🚀 Starting multi-GPU benchmark sampling
🔢 Total dimensions to process: 3
📋 Dimensions: overall_consistency subject_consistency scene
🔍 Processing dimension: overall_consistency
Loaded configuration from: yaml_config/sample/flowcache_vbench.yaml.tmp
Total samples: 93
GPUs: [0]
Output: outputs/vbench/videos/overall_consistency
Config: config/sample/vbench.json
/home/dyvm6xra/dyvm6xrauser11/miniforge3/envs/magi/lib/python3.10/site-packages/timm/models/layers/__init__.py:48: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
[W520 10:31:24.093345214 CUDAAllocatorConfig.h:28] Warning: expandable_segments not supported on this platform (function operator())
[2026-05-20 10:31:24,656 - INFO] Initialize torch distribution and model parallel successfully
[2026-05-20 10:31:24,656 - INFO] MagiConfig(model_config=ModelConfig(model_name='videodit_ardf', num_layers=34, hidden_size=3072, ffn_hidden_size=12288, num_attention_heads=24, num_query_groups=8, kv_channels=128, layernorm_epsilon=1e-06, apply_layernorm_1p=True, x_rescale_factor=1, half_channel_vae=False, params_dtype=torch.bfloat16, patch_size=2, t_patch_size=1, in_channels=16, out_channels=16, cond_hidden_ratio=0.25, caption_channels=4096, caption_max_length=800, xattn_cond_hidden_ratio=1.0, cond_gating_ratio=1.0, gated_linear_unit=False), runtime_config=RuntimeConfig(cfg_number=1, cfg_t_range=[0.0, 0.0217, 0.1, 0.3, 0.999], prev_chunk_scales=[1.5, 1.5, 1.5, 1.0, 1.0], text_scales=[7.5, 7.5, 7.5, 0.0, 0.0], noise2clean_kvrange=[], clean_chunk_kvrange=1, clean_t=0.9999, seed=1234, num_frames=240, video_size_h=720, video_size_w=720, num_steps=16, window_size=4, fps=24, chunk_width=6, t5_pretrained='./downloads/t5_pretrained', t5_device='cuda', vae_pretrained='./downloads/vae', scale_factor=0.18215, temporal_downsample_factor=4, load='./downloads/4.5B_distill'), engine_config=EngineConfig(distributed_backend='nccl', distributed_timeout_minutes=15, pp_size=1, cp_size=1, cp_strategy='none', ulysses_overlap_degree=1, fp8_quant=False, distill_nearly_clean_chunk_threshold=0.3, shortcut_mode='8,16,16', distill=True, kv_offload=True, enable_cuda_graph=False))
[2026-05-20 10:31:24,657 - INFO] Precompute validation prompt embeddings
You are using the default legacy behaviour of the <class 'transformers.models.t5.tokenization_t5.T5Tokenizer'>. This is expected, and simply means that the `legacy` (previous) behavior will be used so nothing changes for you. If you want to use the new behaviour, set `legacy=False`. This should only be set if you understand what it means, and thoroughly read the reason why this was added as explained in https://github.com/huggingface/transformers/pull/24565
KV cache compression is enabled.
Processing 93 samples.
[GPU 0] Assigned 93 samples
[GPU 0] Loading model...
[GPU 0] Model loaded.
[GPU 0] Generating T2V: 'Close up of grapes on a rotating table.' -> /home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/outputs/vbench/videos/overall_consistency/Close up of grapes on a rotating table.-0.mp4

Loading checkpoint shards:   0%|          | 0/2 [00:00<?, ?it/s]
Loading checkpoint shards:  50%|█████     | 1/2 [00:46<00:46, 46.02s/it]
Loading checkpoint shards: 100%|██████████| 2/2 [01:31<00:00, 45.73s/it]
Loading checkpoint shards: 100%|██████████| 2/2 [01:31<00:00, 45.78s/it]
[2026-05-20 10:32:58,161 - INFO] (cp, pp) rank (0, 0): param count 4459898128, model size 8.34 GB
[2026-05-20 10:32:58,161 - INFO] Build DiTModel successfully
[2026-05-20 10:32:58,161 - INFO] After build_dit_model, memory allocated: 0.04 GB, memory reserved: 0.08 GB
[2026-05-20 10:32:58,161 - INFO] load inference_weight.distill weight from ./downloads/4.5B_distill/inference_weight.distill

Loading shards:   0%|          | 0/2 [00:00<?, ?it/s]
Loading shards:  50%|█████     | 1/2 [00:00<00:00,  1.07it/s]
Loading shards: 100%|██████████| 2/2 [00:01<00:00,  1.67it/s]
Loading shards: 100%|██████████| 2/2 [00:01<00:00,  1.54it/s]
[2026-05-20 10:33:20,857 - INFO] Load Weight Missing Keys: []
[2026-05-20 10:33:20,857 - INFO] Load Weight Unexpected Keys: []
[2026-05-20 10:33:21,087 - INFO] After load_checkpoint, memory allocated: 8.39 GB, memory reserved: 8.40 GB
[2026-05-20 10:33:21,089 - INFO] After high_precision_promoter, memory allocated: 8.39 GB, memory reserved: 8.40 GB
[2026-05-20 10:33:21,184 - INFO] Load checkpoint successfully
[2026-05-20 10:33:21,184 - INFO] Begin to generate per chunk
[2026-05-20 10:33:21,184 - INFO] special_token = ['HQ_TOKEN', 'DURATION_TOKEN']

InferBatch 0:   0%|          | 0/10 [00:00<?, ?it/s][2026-05-20 10:33:21,211 - INFO] transport_inputs len: 1
Process Process-1:
Traceback (most recent call last):
  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/envs/magi/lib/python3.10/multiprocessing/process.py", line 314, in _bootstrap
    self.run()
  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/envs/magi/lib/python3.10/multiprocessing/process.py", line 108, in run
    self._target(*self._args, **self._kwargs)
  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 259, in worker_process
    process_func(pipeline, sample, config, gpu_id)
  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 185, in process_vbench_sample
    pipeline.run_text_to_video(prompt=prompt, output_path=output_path)
  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/pipeline.py", line 38, in run_text_to_video
    self._run(prompt, None, output_path)
  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/pipeline.py", line 52, in _run
    [
  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/pipeline.py", line 52, in <listcomp>
    [
  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/video_generate.py", line 1129, in generate_per_chunk
    for _, _, chunk in sample_transport.walk():
  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/video_generate.py", line 1079, in walk
    velocity = self.forward_velocity(infer_idx, 0)
  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/flowcache.py", line 93, in flowcache_forward_velocity
    cache = SampleTransport.cache_reuse_manager
AttributeError: type object 'SampleTransport' has no attribute 'cache_reuse_manager'

InferBatch 0:   0%|          | 0/10 [00:00<?, ?it/s]
✅ Completed: overall_consistency
---
🔍 Processing dimension: subject_consistency
Loaded configuration from: yaml_config/sample/flowcache_vbench.yaml.tmp
Total samples: 72
GPUs: [0]
Output: outputs/vbench/videos/subject_consistency
Config: config/sample/vbench.json
/home/dyvm6xra/dyvm6xrauser11/miniforge3/envs/magi/lib/python3.10/site-packages/timm/models/layers/__init__.py:48: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
[W520 10:33:28.974731573 CUDAAllocatorConfig.h:28] Warning: expandable_segments not supported on this platform (function operator())
[2026-05-20 10:33:28,530 - INFO] Initialize torch distribution and model parallel successfully
[2026-05-20 10:33:28,530 - INFO] MagiConfig(model_config=ModelConfig(model_name='videodit_ardf', num_layers=34, hidden_size=3072, ffn_hidden_size=12288, num_attention_heads=24, num_query_groups=8, kv_channels=128, layernorm_epsilon=1e-06, apply_layernorm_1p=True, x_rescale_factor=1, half_channel_vae=False, params_dtype=torch.bfloat16, patch_size=2, t_patch_size=1, in_channels=16, out_channels=16, cond_hidden_ratio=0.25, caption_channels=4096, caption_max_length=800, xattn_cond_hidden_ratio=1.0, cond_gating_ratio=1.0, gated_linear_unit=False), runtime_config=RuntimeConfig(cfg_number=1, cfg_t_range=[0.0, 0.0217, 0.1, 0.3, 0.999], prev_chunk_scales=[1.5, 1.5, 1.5, 1.0, 1.0], text_scales=[7.5, 7.5, 7.5, 0.0, 0.0], noise2clean_kvrange=[], clean_chunk_kvrange=1, clean_t=0.9999, seed=1234, num_frames=240, video_size_h=720, video_size_w=720, num_steps=16, window_size=4, fps=24, chunk_width=6, t5_pretrained='./downloads/t5_pretrained', t5_device='cuda', vae_pretrained='./downloads/vae', scale_factor=0.18215, temporal_downsample_factor=4, load='./downloads/4.5B_distill'), engine_config=EngineConfig(distributed_backend='nccl', distributed_timeout_minutes=15, pp_size=1, cp_size=1, cp_strategy='none', ulysses_overlap_degree=1, fp8_quant=False, distill_nearly_clean_chunk_threshold=0.3, shortcut_mode='8,16,16', distill=True, kv_offload=True, enable_cuda_graph=False))
[2026-05-20 10:33:28,530 - INFO] Precompute validation prompt embeddings
You are using the default legacy behaviour of the <class 'transformers.models.t5.tokenization_t5.T5Tokenizer'>. This is expected, and simply means that the `legacy` (previous) behavior will be used so nothing changes for you. If you want to use the new behaviour, set `legacy=False`. This should only be set if you understand what it means, and thoroughly read the reason why this was added as explained in https://github.com/huggingface/transformers/pull/24565
KV cache compression is enabled.
Processing 72 samples.
[GPU 0] Assigned 72 samples
[GPU 0] Loading model...
[GPU 0] Model loaded.
[GPU 0] Generating T2V: 'a person swimming in ocean' -> /home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/outputs/vbench/videos/subject_consistency/a person swimming in ocean-0.mp4

Loading checkpoint shards:   0%|          | 0/2 [00:00<?, ?it/s]
Loading checkpoint shards:  50%|█████     | 1/2 [00:10<00:10, 10.92s/it]
Loading checkpoint shards: 100%|██████████| 2/2 [00:19<00:00,  9.50s/it]
Loading checkpoint shards: 100%|██████████| 2/2 [00:19<00:00,  9.71s/it]
[2026-05-20 10:33:49,537 - INFO] (cp, pp) rank (0, 0): param count 4459898128, model size 8.34 GB
[2026-05-20 10:33:49,537 - INFO] Build DiTModel successfully
[2026-05-20 10:33:49,538 - INFO] After build_dit_model, memory allocated: 0.04 GB, memory reserved: 0.08 GB
[2026-05-20 10:33:49,538 - INFO] load inference_weight.distill weight from ./downloads/4.5B_distill/inference_weight.distill

Loading shards:   0%|          | 0/2 [00:00<?, ?it/s]
Loading shards: 100%|██████████| 2/2 [00:00<00:00, 124.40it/s]
[2026-05-20 10:33:51,253 - INFO] Load Weight Missing Keys: []
[2026-05-20 10:33:51,253 - INFO] Load Weight Unexpected Keys: []
[2026-05-20 10:33:51,770 - INFO] After load_checkpoint, memory allocated: 8.39 GB, memory reserved: 8.40 GB
[2026-05-20 10:33:51,773 - INFO] After high_precision_promoter, memory allocated: 8.39 GB, memory reserved: 8.40 GB
[2026-05-20 10:33:51,875 - INFO] Load checkpoint successfully
[2026-05-20 10:33:51,875 - INFO] Begin to generate per chunk
[2026-05-20 10:33:51,875 - INFO] special_token = ['HQ_TOKEN', 'DURATION_TOKEN']

InferBatch 0:   0%|          | 0/10 [00:00<?, ?it/s][2026-05-20 10:33:51,900 - INFO] transport_inputs len: 1
Process Process-1:
Traceback (most recent call last):
  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/envs/magi/lib/python3.10/multiprocessing/process.py", line 314, in _bootstrap
    self.run()
  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/envs/magi/lib/python3.10/multiprocessing/process.py", line 108, in run
    self._target(*self._args, **self._kwargs)
  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 259, in worker_process
    process_func(pipeline, sample, config, gpu_id)
  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 185, in process_vbench_sample
    pipeline.run_text_to_video(prompt=prompt, output_path=output_path)
  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/pipeline.py", line 38, in run_text_to_video
    self._run(prompt, None, output_path)
  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/pipeline.py", line 52, in _run
    [
  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/pipeline.py", line 52, in <listcomp>
    [
  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/video_generate.py", line 1129, in generate_per_chunk
    for _, _, chunk in sample_transport.walk():
  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/video_generate.py", line 1079, in walk
    velocity = self.forward_velocity(infer_idx, 0)
  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/flowcache.py", line 93, in flowcache_forward_velocity
    cache = SampleTransport.cache_reuse_manager
AttributeError: type object 'SampleTransport' has no attribute 'cache_reuse_manager'

InferBatch 0:   0%|          | 0/10 [00:00<?, ?it/s]
✅ Completed: subject_consistency
---
🔍 Processing dimension: scene
Loaded configuration from: yaml_config/sample/flowcache_vbench.yaml.tmp
Total samples: 86
GPUs: [0]
Output: outputs/vbench/videos/scene
Config: config/sample/vbench.json
/home/dyvm6xra/dyvm6xrauser11/miniforge3/envs/magi/lib/python3.10/site-packages/timm/models/layers/__init__.py:48: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
[W520 10:33:58.467950462 CUDAAllocatorConfig.h:28] Warning: expandable_segments not supported on this platform (function operator())
[2026-05-20 10:33:58,023 - INFO] Initialize torch distribution and model parallel successfully
[2026-05-20 10:33:58,023 - INFO] MagiConfig(model_config=ModelConfig(model_name='videodit_ardf', num_layers=34, hidden_size=3072, ffn_hidden_size=12288, num_attention_heads=24, num_query_groups=8, kv_channels=128, layernorm_epsilon=1e-06, apply_layernorm_1p=True, x_rescale_factor=1, half_channel_vae=False, params_dtype=torch.bfloat16, patch_size=2, t_patch_size=1, in_channels=16, out_channels=16, cond_hidden_ratio=0.25, caption_channels=4096, caption_max_length=800, xattn_cond_hidden_ratio=1.0, cond_gating_ratio=1.0, gated_linear_unit=False), runtime_config=RuntimeConfig(cfg_number=1, cfg_t_range=[0.0, 0.0217, 0.1, 0.3, 0.999], prev_chunk_scales=[1.5, 1.5, 1.5, 1.0, 1.0], text_scales=[7.5, 7.5, 7.5, 0.0, 0.0], noise2clean_kvrange=[], clean_chunk_kvrange=1, clean_t=0.9999, seed=1234, num_frames=240, video_size_h=720, video_size_w=720, num_steps=16, window_size=4, fps=24, chunk_width=6, t5_pretrained='./downloads/t5_pretrained', t5_device='cuda', vae_pretrained='./downloads/vae', scale_factor=0.18215, temporal_downsample_factor=4, load='./downloads/4.5B_distill'), engine_config=EngineConfig(distributed_backend='nccl', distributed_timeout_minutes=15, pp_size=1, cp_size=1, cp_strategy='none', ulysses_overlap_degree=1, fp8_quant=False, distill_nearly_clean_chunk_threshold=0.3, shortcut_mode='8,16,16', distill=True, kv_offload=True, enable_cuda_graph=False))
[2026-05-20 10:33:58,023 - INFO] Precompute validation prompt embeddings
You are using the default legacy behaviour of the <class 'transformers.models.t5.tokenization_t5.T5Tokenizer'>. This is expected, and simply means that the `legacy` (previous) behavior will be used so nothing changes for you. If you want to use the new behaviour, set `legacy=False`. This should only be set if you understand what it means, and thoroughly read the reason why this was added as explained in https://github.com/huggingface/transformers/pull/24565
KV cache compression is enabled.
Processing 86 samples.
[GPU 0] Assigned 86 samples
[GPU 0] Loading model...
[GPU 0] Model loaded.
[GPU 0] Generating T2V: 'alley' -> /home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/outputs/vbench/videos/scene/alley-0.mp4

Loading checkpoint shards:   0%|          | 0/2 [00:00<?, ?it/s]
Loading checkpoint shards:  50%|█████     | 1/2 [00:09<00:09,  9.39s/it]
Loading checkpoint shards: 100%|██████████| 2/2 [00:17<00:00,  8.74s/it]
Loading checkpoint shards: 100%|██████████| 2/2 [00:17<00:00,  8.83s/it]
[2026-05-20 10:34:17,215 - INFO] (cp, pp) rank (0, 0): param count 4459898128, model size 8.34 GB
[2026-05-20 10:34:17,216 - INFO] Build DiTModel successfully
[2026-05-20 10:34:17,216 - INFO] After build_dit_model, memory allocated: 0.04 GB, memory reserved: 0.08 GB
[2026-05-20 10:34:17,216 - INFO] load inference_weight.distill weight from ./downloads/4.5B_distill/inference_weight.distill

Loading shards:   0%|          | 0/2 [00:00<?, ?it/s]
Loading shards: 100%|██████████| 2/2 [00:00<00:00, 135.76it/s]
[2026-05-20 10:34:18,927 - INFO] Load Weight Missing Keys: []
[2026-05-20 10:34:18,927 - INFO] Load Weight Unexpected Keys: []
[2026-05-20 10:34:19,154 - INFO] After load_checkpoint, memory allocated: 8.39 GB, memory reserved: 8.40 GB
[2026-05-20 10:34:19,156 - INFO] After high_precision_promoter, memory allocated: 8.39 GB, memory reserved: 8.40 GB
[2026-05-20 10:34:19,251 - INFO] Load checkpoint successfully
[2026-05-20 10:34:19,251 - INFO] Begin to generate per chunk
[2026-05-20 10:34:19,252 - INFO] special_token = ['HQ_TOKEN', 'DURATION_TOKEN']

InferBatch 0:   0%|          | 0/10 [00:00<?, ?it/s][2026-05-20 10:34:19,276 - INFO] transport_inputs len: 1
Process Process-1:
Traceback (most recent call last):
  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/envs/magi/lib/python3.10/multiprocessing/process.py", line 314, in _bootstrap
    self.run()
  File "/home/dyvm6xra/dyvm6xrauser11/miniforge3/envs/magi/lib/python3.10/multiprocessing/process.py", line 108, in run
    self._target(*self._args, **self._kwargs)
  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 259, in worker_process
    process_func(pipeline, sample, config, gpu_id)
  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/sample_video.py", line 185, in process_vbench_sample
    pipeline.run_text_to_video(prompt=prompt, output_path=output_path)
  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/pipeline.py", line 38, in run_text_to_video
    self._run(prompt, None, output_path)
  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/pipeline.py", line 52, in _run
    [
  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/pipeline.py", line 52, in <listcomp>
    [
  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/video_generate.py", line 1129, in generate_per_chunk
    for _, _, chunk in sample_transport.walk():
  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/video_generate.py", line 1079, in walk
    velocity = self.forward_velocity(infer_idx, 0)
  File "/home/dyvm6xra/dyvm6xrauser11/workspace/cz/FlowCache/FlowCache4MAGI-1/inference/pipeline/flowcache.py", line 93, in flowcache_forward_velocity
    cache = SampleTransport.cache_reuse_manager
AttributeError: type object 'SampleTransport' has no attribute 'cache_reuse_manager'

InferBatch 0:   0%|          | 0/10 [00:00<?, ?it/s]
✅ Completed: scene
---
🎉 All sampling tasks completed.