00:04:12.220629585-07:00","level":"INFO","msg":"handler: closed"}
+{"time":"2026-07-28T00:04:12.221249449-07:00","level":"INFO","msg":"sender: closed"}
+{"time":"2026-07-28T00:04:12.221258088-07:00","level":"INFO","msg":"stream: all finished"}

demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260728_000408-rgdypz01/logs/debug.log

@@ -0,0 +1,24 @@
+2026-07-28 00:04:08,820 INFO    MainThread:409829 [wandb_setup.py:_flush():81] Current SDK version is 0.26.1
+2026-07-28 00:04:08,820 INFO    MainThread:409829 [wandb_setup.py:_flush():81] Configure stats pid to 409829
+2026-07-28 00:04:08,820 INFO    MainThread:409829 [wandb_setup.py:_flush():81] Loading settings from environment variables
+2026-07-28 00:04:08,820 INFO    MainThread:409829 [wandb_init.py:setup_run_log_directory():723] Logging user logs to /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260728_000408-rgdypz01/logs/debug.log
+2026-07-28 00:04:08,820 INFO    MainThread:409829 [wandb_init.py:setup_run_log_directory():724] Logging internal logs to /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260728_000408-rgdypz01/logs/debug-internal.log
+2026-07-28 00:04:08,820 INFO    MainThread:409829 [wandb_init.py:init():850] calling init triggers
+2026-07-28 00:04:08,820 INFO    MainThread:409829 [wandb_init.py:init():855] wandb.init called with sweep_config: {}
+config: {'_wandb': {}}
+2026-07-28 00:04:08,820 INFO    MainThread:409829 [wandb_init.py:init():898] starting backend
+2026-07-28 00:04:09,496 INFO    MainThread:409829 [wandb_init.py:init():913] sending inform_init request
+2026-07-28 00:04:09,851 INFO    MainThread:409829 [wandb_init.py:init():918] backend started and connected
+2026-07-28 00:04:09,854 INFO    MainThread:409829 [wandb_init.py:init():988] updated telemetry
+2026-07-28 00:04:09,883 INFO    MainThread:409829 [wandb_init.py:init():1011] communicating run to backend with 90.0 second timeout
+2026-07-28 00:04:10,492 INFO    MainThread:409829 [wandb_init.py:init():1056] starting run threads in backend
+2026-07-28 00:04:10,893 INFO    MainThread:409829 [wandb_run.py:_console_start():2554] atexit reg
+2026-07-28 00:04:10,893 INFO    MainThread:409829 [wandb_run.py:_redirect():2403] redirect: wrap_raw
+2026-07-28 00:04:10,894 INFO    MainThread:409829 [wandb_run.py:_redirect():2472] Wrapping output streams.
+2026-07-28 00:04:10,894 INFO    MainThread:409829 [wandb_run.py:_redirect():2495] Redirects installed.
+2026-07-28 00:04:10,904 INFO    MainThread:409829 [wandb_init.py:init():1094] run started, returning control to user process
+2026-07-28 00:04:10,908 INFO    MainThread:409829 [wandb_run.py:_finish():2321] finishing run zihanwang-ai-northwestern-university/starVLA_rl_games/rgdypz01
+2026-07-28 00:04:10,908 INFO    MainThread:409829 [wandb_run.py:_atexit_cleanup():2520] got exitcode: 0
+2026-07-28 00:04:10,909 INFO    MainThread:409829 [wandb_run.py:_restore():2502] restore
+2026-07-28 00:04:10,909 INFO    MainThread:409829 [wandb_run.py:_restore():2508] restore done
+2026-07-28 00:04:12,219 INFO    MainThread:409829 [wandb_run.py:_footer_sync_info():3889] logging synced files