Add files using upload-large-folder tool

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/logs/vbench2_parallel/Dynamic_Attribute.log

@@ -0,0 +1,31 @@
+[2026-06-22 18:35:43 UTC] COMMAND /usr/local/bin/python -u evaluate.py --videos_path /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/videos/Dynamic_Attribute --dimension Dynamic_Attribute --output_path /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Dynamic_Attribute --full_json_dir /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_selected_info.json --mode vbench_standard
+args: Namespace(output_path='/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Dynamic_Attribute', full_json_dir='/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_selected_info.json', videos_path='/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/videos/Dynamic_Attribute', dimension=['Dynamic_Attribute'], load_ckpt_from_local=None, read_frame=None, mode='vbench_standard', prompt='None', prompt_file=None, category=None, imaging_quality_preprocessing_mode='longer')
+start evaluation
+File exists: /root/ckpts/vbench2/lmms-lab/LLaVA-Video-7B-Qwen2
+Evaluation meta data saved to /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Dynamic_Attribute/results_2026-06-22-18:35:46_full_info.json
+/usr/local/lib/python3.12/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+2026-06-22 18:35:49.047638: E external/local_xla/xla/stream_executor/cuda/cuda_fft.cc:467] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered
+WARNING: All log messages before absl::InitializeLog() is called are written to STDERR
+E0000 00:00:1782153349.067460 3063110 cuda_dnn.cc:8579] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered
+E0000 00:00:1782153349.074275 3063110 cuda_blas.cc:1407] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered
+W0000 00:00:1782153349.092817 3063110 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.
+W0000 00:00:1782153349.092880 3063110 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.
+W0000 00:00:1782153349.092889 3063110 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.
+W0000 00:00:1782153349.092896 3063110 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.
+2026-06-22 18:35:49.097227: I tensorflow/core/platform/cpu_feature_guard.cc:210] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+/nfs/ywang29/Reward_finetuning/VideoX-Fun/VBench/VBench-2.0/vbench2/third_party/LLaVA_NeXT/llava/model/llava_arch.py:215: SyntaxWarning: "is not" with 'int' literal. Did you mean "!="?
+  if slower_img_feat is not 0:
+cur_full_info_path: /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Dynamic_Attribute/results_2026-06-22-18:35:46_full_info.json
+Loaded LLaVA model: /root/ckpts/vbench2/lmms-lab/LLaVA-Video-7B-Qwen2
+You are using a model of type llava to instantiate a model of type llava_qwen. This is not supported for all configurations of models and can yield errors.
+Sliding Window Attention is enabled but not implemented for `sdpa`; unexpected results may be encountered.
+Loading vision tower: google/siglip-so400m-patch14-384
+2026-06-22 18:35:55,625 - accelerate.utils.modeling - INFO - We will use 90% of the memory on device 0 for storing the model, and 10% for the buffer to avoid OOM. You can set `max_memory` in to a higher value to use more memory (at your own risk).
+
+Model Class: LlavaQwenForCausalLM
+
  0%|          | 0/91 [00:00<?, ?it/s]The attention mask is not set and cannot be inferred from input because pad token is same as eos token. As a consequence, you may observe unexpected behavior. Please pass your input's `attention_mask` to obtain reliable results.
+
  1%|          | 1/91 [00:22<34:19, 22.88s/it]
  2%|▏         | 2/91 [00:45<33:35, 22.64s/it]
  3%|▎         | 3/91 [01:07<33:11, 22.63s/it]
  4%|▍         | 4/91 [01:31<33:04, 22.81s/it]
  5%|▌         | 5/91 [01:53<32:34, 22.73s/it]
  7%|▋         | 6/91 [02:16<32:07, 22.67s/it]
  8%|▊         | 7/91 [02:38<31:39, 22.61s/it]
  9%|▉         | 8/91 [03:00<31:05, 22.47s/it]
 10%|▉         | 9/91 [03:23<30:41, 22.46s/it]
 11%|█         | 10/91 [03:45<30:21, 22.49s/it]
 12%|█▏        | 11/91 [04:08<29:58, 22.48s/it]
 13%|█▎        | 12/91 [04:30<29:34, 22.46s/it]
 14%|█▍        | 13/91 [04:53<29:30, 22.70s/it]
 15%|█▌        | 14/91 [05:16<28:58, 22.58s/it]
 16%|█▋        | 15/91 [05:38<28:28, 22.48s/it]
 18%|█▊        | 16/91 [06:00<28:03, 22.45s/it]
 19%|█▊        | 17/91 [06:24<28:04, 22.76s/it]
 20%|█▉        | 18/91 [06:46<27:32, 22.63s/it]
 21%|██        | 19/91 [07:09<27:05, 22.58s/it]
 22%|██▏       | 20/91 [07:32<26:49, 22.67s/it]
 23%|██▎       | 21/91 [07:55<26:35, 22.79s/it]
 24%|██▍       | 22/91 [08:17<26:07, 22.71s/it]
 25%|██▌       | 23/91 [08:40<25:43, 22.70s/it]
 26%|██▋       | 24/91 [09:03<25:21, 22.70s/it]
 27%|██▋       | 25/91 [09:25<24:53, 22.63s/it]
 29%|██▊       | 26/91 [09:47<24:21, 22.49s/it]
 30%|██▉       | 27/91 [10:09<23:53, 22.40s/it]
 31%|███       | 28/91 [10:32<23:33, 22.44s/it]
 32%|███▏      | 29/91 [10:55<23:15, 22.50s/it]
 33%|███▎      | 30/91 [11:17<22:54, 22.53s/it]
 34%|███▍      | 31/91 [11:40<22:40, 22.67s/it]
 35%|███▌      | 32/91 [11:49<18:21, 18.67s/it]
 36%|███▋      | 33/91 [11:59<15:21, 15.88s/it]
 37%|███▋      | 34/91 [12:08<13:10, 13.87s/it]
 38%|███▊      | 35/91 [12:17<11:41, 12.52s/it]
 40%|███▉      | 36/91 [12:27<10:37, 11.59s/it]
 41%|████      | 37/91 [12:36<09:46, 10.86s/it]
 42%|████▏     | 38/91 [12:46<09:21, 10.59s/it]
 43%|████▎     | 39/91 [12:55<08:51, 10.21s/it]
 44%|████▍     | 40/91 [13:05<08:29,  9.98s/it]
 45%|████▌     | 41/91 [13:14<08:09,  9.78s/it]
 46%|████▌     | 42/91 [13:24<07:55,  9.70s/it]
 47%|████▋     | 43/91 [13:33<07:36,  9.52s/it]
 48%|████▊     | 44/91 [13:42<07:26,  9.51s/it]
 49%|████▉     | 45/91 [13:52<07:16,  9.50s/it]
 51%|█████     | 46/91 [14:01<07:12,  9.62s/it]
 52%|█████▏    | 47/91 [14:11<07:03,  9.62s/it]
 53%|█████▎    | 48/91 [14:20<06:49,  9.53s/it]
 54%|█████▍    | 49/91 [14:30<06:39,  9.52s/it]
 55%|█████▍    | 50/91 [14:39<06:28,  9.48s/it]
 56%|█████▌    | 51/91 [14:49<06:24,  9.62s/it]
 57%|█████▋    | 52/91 [14:59<06:11,  9.53s/it]
 58%|█████▊    | 53/91 [15:08<06:06,  9.65s/it]
 59%|█████▉    | 54/91 [15:18<05:54,  9.58s/it]
 60%|██████    | 55/91 [15:28<05:50,  9.74s/it]
 62%|██████▏   | 56/91 [15:38<05:39,  9.69s/it]
 63%|██████▎   | 57/91 [15:47<05:27,  9.62s/it]
 64%|██████▎   | 58/91 [15:57<05:22,  9.78s/it]
 65%|██████▍   | 59/91 [16:07<05:11,  9.72s/it]
 66%|██████▌   | 60/91 [16:17<05:01,  9.73s/it]
 67%|██████▋   | 61/91 [16:26<04:48,  9.62s/it]
 68%|██████▊   | 62/91 [16:35<04:37,  9.57s/it]
 69%|██████▉   | 63/91 [16:45<04:32,  9.72s/it]
 70%|███████   | 64/91 [16:55<04:20,  9.64s/it]
 71%|███████▏  | 65/91 [17:04<04:09,  9.60s/it]
 73%|███████▎  | 66/91 [17:14<04:02,  9.72s/it]
 74%|███████▎  | 67/91 [17:24<03:54,  9.78s/it]
 75%|███████▍  | 68/91 [17:47<05:13, 13.63s/it]
 76%|███████▌  | 69/91 [18:09<05:56, 16.21s/it]
 77%|███████▋  | 70/91 [18:31<06:18, 18.02s/it]
 78%|███████▊  | 71/91 [18:54<06:25, 19.30s/it]
 79%|███████▉  | 72/91 [19:16<06:23, 20.19s/it]
 80%|████████  | 73/91 [19:38<06:14, 20.82s/it]
 81%|████████▏ | 74/91 [20:01<06:01, 21.27s/it]
 82%|████████▏ | 75/91 [20:23<05:46, 21.68s/it]
 84%|████████▎ | 76/91 [20:46<05:28, 21.88s/it]
 85%|████████▍ | 77/91 [21:08<05:09, 22.07s/it]
 86%|████████▌ | 78/91 [21:30<04:48, 22.17s/it]
 87%|████████▋ | 79/91 [21:53<04:27, 22.32s/it]
 88%|████████▊ | 80/91 [22:15<04:05, 22.30s/it]
 89%|████████▉ | 81/91 [22:38<03:44, 22.47s/it]
 90%|█████████ | 82/91 [23:01<03:21, 22.41s/it]
 91%|█████████ | 83/91 [23:23<02:59, 22.46s/it]
 92%|█████████▏| 84/91 [23:46<02:37, 22.45s/it]
 93%|█████████▎| 85/91 [24:08<02:14, 22.42s/it]
 95%|█████████▍| 86/91 [24:30<01:51, 22.40s/it]
 96%|█████████▌| 87/91 [24:52<01:29, 22.36s/it]
 97%|█████████▋| 88/91 [25:15<01:07, 22.48s/it]
 98%|█████████▊| 89/91 [25:37<00:44, 22.41s/it]
 99%|█████████▉| 90/91 [26:00<00:22, 22.48s/it]
+Evaluation results saved to /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Dynamic_Attribute/results_2026-06-22-18:35:46_eval_results.json
+done

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/logs/vbench2_parallel/Dynamic_Spatial_Relationship.log

@@ -0,0 +1,31 @@
+[2026-06-22 19:03:13 UTC] COMMAND /usr/local/bin/python -u evaluate.py --videos_path /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/videos/Dynamic_Spatial_Relationship --dimension Dynamic_Spatial_Relationship --output_path /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Dynamic_Spatial_Relationship --full_json_dir /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_selected_info.json --mode vbench_standard
+args: Namespace(output_path='/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Dynamic_Spatial_Relationship', full_json_dir='/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_selected_info.json', videos_path='/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/videos/Dynamic_Spatial_Relationship', dimension=['Dynamic_Spatial_Relationship'], load_ckpt_from_local=None, read_frame=None, mode='vbench_standard', prompt='None', prompt_file=None, category=None, imaging_quality_preprocessing_mode='longer')
+start evaluation
+File exists: /root/ckpts/vbench2/lmms-lab/LLaVA-Video-7B-Qwen2
+Evaluation meta data saved to /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Dynamic_Spatial_Relationship/results_2026-06-22-19:03:17_full_info.json
+/usr/local/lib/python3.12/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+2026-06-22 19:03:19.835619: E external/local_xla/xla/stream_executor/cuda/cuda_fft.cc:467] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered
+WARNING: All log messages before absl::InitializeLog() is called are written to STDERR
+E0000 00:00:1782154999.857188 3076605 cuda_dnn.cc:8579] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered
+E0000 00:00:1782154999.864156 3076605 cuda_blas.cc:1407] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered
+W0000 00:00:1782154999.888978 3076605 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.
+W0000 00:00:1782154999.889051 3076605 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.
+W0000 00:00:1782154999.889062 3076605 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.
+W0000 00:00:1782154999.889071 3076605 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.
+2026-06-22 19:03:19.893790: I tensorflow/core/platform/cpu_feature_guard.cc:210] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+/nfs/ywang29/Reward_finetuning/VideoX-Fun/VBench/VBench-2.0/vbench2/third_party/LLaVA_NeXT/llava/model/llava_arch.py:215: SyntaxWarning: "is not" with 'int' literal. Did you mean "!="?
+  if slower_img_feat is not 0:
+cur_full_info_path: /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Dynamic_Spatial_Relationship/results_2026-06-22-19:03:17_full_info.json
+Loaded LLaVA model: /root/ckpts/vbench2/lmms-lab/LLaVA-Video-7B-Qwen2
+You are using a model of type llava to instantiate a model of type llava_qwen. This is not supported for all configurations of models and can yield errors.
+Sliding Window Attention is enabled but not implemented for `sdpa`; unexpected results may be encountered.
+Loading vision tower: google/siglip-so400m-patch14-384
+2026-06-22 19:03:27,319 - accelerate.utils.modeling - INFO - We will use 90% of the memory on device 0 for storing the model, and 10% for the buffer to avoid OOM. You can set `max_memory` in to a higher value to use more memory (at your own risk).
+
+Model Class: LlavaQwenForCausalLM
+
  0%|          | 0/69 [00:00<?, ?it/s]The attention mask is not set and cannot be inferred from input because pad token is same as eos token. As a consequence, you may observe unexpected behavior. Please pass your input's `attention_mask` to obtain reliable results.
+
  1%|▏         | 1/69 [00:07<08:35,  7.57s/it]
  3%|▎         | 2/69 [00:13<07:41,  6.89s/it]
  4%|▍         | 3/69 [00:20<07:24,  6.73s/it]
  6%|▌         | 4/69 [00:27<07:12,  6.65s/it]
  7%|▋         | 5/69 [00:33<07:05,  6.64s/it]
  9%|▊         | 6/69 [00:40<06:56,  6.62s/it]
 10%|█         | 7/69 [00:46<06:48,  6.59s/it]
 12%|█▏        | 8/69 [00:53<06:41,  6.58s/it]
 13%|█▎        | 9/69 [00:59<06:32,  6.54s/it]
 14%|█▍        | 10/69 [01:06<06:31,  6.63s/it]
 16%|█▌        | 11/69 [01:13<06:20,  6.55s/it]
 17%|█▋        | 12/69 [01:19<06:10,  6.50s/it]
 19%|█▉        | 13/69 [01:25<06:02,  6.47s/it]
 20%|██        | 14/69 [01:32<05:55,  6.46s/it]
 22%|██▏       | 15/69 [01:38<05:48,  6.45s/it]
 23%|██▎       | 16/69 [01:45<05:42,  6.47s/it]
 25%|██▍       | 17/69 [01:51<05:40,  6.55s/it]
 26%|██▌       | 18/69 [01:58<05:31,  6.50s/it]
 28%|██▊       | 19/69 [02:04<05:26,  6.52s/it]
 29%|██▉       | 20/69 [02:11<05:17,  6.47s/it]
 30%|███       | 21/69 [02:17<05:09,  6.44s/it]
 32%|███▏      | 22/69 [02:24<05:07,  6.54s/it]
 33%|███▎      | 23/69 [02:31<05:03,  6.60s/it]
 35%|███▍      | 24/69 [02:37<04:54,  6.55s/it]
 36%|███▌      | 25/69 [02:44<04:50,  6.60s/it]
 38%|███▊      | 26/69 [02:50<04:41,  6.56s/it]
 39%|███▉      | 27/69 [02:57<04:35,  6.56s/it]
 41%|████      | 28/69 [03:03<04:29,  6.56s/it]
 42%|████▏     | 29/69 [03:10<04:21,  6.53s/it]
 43%|████▎     | 30/69 [03:16<04:16,  6.57s/it]
 45%|████▍     | 31/69 [03:23<04:10,  6.59s/it]
 46%|████▋     | 32/69 [03:30<04:08,  6.71s/it]
 48%|████▊     | 33/69 [03:37<04:01,  6.72s/it]
 49%|████▉     | 34/69 [03:44<03:55,  6.72s/it]
 51%|█████     | 35/69 [03:50<03:48,  6.71s/it]
 52%|█████▏    | 36/69 [03:57<03:38,  6.63s/it]
 54%|█████▎    | 37/69 [04:03<03:29,  6.54s/it]
 55%|█████▌    | 38/69 [04:09<03:21,  6.49s/it]
 57%|█████▋    | 39/69 [04:16<03:14,  6.48s/it]
 58%|█████▊    | 40/69 [04:22<03:08,  6.49s/it]
 59%|█████▉    | 41/69 [04:29<03:02,  6.53s/it]
 61%|██████    | 42/69 [04:35<02:56,  6.52s/it]
 62%|██████▏   | 43/69 [04:42<02:47,  6.46s/it]
 64%|██████▍   | 44/69 [04:48<02:41,  6.46s/it]
 65%|██████▌   | 45/69 [04:55<02:37,  6.56s/it]
 67%|██████▋   | 46/69 [05:02<02:30,  6.56s/it]
 68%|██████▊   | 47/69 [05:08<02:25,  6.60s/it]
 70%|██████▉   | 48/69 [05:15<02:18,  6.57s/it]
 71%|███████   | 49/69 [05:21<02:09,  6.50s/it]
 72%|███████▏  | 50/69 [05:28<02:03,  6.48s/it]
 74%|███████▍  | 51/69 [05:34<01:55,  6.42s/it]
 75%|███████▌  | 52/69 [05:40<01:49,  6.46s/it]
 77%|███████▋  | 53/69 [05:47<01:43,  6.50s/it]
 78%|███████▊  | 54/69 [05:53<01:37,  6.48s/it]
 80%|███████▉  | 55/69 [06:00<01:30,  6.45s/it]
 81%|████████  | 56/69 [06:06<01:23,  6.44s/it]
 83%|████████▎ | 57/69 [06:13<01:17,  6.46s/it]
 84%|████████▍ | 58/69 [06:19<01:11,  6.47s/it]
 86%|████████▌ | 59/69 [06:26<01:04,  6.48s/it]
 87%|████████▋ | 60/69 [06:32<00:58,  6.50s/it]
 88%|████████▊ | 61/69 [06:39<00:51,  6.44s/it]
 90%|████████▉ | 62/69 [06:45<00:45,  6.49s/it]
 91%|█████████▏| 63/69 [06:51<00:38,  6.43s/it]
 93%|█████████▎| 64/69 [06:58<00:32,  6.54s/it]
 94%|█████████▍| 65/69 [07:05<00:26,  6.55s/it]
 96%|█████████▌| 66/69 [07:12<00:19,  6.59s/it]
 97%|█████████▋| 67/69 [07:18<00:13,  6.63s/it]
 99%|█████████▊| 68/69 [07:25<00:06,  6.61s/it]
+Evaluation results saved to /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Dynamic_Spatial_Relationship/results_2026-06-22-19:03:17_eval_results.json
+done

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/logs/vbench2_parallel/Human_Anatomy_shard_1.log

@@ -0,0 +1,34 @@
+[2026-06-22 21:20:21 UTC] COMMAND /usr/local/bin/python -u evaluate.py --videos_path /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/videos/Human_Anatomy --dimension Human_Anatomy --output_path /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Human_Anatomy_shards/2026-06-22-21:20:21/shard_1 --full_json_dir /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Human_Anatomy_shards/2026-06-22-21:20:21/full_info_shard_1.json --mode vbench_standard
+args: Namespace(output_path='/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Human_Anatomy_shards/2026-06-22-21:20:21/shard_1', full_json_dir='/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Human_Anatomy_shards/2026-06-22-21:20:21/full_info_shard_1.json', videos_path='/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/videos/Human_Anatomy', dimension=['Human_Anatomy'], load_ckpt_from_local=None, read_frame=None, mode='vbench_standard', prompt='None', prompt_file=None, category=None, imaging_quality_preprocessing_mode='longer')
+start evaluation
+Evaluation meta data saved to /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Human_Anatomy_shards/2026-06-22-21:20:21/shard_1/results_2026-06-22-21:20:24_full_info.json
+/usr/local/lib/python3.12/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+06/22 21:20:28 - mmengine - INFO - Adafactor is already registered in optimizer at torch.optim. Registration ignored.
+/usr/local/lib/python3.12/site-packages/imagecorruptions/corruptions.py:17: UserWarning: pkg_resources is deprecated as an API. See https://setuptools.pypa.io/en/latest/pkg_resources.html. The pkg_resources package is slated for removal as early as 2025-11-30. Refrain from using this package or pin to Setuptools<81.
+  from pkg_resources import resource_filename
+cur_full_info_path: /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Human_Anatomy_shards/2026-06-22-21:20:21/shard_1/results_2026-06-22-21:20:24_full_info.json
+2026-06-22 21:20:31.981327: E external/local_xla/xla/stream_executor/cuda/cuda_fft.cc:467] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered
+WARNING: All log messages before absl::InitializeLog() is called are written to STDERR
+E0000 00:00:1782163232.000442 3196082 cuda_dnn.cc:8579] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered
+E0000 00:00:1782163232.007091 3196082 cuda_blas.cc:1407] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered
+W0000 00:00:1782163232.024756 3196082 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.
+W0000 00:00:1782163232.024798 3196082 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.
+W0000 00:00:1782163232.024807 3196082 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.
+W0000 00:00:1782163232.024815 3196082 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.
+Loads checkpoint by local backend from path: /root/ckpts/vbench2/YOLO-World/yolo_world_v2_xl_obj365v1_goldg_cc3mlite_pretrain-5daf1395.pth
+2026-06-22 21:20:36,874 - lvis.lvis - INFO - Loading annotations.
+2026-06-22 21:20:37,605 - lvis.lvis - INFO - Creating index.
+2026-06-22 21:20:37,640 - lvis.lvis - INFO - Index created.
+Loads checkpoint by local backend from path: /root/ckpts/vbench2/YOLO-World/yolo_world_v2_xl_obj365v1_goldg_cc3mlite_pretrain-5daf1395.pth
+2026-06-22 21:20:42,084 - lvis.lvis - INFO - Loading annotations.
+2026-06-22 21:20:42,835 - lvis.lvis - INFO - Creating index.
+2026-06-22 21:20:42,942 - lvis.lvis - INFO - Index created.
+=> merge config from vbench2/third_party/ViTDetector/simmim_finetune__vit_base__img224__800ep.yaml
+/usr/local/lib/python3.12/site-packages/torch/functional.py:505: UserWarning: torch.meshgrid: in an upcoming release, it will be required to pass the indexing argument. (Triggered internally at /pytorch/aten/src/ATen/native/TensorShape.cpp:4317.)
+  return _VF.meshgrid(tensors, **kwargs)  # type: ignore[attr-defined]
+=> merge config from vbench2/third_party/ViTDetector/simmim_finetune__vit_base__img224__800ep.yaml
+=> merge config from vbench2/third_party/ViTDetector/simmim_finetune__vit_base__img224__800ep.yaml
+
  0%|          | 0/180 [00:00<?, ?it/s]
  1%|          | 1/180 [02:08<6:24:27, 128.87s/it]
  1%|          | 2/180 [04:12<6:12:42, 125.63s/it]
  2%|▏         | 3/180 [05:35<5:13:03, 106.12s/it]
  2%|▏         | 4/180 [07:02<4:49:58, 98.85s/it] 
  3%|▎         | 5/180 [08:38<4:45:08, 97.76s/it]
  3%|▎         | 6/180 [10:19<4:46:09, 98.68s/it]
  4%|▍         | 7/180 [12:33<5:18:16, 110.38s/it]
  4%|▍         | 8/180 [16:58<7:37:33, 159.61s/it]
  5%|▌         | 9/180 [19:44<7:40:45, 161.67s/it]
  6%|▌         | 10/180 [20:58<6:20:43, 134.37s/it]
  6%|▌         | 11/180 [22:08<5:23:45, 114.95s/it]
  7%|▋         | 12/180 [23:12<4:38:15, 99.38s/it] 
  7%|▋         | 13/180 [25:00<4:44:00, 102.04s/it]
  8%|▊         | 14/180 [26:39<4:39:08, 100.89s/it]
  8%|▊         | 15/180 [28:19<4:37:14, 100.81s/it]
  9%|▉         | 16/180 [32:07<6:19:35, 138.88s/it]
  9%|▉         | 17/180 [35:06<6:50:23, 151.06s/it]
 10%|█         | 18/180 [38:20<7:22:29, 163.88s/it]
 11%|█         | 19/180 [39:56<6:25:40, 143.73s/it]
 11%|█         | 20/180 [41:39<5:50:40, 131.51s/it]
 12%|█▏        | 21/180 [43:14<5:18:51, 120.32s/it]
 12%|█▏        | 22/180 [44:41<4:50:24, 110.28s/it]
 13%|█▎        | 23/180 [46:11<4:33:02, 104.35s/it]
 13%|█▎        | 24/180 [47:27<4:09:10, 95.83s/it] 
 14%|█▍        | 25/180 [50:29<5:13:54, 121.52s/it]
 14%|█▍        | 26/180 [52:06<4:53:45, 114.45s/it]
 15%|█▌        | 27/180 [54:12<5:00:20, 117.78s/it]
 16%|█▌        | 28/180 [56:15<5:02:29, 119.41s/it]
 16%|█▌        | 29/180 [57:31<4:27:51, 106.44s/it]
 17%|█▋        | 30/180 [58:42<3:58:54, 95.56s/it] 
 17%|█▋        | 31/180 [1:00:40<4:14:18, 102.41s/it]
 18%|█▊        | 32/180 [1:02:27<4:16:04, 103.81s/it]
 18%|█▊        | 33/180 [1:04:20<4:21:22, 106.69s/it]
 19%|█▉        | 34/180 [1:06:13<4:23:46, 108.40s/it]
 19%|█▉        | 35/180 [1:09:45<5:36:58, 139.44s/it]
 20%|██        | 36/180 [1:10:44<4:37:17, 115.54s/it]
 21%|██        | 37/180 [1:12:23<4:23:22, 110.50s/it]
 21%|██        | 38/180 [1:17:00<6:19:26, 160.33s/it]
 22%|██▏       | 39/180 [1:19:25<6:05:58, 155.73s/it]
 22%|██▏       | 40/180 [1:20:54<5:16:47, 135.77s/it]
 23%|██▎       | 41/180 [1:22:47<4:58:55, 129.03s/it]
 23%|██▎       | 42/180 [1:24:11<4:25:07, 115.27s/it]
 24%|██▍       | 43/180 [1:25:35<4:02:19, 106.13s/it]
 24%|██▍       | 44/180 [1:27:16<3:56:48, 104.48s/it]
 25%|██▌       | 45/180 [1:29:05<3:57:54, 105.74s/it]
 26%|██▌       | 46/180 [1:34:28<6:22:07, 171.10s/it]
 26%|██▌       | 47/180 [1:37:39<6:32:15, 176.95s/it]
 27%|██▋       | 48/180 [1:40:56<6:42:40, 183.04s/it]
 27%|██▋       | 49/180 [1:42:57<5:59:02, 164.45s/it]
 28%|██▊       | 50/180 [1:44:46<5:20:29, 147.92s/it]
 28%|██▊       | 51/180 [1:46:06<4:34:01, 127.46s/it]
 29%|██▉       | 52/180 [1:47:38<4:09:18, 116.86s/it]
 29%|██▉       | 53/180 [1:48:59<3:44:22, 106.01s/it]
 30%|███       | 54/180 [1:50:40<3:39:35, 104.57s/it]
 31%|███       | 55/180 [1:51:26<3:00:55, 86.84s/it] 
 31%|███       | 56/180 [1:52:43<2:53:47, 84.09s/it]
 32%|███▏      | 57/180 [1:53:33<2:30:58, 73.64s/it]
 32%|███▏      | 58/180 [1:56:04<3:17:04, 96.92s/it]
 33%|███▎      | 59/180 [1:57:45<3:17:48, 98.09s/it]
 33%|███▎      | 60/180 [1:59:12<3:09:55, 94.97s/it]
 34%|███▍      | 61/180 [2:00:39<3:03:16, 92.41s/it]
 34%|███▍      | 62/180 [2:01:56<2:52:40, 87.80s/it]
 35%|███▌      | 63/180 [2:03:00<2:37:09, 80.59s/it]
 36%|███▌      | 64/180 [2:04:17<2:33:47, 79.55s/it]
 36%|███▌      | 65/180 [2:06:53<3:16:45, 102.66s/it]
 37%|███▋      | 66/180 [2:08:39<3:16:41, 103.53s/it]
 37%|███▋      | 67/180 [2:10:13<3:09:39, 100.70s/it]
 38%|███▊      | 68/180 [2:11:44<3:02:27, 97.75s/it] 
 38%|███▊      | 69/180 [2:12:14<2:23:12, 77.41s/it]
 39%|███▉      | 70/180 [2:13:15<2:13:13, 72.67s/it]
 39%|███▉      | 71/180 [2:14:19<2:06:56, 69.88s/it]
 40%|████      | 72/180 [2:15:28<2:05:12, 69.56s/it]
 41%|████      | 73/180 [2:16:29<1:59:48, 67.19s/it]
 41%|████      | 74/180 [2:17:38<1:59:41, 67.75s/it]
 42%|████▏     | 75/180 [2:18:50<2:00:51, 69.06s/it]
 42%|████▏     | 76/180 [2:20:10<2:04:54, 72.07s/it]
 43%|████▎     | 77/180 [2:21:39<2:12:48, 77.36s/it]
 43%|████▎     | 78/180 [2:22:46<2:06:08, 74.20s/it]
 44%|████▍     | 79/180 [2:24:24<2:17:00, 81.39s/it]
 44%|████▍     | 80/180 [2:26:29<2:37:24, 94.45s/it]
 45%|████▌     | 81/180 [2:27:37<2:22:40, 86.47s/it]
 46%|████▌     | 82/180 [2:28:28<2:03:43, 75.75s/it]
 46%|████▌     | 83/180 [2:29:39<2:00:25, 74.49s/it]
 47%|████▋     | 84/180 [2:30:51<1:57:38, 73.53s/it]
 47%|████▋     | 85/180 [2:31:56<1:52:43, 71.20s/it]
 48%|████▊     | 86/180 [2:33:04<1:49:46, 70.07s/it]
 48%|████▊     | 87/180 [2:34:25<1:53:40, 73.33s/it]
 49%|████▉     | 88/180 [2:36:20<2:11:52, 86.00s/it]
 49%|████▉     | 89/180 [2:38:17<2:24:30, 95.28s/it]
 50%|█████     | 90/180 [2:40:02<2:27:06, 98.07s/it]
 51%|█████     | 91/180 [2:41:43<2:26:52, 99.02s/it]
 51%|█████     | 92/180 [2:43:04<2:17:07, 93.50s/it]
 52%|█████▏    | 93/180 [2:44:33<2:13:42, 92.21s/it]
 52%|█████▏    | 94/180 [2:45:57<2:08:30, 89.65s/it]
 53%|█████▎    | 95/180 [2:46:52<1:52:39, 79.53s/it]
 53%|█████▎    | 96/180 [2:48:03<1:47:36, 76.86s/it]
 54%|█████▍    | 97/180 [2:49:29<1:50:14, 79.70s/it]
 54%|█████▍    | 98/180 [2:51:13<1:58:46, 86.91s/it]
 55%|█████▌    | 99/180 [2:53:00<2:05:27, 92.93s/it]
 56%|█████▌    | 100/180 [2:54:40<2:06:44, 95.06s/it]
 56%|█████▌    | 101/180 [2:56:28<2:10:19, 98.98s/it]
 57%|█████▋    | 102/180 [2:58:32<2:18:10, 106.28s/it]
 57%|█████▋    | 103/180 [2:59:54<2:07:10, 99.09s/it] 
 58%|█████▊    | 104/180 [3:01:11<1:57:10, 92.50s/it]
 58%|█████▊    | 105/180 [3:02:18<1:46:03, 84.85s/it]
 59%|█████▉    | 106/180 [3:04:03<1:52:08, 90.92s/it]
 59%|█████▉    | 107/180 [3:04:13<1:21:01, 66.59s/it]
 60%|██████    | 108/180 [3:05:34<1:24:57, 70.80s/it]
 61%|██████    | 109/180 [3:07:33<1:41:04, 85.41s/it]
 61%|██████    | 110/180 [3:08:55<1:38:33, 84.48s/it]
 62%|██████▏   | 111/180 [3:10:43<1:45:00, 91.31s/it]
 62%|██████▏   | 112/180 [3:12:22<1:46:04, 93.59s/it]
 63%|██████▎   | 113/180 [3:14:06<1:48:02, 96.75s/it]
 63%|██████▎   | 114/180 [3:15:34<1:43:38, 94.22s/it]
 64%|██████▍   | 115/180 [3:16:44<1:34:11, 86.95s/it]
 64%|██████▍   | 116/180 [3:17:49<1:25:50, 80.47s/it]
 65%|██████▌   | 117/180 [3:19:06<1:23:18, 79.34s/it]
 66%|██████▌   | 118/180 [3:20:52<1:30:11, 87.28s/it]
 66%|██████▌   | 119/180 [3:22:35<1:33:33, 92.02s/it]
 67%|██████▋   | 120/180 [3:24:09<1:32:37, 92.63s/it]
 67%|██████▋   | 121/180 [3:25:26<1:26:22, 87.84s/it]
 68%|██████▊   | 122/180 [3:26:28<1:17:30, 80.17s/it]
 68%|██████▊   | 123/180 [3:27:10<1:05:18, 68.75s/it]
 69%|██████▉   | 124/180 [3:28:23<1:05:30, 70.18s/it]
 69%|██████▉   | 125/180 [3:30:01<1:11:56, 78.48s/it]
 70%|███████   | 126/180 [3:31:23<1:11:30, 79.46s/it]
 71%|███████   | 127/180 [3:33:44<1:26:36, 98.04s/it]
 71%|███████   | 128/180 [3:35:23<1:25:09, 98.26s/it]
 72%|███████▏  | 129/180 [3:37:31<1:30:56, 107.00s/it]
 72%|███████▏  | 130/180 [3:38:41<1:20:03, 96.07s/it] 
 73%|███████▎  | 131/180 [3:41:38<1:38:15, 120.31s/it]
 73%|███████▎  | 132/180 [3:44:23<1:47:02, 133.80s/it]
 74%|███████▍  | 133/180 [3:45:43<1:32:00, 117.46s/it]
 74%|███████▍  | 134/180 [3:47:17<1:24:36, 110.37s/it]
 75%|███████▌  | 135/180 [3:48:21<1:12:33, 96.75s/it] 
 76%|███████▌  | 136/180 [3:50:53<1:22:53, 113.03s/it]
 76%|███████▌  | 137/180 [3:52:41<1:20:03, 111.70s/it]
 77%|███████▋  | 138/180 [3:55:12<1:26:29, 123.56s/it]
 77%|███████▋  | 139/180 [3:57:19<1:24:58, 124.36s/it]
 78%|███████▊  | 140/180 [4:00:53<1:40:58, 151.47s/it]
 78%|███████▊  | 141/180 [4:03:13<1:36:08, 147.92s/it]
 79%|███████▉  | 142/180 [4:06:35<1:43:58, 164.16s/it]
 79%|███████▉  | 143/180 [4:10:25<1:53:19, 183.77s/it]
 80%|████████  | 144/180 [4:14:10<1:57:46, 196.29s/it]
 81%|████████  | 145/180 [4:16:24<1:43:35, 177.57s/it]
 81%|████████  | 146/180 [4:17:58<1:26:26, 152.54s/it]
 82%|████████▏ | 147/180 [4:20:52<1:27:30, 159.10s/it]
 82%|████████▏ | 148/180 [4:23:16<1:22:26, 154.57s/it]
 83%|████████▎ | 149/180 [4:25:04<1:12:30, 140.34s/it]
 83%|████████▎ | 150/180 [4:26:32<1:02:24, 124.82s/it]
 84%|████████▍ | 151/180 [4:29:18<1:06:19, 137.23s/it]
 84%|████████▍ | 152/180 [4:30:47<57:13, 122.61s/it]  
 85%|████████▌ | 153/180 [4:33:28<1:00:18, 134.02s/it]
 86%|████████▌ | 154/180 [4:36:25<1:03:46, 147.16s/it]
 86%|████████▌ | 155/180 [4:38:21<57:21, 137.67s/it]  
 87%|████████▋ | 156/180 [4:40:49<56:18, 140.76s/it]
 87%|████████▋ | 157/180 [4:42:41<50:40, 132.18s/it]
 88%|████████▊ | 158/180 [4:43:47<41:09, 112.24s/it]
 88%|████████▊ | 159/180 [4:46:14<43:00, 122.89s/it]
 89%|████████▉ | 160/180 [4:49:17<46:55, 140.78s/it]
 89%|████████▉ | 161/180 [4:52:48<51:14, 161.80s/it]
 90%|█████████ | 162/180 [4:55:55<50:50, 169.49s/it]
 91%|█████████ | 163/180 [4:58:04<44:31, 157.13s/it]
 91%|█████████ | 164/180 [5:01:34<46:11, 173.22s/it]
 92%|█████████▏| 165/180 [5:05:33<48:13, 192.92s/it]
 92%|█████████▏| 166/180 [5:07:40<40:23, 173.12s/it]
 93%|█████████▎| 167/180 [5:12:14<44:04, 203.41s/it]
 93%|█████████▎| 168/180 [5:16:53<45:13, 226.10s/it]
 94%|█████████▍| 169/180 [5:19:57<39:07, 213.45s/it]
 94%|█████████▍| 170/180 [5:23:19<35:00, 210.10s/it]
 95%|█████████▌| 171/180 [5:25:01<26:36, 177.39s/it]
 96%|█████████▌| 172/180 [5:27:48<23:16, 174.52s/it]
 96%|█████████▌| 173/180 [5:30:47<20:29, 175.66s/it]
 97%|█████████▋| 174/180 [5:33:21<16:55, 169.24s/it]
 97%|█████████▋| 175/180 [5:36:35<14:43, 176.70s/it]
 98%|█████████▊| 176/180 [5:39:00<11:08, 167.14s/it]
 98%|█████████▊| 177/180 [5:41:25<08:01, 160.65s/it]
 99%|█████████▉| 178/180 [5:42:22<04:18, 129.36s/it]
 99%|█████████▉| 179/180 [5:43:04<01:43, 103.30s/it]
+Evaluation results saved to /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Human_Anatomy_shards/2026-06-22-21:20:21/shard_1/results_2026-06-22-21:20:24_eval_results.json
+done

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/logs/vbench2_parallel/Human_Clothes.log

@@ -0,0 +1,33 @@
+[2026-06-22 19:11:43 UTC] COMMAND /usr/local/bin/python -u evaluate.py --videos_path /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/videos/Human_Clothes --dimension Human_Clothes --output_path /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Human_Clothes --full_json_dir /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_selected_info.json --mode vbench_standard
+args: Namespace(output_path='/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Human_Clothes', full_json_dir='/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_selected_info.json', videos_path='/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/videos/Human_Clothes', dimension=['Human_Clothes'], load_ckpt_from_local=None, read_frame=None, mode='vbench_standard', prompt='None', prompt_file=None, category=None, imaging_quality_preprocessing_mode='longer')
+start evaluation
+File exists: /root/ckpts/vbench2/lmms-lab/LLaVA-Video-7B-Qwen2
+Evaluation meta data saved to /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Human_Clothes/results_2026-06-22-19:11:46_full_info.json
+/usr/local/lib/python3.12/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+2026-06-22 19:11:48.878771: E external/local_xla/xla/stream_executor/cuda/cuda_fft.cc:467] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered
+WARNING: All log messages before absl::InitializeLog() is called are written to STDERR
+E0000 00:00:1782155508.898973 3081735 cuda_dnn.cc:8579] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered
+E0000 00:00:1782155508.905762 3081735 cuda_blas.cc:1407] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered
+W0000 00:00:1782155508.924521 3081735 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.
+W0000 00:00:1782155508.924564 3081735 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.
+W0000 00:00:1782155508.924574 3081735 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.
+W0000 00:00:1782155508.924582 3081735 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.
+2026-06-22 19:11:48.929038: I tensorflow/core/platform/cpu_feature_guard.cc:210] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+/nfs/ywang29/Reward_finetuning/VideoX-Fun/VBench/VBench-2.0/vbench2/third_party/LLaVA_NeXT/llava/model/llava_arch.py:215: SyntaxWarning: "is not" with 'int' literal. Did you mean "!="?
+  if slower_img_feat is not 0:
+/nfs/ywang29/Reward_finetuning/VideoX-Fun/VBench/VBench-2.0/vbench2/third_party/LLaVA_NeXT/llava/model/llava_arch.py:215: SyntaxWarning: "is not" with 'int' literal. Did you mean "!="?
+  if slower_img_feat is not 0:
+cur_full_info_path: /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Human_Clothes/results_2026-06-22-19:11:46_full_info.json
+Loaded LLaVA model: /root/ckpts/vbench2/lmms-lab/LLaVA-Video-7B-Qwen2
+You are using a model of type llava to instantiate a model of type llava_qwen. This is not supported for all configurations of models and can yield errors.
+Sliding Window Attention is enabled but not implemented for `sdpa`; unexpected results may be encountered.
+Loading vision tower: google/siglip-so400m-patch14-384
+2026-06-22 19:11:55,251 - accelerate.utils.modeling - INFO - We will use 90% of the memory on device 0 for storing the model, and 10% for the buffer to avoid OOM. You can set `max_memory` in to a higher value to use more memory (at your own risk).
+
+Model Class: LlavaQwenForCausalLM
+
  0%|          | 0/75 [00:00<?, ?it/s]The attention mask is not set and cannot be inferred from input because pad token is same as eos token. As a consequence, you may observe unexpected behavior. Please pass your input's `attention_mask` to obtain reliable results.
+
  1%|▏         | 1/75 [00:23<28:39, 23.24s/it]
  3%|▎         | 2/75 [00:45<27:18, 22.44s/it]
  4%|▍         | 3/75 [01:07<26:48, 22.34s/it]
  5%|▌         | 4/75 [01:29<26:20, 22.26s/it]
  7%|▋         | 5/75 [01:51<25:50, 22.15s/it]
  8%|▊         | 6/75 [02:13<25:23, 22.08s/it]
  9%|▉         | 7/75 [02:35<24:58, 22.04s/it]
 11%|█         | 8/75 [02:57<24:36, 22.04s/it]
 12%|█▏        | 9/75 [03:19<24:15, 22.05s/it]
 13%|█▎        | 10/75 [03:41<23:53, 22.05s/it]
 15%|█▍        | 11/75 [03:59<22:06, 20.72s/it]
 16%|█▌        | 12/75 [04:21<22:10, 21.12s/it]
 17%|█▋        | 13/75 [04:43<22:05, 21.38s/it]
 19%|█▊        | 14/75 [05:05<21:57, 21.60s/it]
 20%|██        | 15/75 [05:23<20:27, 20.46s/it]
 21%|██▏       | 16/75 [05:45<20:38, 21.00s/it]
 23%|██▎       | 17/75 [05:58<18:06, 18.73s/it]
 24%|██▍       | 18/75 [06:20<18:45, 19.75s/it]
 25%|██▌       | 19/75 [06:43<19:06, 20.47s/it]
 27%|██▋       | 20/75 [07:05<19:12, 20.96s/it]
 28%|██▊       | 21/75 [07:27<19:11, 21.32s/it]
 29%|██▉       | 22/75 [07:49<19:04, 21.60s/it]
 31%|███       | 23/75 [08:11<18:55, 21.83s/it]
 32%|███▏      | 24/75 [08:34<18:40, 21.97s/it]
 33%|███▎      | 25/75 [08:56<18:20, 22.02s/it]
 35%|███▍      | 26/75 [09:18<17:59, 22.04s/it]
 36%|███▌      | 27/75 [09:40<17:39, 22.08s/it]
 37%|███▋      | 28/75 [10:02<17:19, 22.12s/it]
 39%|███▊      | 29/75 [10:25<16:58, 22.14s/it]
 40%|████      | 30/75 [10:47<16:36, 22.15s/it]
 41%|████▏     | 31/75 [11:05<15:16, 20.83s/it]
 43%|████▎     | 32/75 [11:18<13:19, 18.60s/it]
 44%|████▍     | 33/75 [11:36<12:50, 18.35s/it]
 45%|████▌     | 34/75 [11:58<13:19, 19.49s/it]
 47%|████▋     | 35/75 [12:20<13:29, 20.24s/it]
 48%|████▊     | 36/75 [12:38<12:40, 19.49s/it]
 49%|████▉     | 37/75 [13:00<12:49, 20.25s/it]
 51%|█████     | 38/75 [13:22<12:50, 20.82s/it]
 52%|█████▏    | 39/75 [13:39<11:55, 19.86s/it]
 53%|█████▎    | 40/75 [13:57<11:12, 19.20s/it]
 55%|█████▍    | 41/75 [14:19<11:22, 20.08s/it]
 56%|█████▌    | 42/75 [14:41<11:23, 20.70s/it]
 57%|█████▋    | 43/75 [15:04<11:17, 21.17s/it]
 59%|█████▊    | 44/75 [15:26<11:04, 21.42s/it]
 60%|██████    | 45/75 [15:48<10:49, 21.63s/it]
 61%|██████▏   | 46/75 [16:10<10:31, 21.78s/it]
 63%|██████▎   | 47/75 [16:32<10:13, 21.91s/it]
 64%|██████▍   | 48/75 [16:54<09:53, 21.97s/it]
 65%|██████▌   | 49/75 [17:16<09:32, 22.00s/it]
 67%|██████▋   | 50/75 [17:38<09:10, 22.03s/it]
 68%|██████▊   | 51/75 [18:00<08:49, 22.05s/it]
 69%|██████▉   | 52/75 [18:22<08:27, 22.06s/it]
 71%|███████   | 53/75 [18:45<08:05, 22.08s/it]
 72%|███████▏  | 54/75 [19:07<07:43, 22.06s/it]
 73%|███████▎  | 55/75 [19:29<07:20, 22.04s/it]
 75%|███████▍  | 56/75 [19:51<06:58, 22.04s/it]
 76%|███████▌  | 57/75 [20:13<06:37, 22.10s/it]
 77%|███████▋  | 58/75 [20:35<06:16, 22.13s/it]
 79%|███████▊  | 59/75 [20:57<05:54, 22.13s/it]
 80%|████████  | 60/75 [21:19<05:32, 22.15s/it]
 81%|████████▏ | 61/75 [21:42<05:10, 22.19s/it]
 83%|████████▎ | 62/75 [22:04<04:48, 22.21s/it]
 84%|████████▍ | 63/75 [22:26<04:26, 22.21s/it]
 85%|████████▌ | 64/75 [22:48<04:04, 22.23s/it]
 87%|████████▋ | 65/75 [23:11<03:42, 22.22s/it]
 88%|████████▊ | 66/75 [23:33<03:20, 22.23s/it]
 89%|████████▉ | 67/75 [23:55<02:58, 22.27s/it]
 91%|█████████ | 68/75 [24:18<02:35, 22.27s/it]
 92%|█████████▏| 69/75 [24:40<02:13, 22.27s/it]
 93%|█████████▎| 70/75 [25:02<01:51, 22.25s/it]
 95%|█████████▍| 71/75 [25:24<01:29, 22.27s/it]
 96%|█████████▌| 72/75 [25:47<01:06, 22.26s/it]
 97%|█████████▋| 73/75 [26:09<00:44, 22.25s/it]
 99%|█████████▊| 74/75 [26:31<00:22, 22.24s/it]
+Evaluation results saved to /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Human_Clothes/results_2026-06-22-19:11:46_eval_results.json
+done

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/logs/vbench2_parallel/Instance_Preservation_shard_2.log

@@ -0,0 +1,47 @@
+[2026-06-23 03:08:23 UTC] COMMAND /usr/local/bin/python -u evaluate.py --videos_path /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/videos/Instance_Preservation --dimension Instance_Preservation --output_path /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Instance_Preservation_shards/2026-06-23-03:08:22/shard_2 --full_json_dir /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Instance_Preservation_shards/2026-06-23-03:08:22/full_info_shard_2.json --mode vbench_standard
+args: Namespace(output_path='/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Instance_Preservation_shards/2026-06-23-03:08:22/shard_2', full_json_dir='/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Instance_Preservation_shards/2026-06-23-03:08:22/full_info_shard_2.json', videos_path='/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/videos/Instance_Preservation', dimension=['Instance_Preservation'], load_ckpt_from_local=None, read_frame=None, mode='vbench_standard', prompt='None', prompt_file=None, category=None, imaging_quality_preprocessing_mode='longer')
+start evaluation
+Evaluation meta data saved to /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Instance_Preservation_shards/2026-06-23-03:08:22/shard_2/results_2026-06-23-03:08:26_full_info.json
+/usr/local/lib/python3.12/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+2026-06-23 03:08:29.425724: E external/local_xla/xla/stream_executor/cuda/cuda_fft.cc:467] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered
+WARNING: All log messages before absl::InitializeLog() is called are written to STDERR
+E0000 00:00:1782184109.445022 3295554 cuda_dnn.cc:8579] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered
+E0000 00:00:1782184109.451721 3295554 cuda_blas.cc:1407] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered
+W0000 00:00:1782184109.469753 3295554 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.
+W0000 00:00:1782184109.469794 3295554 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.
+W0000 00:00:1782184109.469803 3295554 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.
+W0000 00:00:1782184109.469811 3295554 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.
+[INFO:swift] Successfully registered `/nfs/ywang29/Reward_finetuning/VideoX-Fun/VBench/VBench-2.0/vbench2/third_party/Instance_detector/swift/llm/dataset/data/dataset_info.json`.
+cur_full_info_path: /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Instance_Preservation_shards/2026-06-23-03:08:22/shard_2/results_2026-06-23-03:08:26_full_info.json
+[INFO:swift] Successfully loaded /root/ckpts/vbench2/instance_anomaly_detector/model/args.json.
+[INFO:swift] Loading the model using model_dir: /root/ckpts/huggingface/Qwen/Qwen2.5-VL-3B-Instruct
+[INFO:swift] Setting torch_dtype: torch.bfloat16
+Using a slow image processor as `use_fast` is unset and a slow processor was saved with this model. `use_fast=True` will be the default behavior in v4.52, even if the model was saved with a slow processor. This will result in minor differences in outputs. You'll still be able to use a slow processor with `use_fast=False`.
+[INFO:swift] model_kwargs: {'device_map': 'cuda:0'}
+
+[INFO:swift] Setting image_factor: 28. You can adjust this hyperparameter through the environment variable: `IMAGE_FACTOR`.
+[INFO:swift] Setting min_pixels: 3136. You can adjust this hyperparameter through the environment variable: `MIN_PIXELS`.
+[INFO:swift] Using environment variable `MAX_PIXELS`, Setting max_pixels: 1003520.
+[INFO:swift] Setting max_ratio: 200. You can adjust this hyperparameter through the environment variable: `MAX_RATIO`.
+[INFO:swift] Using environment variable `VIDEO_MIN_PIXELS`, Setting video_min_pixels: 100352.
+[INFO:swift] Using environment variable `VIDEO_MAX_PIXELS`, Setting video_max_pixels: 602112.
+[INFO:swift] Using environment variable `VIDEO_TOTAL_PIXELS`, Setting video_total_pixels: 90316800.
+[INFO:swift] Setting frame_factor: 2. You can adjust this hyperparameter through the environment variable: `FRAME_FACTOR`.
+[INFO:swift] Setting fps: 2.0. You can adjust this hyperparameter through the environment variable: `FPS`.
+[INFO:swift] Using environment variable `FPS_MIN_FRAMES`, Setting fps_min_frames: 32.
+[INFO:swift] Using environment variable `FPS_MAX_FRAMES`, Setting fps_max_frames: 50.
+[INFO:swift] default_system: 'You are a helpful and harmless assistant.'
+[INFO:swift] max_length: 128000
+[INFO:swift] response_prefix: ''
+[INFO:swift] agent_template: hermes
+[INFO:swift] norm_bbox: none
+[INFO:swift] default_system: 'You are a helpful and harmless assistant.'
+[INFO:swift] max_length: 128000
+[INFO:swift] response_prefix: ''
+[INFO:swift] agent_template: hermes
+[INFO:swift] norm_bbox: none
+
  0%|          | 0/8 [00:00<?, ?it/s]qwen-vl-utils using decord to read video.
+
 12%|█▎        | 1/8 [01:49<12:44, 109.17s/it]
 25%|██▌       | 2/8 [03:36<10:49, 108.25s/it]
 38%|███▊      | 3/8 [05:24<09:00, 108.15s/it]
 50%|█████     | 4/8 [07:03<06:57, 104.29s/it]
 62%|██████▎   | 5/8 [08:50<05:15, 105.22s/it]
 75%|███████▌  | 6/8 [10:36<03:31, 105.74s/it]
 88%|████████▊ | 7/8 [12:06<01:40, 100.57s/it]
+Evaluation results saved to /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Instance_Preservation_shards/2026-06-23-03:08:22/shard_2/results_2026-06-23-03:08:26_eval_results.json
+done

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/logs/vbench2_parallel/Instance_Preservation_shard_5.log

@@ -0,0 +1,47 @@
+[2026-06-23 03:08:23 UTC] COMMAND /usr/local/bin/python -u evaluate.py --videos_path /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/videos/Instance_Preservation --dimension Instance_Preservation --output_path /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Instance_Preservation_shards/2026-06-23-03:08:22/shard_5 --full_json_dir /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Instance_Preservation_shards/2026-06-23-03:08:22/full_info_shard_5.json --mode vbench_standard
+args: Namespace(output_path='/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Instance_Preservation_shards/2026-06-23-03:08:22/shard_5', full_json_dir='/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Instance_Preservation_shards/2026-06-23-03:08:22/full_info_shard_5.json', videos_path='/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/videos/Instance_Preservation', dimension=['Instance_Preservation'], load_ckpt_from_local=None, read_frame=None, mode='vbench_standard', prompt='None', prompt_file=None, category=None, imaging_quality_preprocessing_mode='longer')
+start evaluation
+Evaluation meta data saved to /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Instance_Preservation_shards/2026-06-23-03:08:22/shard_5/results_2026-06-23-03:08:26_full_info.json
+/usr/local/lib/python3.12/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+2026-06-23 03:08:29.538846: E external/local_xla/xla/stream_executor/cuda/cuda_fft.cc:467] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered
+WARNING: All log messages before absl::InitializeLog() is called are written to STDERR
+E0000 00:00:1782184109.557629 3295557 cuda_dnn.cc:8579] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered
+E0000 00:00:1782184109.564044 3295557 cuda_blas.cc:1407] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered
+W0000 00:00:1782184109.581063 3295557 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.
+W0000 00:00:1782184109.581108 3295557 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.
+W0000 00:00:1782184109.581117 3295557 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.
+W0000 00:00:1782184109.581125 3295557 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.
+[INFO:swift] Successfully registered `/nfs/ywang29/Reward_finetuning/VideoX-Fun/VBench/VBench-2.0/vbench2/third_party/Instance_detector/swift/llm/dataset/data/dataset_info.json`.
+cur_full_info_path: /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Instance_Preservation_shards/2026-06-23-03:08:22/shard_5/results_2026-06-23-03:08:26_full_info.json
+[INFO:swift] Successfully loaded /root/ckpts/vbench2/instance_anomaly_detector/model/args.json.
+[INFO:swift] Loading the model using model_dir: /root/ckpts/huggingface/Qwen/Qwen2.5-VL-3B-Instruct
+[INFO:swift] Setting torch_dtype: torch.bfloat16
+Using a slow image processor as `use_fast` is unset and a slow processor was saved with this model. `use_fast=True` will be the default behavior in v4.52, even if the model was saved with a slow processor. This will result in minor differences in outputs. You'll still be able to use a slow processor with `use_fast=False`.
+[INFO:swift] model_kwargs: {'device_map': 'cuda:0'}
+
+[INFO:swift] Setting image_factor: 28. You can adjust this hyperparameter through the environment variable: `IMAGE_FACTOR`.
+[INFO:swift] Setting min_pixels: 3136. You can adjust this hyperparameter through the environment variable: `MIN_PIXELS`.
+[INFO:swift] Using environment variable `MAX_PIXELS`, Setting max_pixels: 1003520.
+[INFO:swift] Setting max_ratio: 200. You can adjust this hyperparameter through the environment variable: `MAX_RATIO`.
+[INFO:swift] Using environment variable `VIDEO_MIN_PIXELS`, Setting video_min_pixels: 100352.
+[INFO:swift] Using environment variable `VIDEO_MAX_PIXELS`, Setting video_max_pixels: 602112.
+[INFO:swift] Using environment variable `VIDEO_TOTAL_PIXELS`, Setting video_total_pixels: 90316800.
+[INFO:swift] Setting frame_factor: 2. You can adjust this hyperparameter through the environment variable: `FRAME_FACTOR`.
+[INFO:swift] Setting fps: 2.0. You can adjust this hyperparameter through the environment variable: `FPS`.
+[INFO:swift] Using environment variable `FPS_MIN_FRAMES`, Setting fps_min_frames: 32.
+[INFO:swift] Using environment variable `FPS_MAX_FRAMES`, Setting fps_max_frames: 50.
+[INFO:swift] default_system: 'You are a helpful and harmless assistant.'
+[INFO:swift] max_length: 128000
+[INFO:swift] response_prefix: ''
+[INFO:swift] agent_template: hermes
+[INFO:swift] norm_bbox: none
+[INFO:swift] default_system: 'You are a helpful and harmless assistant.'
+[INFO:swift] max_length: 128000
+[INFO:swift] response_prefix: ''
+[INFO:swift] agent_template: hermes
+[INFO:swift] norm_bbox: none
+
  0%|          | 0/8 [00:00<?, ?it/s]qwen-vl-utils using decord to read video.
+
 12%|█▎        | 1/8 [01:50<12:53, 110.51s/it]
 25%|██▌       | 2/8 [03:40<10:59, 109.98s/it]
 38%|███▊      | 3/8 [05:17<08:42, 104.41s/it]
 50%|█████     | 4/8 [07:04<07:01, 105.26s/it]
 62%|██████▎   | 5/8 [08:52<05:18, 106.25s/it]
 75%|███████▌  | 6/8 [10:39<03:32, 106.40s/it]
 88%|████████▊ | 7/8 [12:26<01:46, 106.77s/it]
+Evaluation results saved to /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Instance_Preservation_shards/2026-06-23-03:08:22/shard_5/results_2026-06-23-03:08:26_eval_results.json
+done

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/logs/vbench2_parallel/Mechanics.log

@@ -0,0 +1,31 @@
+[2026-06-22 19:55:43 UTC] COMMAND /usr/local/bin/python -u evaluate.py --videos_path /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/videos/Mechanics --dimension Mechanics --output_path /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Mechanics --full_json_dir /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_selected_info.json --mode vbench_standard
+args: Namespace(output_path='/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Mechanics', full_json_dir='/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_selected_info.json', videos_path='/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/videos/Mechanics', dimension=['Mechanics'], load_ckpt_from_local=None, read_frame=None, mode='vbench_standard', prompt='None', prompt_file=None, category=None, imaging_quality_preprocessing_mode='longer')
+start evaluation
+File exists: /root/ckpts/vbench2/lmms-lab/LLaVA-Video-7B-Qwen2
+Evaluation meta data saved to /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Mechanics/results_2026-06-22-19:55:46_full_info.json
+/usr/local/lib/python3.12/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+2026-06-22 19:55:48.519532: E external/local_xla/xla/stream_executor/cuda/cuda_fft.cc:467] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered
+WARNING: All log messages before absl::InitializeLog() is called are written to STDERR
+E0000 00:00:1782158148.538555 3113593 cuda_dnn.cc:8579] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered
+E0000 00:00:1782158148.545105 3113593 cuda_blas.cc:1407] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered
+W0000 00:00:1782158148.562539 3113593 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.
+W0000 00:00:1782158148.562582 3113593 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.
+W0000 00:00:1782158148.562591 3113593 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.
+W0000 00:00:1782158148.562600 3113593 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.
+2026-06-22 19:55:48.566771: I tensorflow/core/platform/cpu_feature_guard.cc:210] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+/nfs/ywang29/Reward_finetuning/VideoX-Fun/VBench/VBench-2.0/vbench2/third_party/LLaVA_NeXT/llava/model/llava_arch.py:215: SyntaxWarning: "is not" with 'int' literal. Did you mean "!="?
+  if slower_img_feat is not 0:
+cur_full_info_path: /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Mechanics/results_2026-06-22-19:55:46_full_info.json
+Loaded LLaVA model: /root/ckpts/vbench2/lmms-lab/LLaVA-Video-7B-Qwen2
+You are using a model of type llava to instantiate a model of type llava_qwen. This is not supported for all configurations of models and can yield errors.
+Sliding Window Attention is enabled but not implemented for `sdpa`; unexpected results may be encountered.
+Loading vision tower: google/siglip-so400m-patch14-384
+2026-06-22 19:55:54,598 - accelerate.utils.modeling - INFO - We will use 90% of the memory on device 0 for storing the model, and 10% for the buffer to avoid OOM. You can set `max_memory` in to a higher value to use more memory (at your own risk).
+
+Model Class: LlavaQwenForCausalLM
+
  0%|          | 0/57 [00:00<?, ?it/s]The attention mask is not set and cannot be inferred from input because pad token is same as eos token. As a consequence, you may observe unexpected behavior. Please pass your input's `attention_mask` to obtain reliable results.
+
  2%|▏         | 1/57 [00:16<15:17, 16.38s/it]
  4%|▎         | 2/57 [00:31<14:29, 15.80s/it]
  5%|▌         | 3/57 [00:47<14:06, 15.67s/it]
  7%|▋         | 4/57 [01:02<13:48, 15.64s/it]
  9%|▉         | 5/57 [01:18<13:30, 15.58s/it]
 11%|█         | 6/57 [01:33<13:14, 15.57s/it]
 12%|█▏        | 7/57 [01:49<12:55, 15.52s/it]
 14%|█▍        | 8/57 [02:04<12:41, 15.55s/it]
 16%|█▌        | 9/57 [02:13<10:47, 13.48s/it]
 18%|█▊        | 10/57 [02:22<09:26, 12.05s/it]
 19%|█▉        | 11/57 [02:31<08:28, 11.06s/it]
 21%|██        | 12/57 [02:40<07:48, 10.41s/it]
 23%|██▎       | 13/57 [02:55<08:44, 11.93s/it]
 25%|██▍       | 14/57 [03:04<07:56, 11.08s/it]
 26%|██▋       | 15/57 [03:13<07:15, 10.38s/it]
 28%|██▊       | 16/57 [03:29<08:08, 11.92s/it]
 30%|██▉       | 17/57 [03:44<08:40, 13.01s/it]
 32%|███▏      | 18/57 [04:00<08:57, 13.77s/it]
 33%|███▎      | 19/57 [04:15<09:03, 14.31s/it]
 35%|███▌      | 20/57 [04:31<09:01, 14.65s/it]
 37%|███▋      | 21/57 [04:46<08:54, 14.85s/it]
 39%|███▊      | 22/57 [04:55<07:37, 13.07s/it]
 40%|████      | 23/57 [05:04<06:41, 11.79s/it]
 42%|████▏     | 24/57 [05:13<06:00, 10.92s/it]
 44%|████▍     | 25/57 [05:22<05:29, 10.29s/it]
 46%|████▌     | 26/57 [05:30<05:04,  9.81s/it]
 47%|████▋     | 27/57 [05:39<04:45,  9.51s/it]
 49%|████▉     | 28/57 [05:48<04:29,  9.29s/it]
 51%|█████     | 29/57 [05:57<04:17,  9.19s/it]
 53%|█████▎    | 30/57 [06:12<04:57, 11.03s/it]
 54%|█████▍    | 31/57 [06:28<05:20, 12.33s/it]
 56%|█████▌    | 32/57 [06:43<05:31, 13.24s/it]
 58%|█████▊    | 33/57 [06:58<05:33, 13.88s/it]
 60%|█████▉    | 34/57 [07:14<05:29, 14.31s/it]
 61%|██████▏   | 35/57 [07:29<05:21, 14.61s/it]
 63%|██████▎   | 36/57 [07:44<05:10, 14.79s/it]
 65%|██████▍   | 37/57 [07:59<04:58, 14.94s/it]
 67%|██████▋   | 38/57 [08:15<04:46, 15.07s/it]
 68%|██████▊   | 39/57 [08:30<04:33, 15.18s/it]
 70%|███████   | 40/57 [08:46<04:19, 15.26s/it]
 72%|███████▏  | 41/57 [09:01<04:04, 15.30s/it]
 74%|███████▎  | 42/57 [09:16<03:49, 15.28s/it]
 75%|███████▌  | 43/57 [09:32<03:34, 15.32s/it]
 77%|███████▋  | 44/57 [09:47<03:19, 15.33s/it]
 79%|███████▉  | 45/57 [10:02<03:04, 15.36s/it]
 81%|████████  | 46/57 [10:18<02:49, 15.41s/it]
 82%|████████▏ | 47/57 [10:34<02:34, 15.48s/it]
 84%|████████▍ | 48/57 [10:49<02:19, 15.46s/it]
 86%|████████▌ | 49/57 [11:05<02:03, 15.47s/it]
 88%|████████▊ | 50/57 [11:20<01:48, 15.45s/it]
 89%|████████▉ | 51/57 [11:35<01:32, 15.42s/it]
 91%|█████████ | 52/57 [11:51<01:16, 15.38s/it]
 93%|█████████▎| 53/57 [12:06<01:01, 15.35s/it]
 95%|█████████▍| 54/57 [12:21<00:46, 15.35s/it]
 96%|█████████▋| 55/57 [12:37<00:30, 15.34s/it]
 98%|█████████▊| 56/57 [12:52<00:15, 15.32s/it]
+Evaluation results saved to /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Mechanics/results_2026-06-22-19:55:46_eval_results.json
+done

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/logs/vbench2_parallel/Motion_Rationality.log

@@ -0,0 +1,31 @@
+[2026-06-22 20:41:13 UTC] COMMAND /usr/local/bin/python -u evaluate.py --videos_path /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/videos/Motion_Rationality --dimension Motion_Rationality --output_path /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Motion_Rationality --full_json_dir /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_selected_info.json --mode vbench_standard
+args: Namespace(output_path='/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Motion_Rationality', full_json_dir='/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_selected_info.json', videos_path='/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/videos/Motion_Rationality', dimension=['Motion_Rationality'], load_ckpt_from_local=None, read_frame=None, mode='vbench_standard', prompt='None', prompt_file=None, category=None, imaging_quality_preprocessing_mode='longer')
+start evaluation
+File exists: /root/ckpts/vbench2/lmms-lab/LLaVA-Video-7B-Qwen2
+Evaluation meta data saved to /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Motion_Rationality/results_2026-06-22-20:41:16_full_info.json
+/usr/local/lib/python3.12/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+2026-06-22 20:41:18.660701: E external/local_xla/xla/stream_executor/cuda/cuda_fft.cc:467] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered
+WARNING: All log messages before absl::InitializeLog() is called are written to STDERR
+E0000 00:00:1782160878.679842 3124928 cuda_dnn.cc:8579] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered
+E0000 00:00:1782160878.686407 3124928 cuda_blas.cc:1407] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered
+W0000 00:00:1782160878.703834 3124928 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.
+W0000 00:00:1782160878.703874 3124928 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.
+W0000 00:00:1782160878.703883 3124928 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.
+W0000 00:00:1782160878.703891 3124928 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.
+2026-06-22 20:41:18.708104: I tensorflow/core/platform/cpu_feature_guard.cc:210] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+/nfs/ywang29/Reward_finetuning/VideoX-Fun/VBench/VBench-2.0/vbench2/third_party/LLaVA_NeXT/llava/model/llava_arch.py:215: SyntaxWarning: "is not" with 'int' literal. Did you mean "!="?
+  if slower_img_feat is not 0:
+cur_full_info_path: /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Motion_Rationality/results_2026-06-22-20:41:16_full_info.json
+Loaded LLaVA model: /root/ckpts/vbench2/lmms-lab/LLaVA-Video-7B-Qwen2
+You are using a model of type llava to instantiate a model of type llava_qwen. This is not supported for all configurations of models and can yield errors.
+Sliding Window Attention is enabled but not implemented for `sdpa`; unexpected results may be encountered.
+Loading vision tower: google/siglip-so400m-patch14-384
+2026-06-22 20:41:25,855 - accelerate.utils.modeling - INFO - We will use 90% of the memory on device 0 for storing the model, and 10% for the buffer to avoid OOM. You can set `max_memory` in to a higher value to use more memory (at your own risk).
+
+Model Class: LlavaQwenForCausalLM
+
  0%|          | 0/58 [00:00<?, ?it/s]The attention mask is not set and cannot be inferred from input because pad token is same as eos token. As a consequence, you may observe unexpected behavior. Please pass your input's `attention_mask` to obtain reliable results.
+
  2%|▏         | 1/58 [00:29<28:07, 29.61s/it]
  3%|▎         | 2/58 [00:58<27:00, 28.94s/it]
  5%|▌         | 3/58 [01:26<26:23, 28.80s/it]
  7%|▋         | 4/58 [01:55<25:49, 28.69s/it]
  9%|▊         | 5/58 [02:23<25:19, 28.68s/it]
 10%|█         | 6/58 [02:52<24:50, 28.67s/it]
 12%|█▏        | 7/58 [03:21<24:21, 28.66s/it]
 14%|█▍        | 8/58 [03:49<23:54, 28.69s/it]
 16%|█▌        | 9/58 [04:18<23:25, 28.69s/it]
 17%|█▋        | 10/58 [04:47<23:06, 28.89s/it]
 19%|█▉        | 11/58 [05:16<22:35, 28.84s/it]
 21%|██        | 12/58 [05:45<22:05, 28.82s/it]
 22%|██▏       | 13/58 [06:14<21:37, 28.83s/it]
 24%|██▍       | 14/58 [06:42<21:05, 28.76s/it]
 26%|██▌       | 15/58 [07:12<20:41, 28.86s/it]
 28%|██▊       | 16/58 [07:40<20:11, 28.84s/it]
 29%|██▉       | 17/58 [08:09<19:40, 28.80s/it]
 31%|███       | 18/58 [08:38<19:10, 28.77s/it]
 33%|███▎      | 19/58 [09:06<18:41, 28.74s/it]
 34%|███▍      | 20/58 [09:35<18:11, 28.72s/it]
 36%|███▌      | 21/58 [10:04<17:43, 28.73s/it]
 38%|███▊      | 22/58 [10:33<17:14, 28.72s/it]
 40%|███▉      | 23/58 [11:01<16:46, 28.75s/it]
 41%|████▏     | 24/58 [11:30<16:16, 28.73s/it]
 43%|████▎     | 25/58 [11:59<15:49, 28.77s/it]
 45%|████▍     | 26/58 [12:27<15:18, 28.71s/it]
 47%|████▋     | 27/58 [12:56<14:50, 28.74s/it]
 48%|████▊     | 28/58 [13:25<14:20, 28.70s/it]
 50%|█████     | 29/58 [13:54<13:53, 28.73s/it]
 52%|█████▏    | 30/58 [14:22<13:24, 28.73s/it]
 53%|█████▎    | 31/58 [14:51<12:55, 28.73s/it]
 55%|█████▌    | 32/58 [15:20<12:28, 28.80s/it]
 57%|█████▋    | 33/58 [15:49<12:00, 28.82s/it]
 59%|█████▊    | 34/58 [16:18<11:30, 28.78s/it]
 60%|██████    | 35/58 [16:46<11:00, 28.73s/it]
 62%|██████▏   | 36/58 [17:15<10:32, 28.76s/it]
 64%|██████▍   | 37/58 [17:44<10:03, 28.74s/it]
 66%|██████▌   | 38/58 [18:13<09:34, 28.74s/it]
 67%|██████▋   | 39/58 [18:42<09:12, 29.06s/it]
 69%|██████▉   | 40/58 [19:11<08:40, 28.90s/it]
 71%|███████   | 41/58 [19:39<08:09, 28.81s/it]
 72%|███████▏  | 42/58 [20:08<07:39, 28.70s/it]
 74%|███████▍  | 43/58 [20:36<07:09, 28.65s/it]
 76%|███████▌  | 44/58 [21:05<06:40, 28.62s/it]
 78%|███████▊  | 45/58 [21:34<06:12, 28.62s/it]
 79%|███████▉  | 46/58 [22:02<05:43, 28.66s/it]
 81%|████████  | 47/58 [22:31<05:14, 28.61s/it]
 83%|████████▎ | 48/58 [22:59<04:46, 28.62s/it]
 84%|████████▍ | 49/58 [23:28<04:18, 28.69s/it]
 86%|████████▌ | 50/58 [23:57<03:49, 28.64s/it]
 88%|████████▊ | 51/58 [24:25<03:20, 28.63s/it]
 90%|████████▉ | 52/58 [24:54<02:51, 28.65s/it]
 91%|█████████▏| 53/58 [25:23<02:23, 28.69s/it]
 93%|█████████▎| 54/58 [25:52<01:54, 28.66s/it]
 95%|█████████▍| 55/58 [26:20<01:25, 28.66s/it]
 97%|█████████▋| 56/58 [26:49<00:57, 28.65s/it]
 98%|█████████▊| 57/58 [27:17<00:28, 28.64s/it]
+Evaluation results saved to /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Motion_Rationality/results_2026-06-22-20:41:16_eval_results.json
+done

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/logs/vbench2_parallel/Thermotics.log

@@ -0,0 +1,31 @@
+[2026-06-22 21:06:43 UTC] COMMAND /usr/local/bin/python -u evaluate.py --videos_path /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/videos/Thermotics --dimension Thermotics --output_path /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Thermotics --full_json_dir /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_selected_info.json --mode vbench_standard
+args: Namespace(output_path='/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Thermotics', full_json_dir='/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_selected_info.json', videos_path='/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/videos/Thermotics', dimension=['Thermotics'], load_ckpt_from_local=None, read_frame=None, mode='vbench_standard', prompt='None', prompt_file=None, category=None, imaging_quality_preprocessing_mode='longer')
+start evaluation
+File exists: /root/ckpts/vbench2/lmms-lab/LLaVA-Video-7B-Qwen2
+Evaluation meta data saved to /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Thermotics/results_2026-06-22-21:06:46_full_info.json
+/usr/local/lib/python3.12/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
+  warnings.warn(
+2026-06-22 21:06:48.576008: E external/local_xla/xla/stream_executor/cuda/cuda_fft.cc:467] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered
+WARNING: All log messages before absl::InitializeLog() is called are written to STDERR
+E0000 00:00:1782162408.595259 3192887 cuda_dnn.cc:8579] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered
+E0000 00:00:1782162408.601952 3192887 cuda_blas.cc:1407] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered
+W0000 00:00:1782162408.619931 3192887 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.
+W0000 00:00:1782162408.619983 3192887 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.
+W0000 00:00:1782162408.619992 3192887 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.
+W0000 00:00:1782162408.620000 3192887 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.
+2026-06-22 21:06:48.624329: I tensorflow/core/platform/cpu_feature_guard.cc:210] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+/nfs/ywang29/Reward_finetuning/VideoX-Fun/VBench/VBench-2.0/vbench2/third_party/LLaVA_NeXT/llava/model/llava_arch.py:215: SyntaxWarning: "is not" with 'int' literal. Did you mean "!="?
+  if slower_img_feat is not 0:
+cur_full_info_path: /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Thermotics/results_2026-06-22-21:06:46_full_info.json
+Loaded LLaVA model: /root/ckpts/vbench2/lmms-lab/LLaVA-Video-7B-Qwen2
+You are using a model of type llava to instantiate a model of type llava_qwen. This is not supported for all configurations of models and can yield errors.
+Sliding Window Attention is enabled but not implemented for `sdpa`; unexpected results may be encountered.
+Loading vision tower: google/siglip-so400m-patch14-384
+2026-06-22 21:06:54,500 - accelerate.utils.modeling - INFO - We will use 90% of the memory on device 0 for storing the model, and 10% for the buffer to avoid OOM. You can set `max_memory` in to a higher value to use more memory (at your own risk).
+
+Model Class: LlavaQwenForCausalLM
+
  0%|          | 0/50 [00:00<?, ?it/s]The attention mask is not set and cannot be inferred from input because pad token is same as eos token. As a consequence, you may observe unexpected behavior. Please pass your input's `attention_mask` to obtain reliable results.
+
  2%|▏         | 1/50 [00:16<13:06, 16.05s/it]
  4%|▍         | 2/50 [00:31<12:39, 15.83s/it]
  6%|▌         | 3/50 [00:47<12:19, 15.73s/it]
  8%|▊         | 4/50 [01:02<12:01, 15.67s/it]
 10%|█         | 5/50 [01:18<11:45, 15.67s/it]
 12%|█▏        | 6/50 [01:34<11:33, 15.77s/it]
 14%|█▍        | 7/50 [01:50<11:16, 15.74s/it]
 16%|█▌        | 8/50 [02:05<11:00, 15.72s/it]
 18%|█▊        | 9/50 [02:21<10:43, 15.69s/it]
 20%|██        | 10/50 [02:37<10:26, 15.66s/it]
 22%|██▏       | 11/50 [02:46<08:53, 13.68s/it]
 24%|██▍       | 12/50 [03:02<09:03, 14.29s/it]
 26%|██▌       | 13/50 [03:17<09:04, 14.71s/it]
 28%|██▊       | 14/50 [03:33<09:00, 15.02s/it]
 30%|███       | 15/50 [03:49<08:52, 15.20s/it]
 32%|███▏      | 16/50 [04:04<08:40, 15.32s/it]
 34%|███▍      | 17/50 [04:20<08:27, 15.39s/it]
 36%|███▌      | 18/50 [04:35<08:14, 15.47s/it]
 38%|███▊      | 19/50 [04:51<08:00, 15.50s/it]
 40%|████      | 20/50 [05:00<06:47, 13.60s/it]
 42%|████▏     | 21/50 [05:16<06:51, 14.20s/it]
 44%|████▍     | 22/50 [05:31<06:49, 14.61s/it]
 46%|████▌     | 23/50 [05:41<05:51, 13.00s/it]
 48%|████▊     | 24/50 [05:56<05:57, 13.76s/it]
 50%|█████     | 25/50 [06:12<05:57, 14.32s/it]
 52%|█████▏    | 26/50 [06:27<05:53, 14.72s/it]
 54%|█████▍    | 27/50 [06:43<05:44, 14.97s/it]
 56%|█████▌    | 28/50 [06:58<05:33, 15.15s/it]
 58%|█████▊    | 29/50 [07:14<05:20, 15.27s/it]
 60%|██████    | 30/50 [07:30<05:07, 15.37s/it]
 62%|██████▏   | 31/50 [07:45<04:52, 15.41s/it]
 64%|██████▍   | 32/50 [08:01<04:38, 15.45s/it]
 66%|██████▌   | 33/50 [08:16<04:22, 15.46s/it]
 68%|██████▊   | 34/50 [08:32<04:07, 15.47s/it]
 70%|███████   | 35/50 [08:47<03:52, 15.49s/it]
 72%|███████▏  | 36/50 [09:03<03:37, 15.52s/it]
 74%|███████▍  | 37/50 [09:18<03:21, 15.52s/it]
 76%|███████▌  | 38/50 [09:40<03:29, 17.50s/it]
 78%|███████▊  | 39/50 [09:56<03:06, 16.92s/it]
 80%|████████  | 40/50 [10:12<02:45, 16.53s/it]
 82%|████████▏ | 41/50 [10:27<02:26, 16.26s/it]
 84%|████████▍ | 42/50 [10:43<02:08, 16.06s/it]
 86%|████████▌ | 43/50 [10:52<01:37, 13.98s/it]
 88%|████████▊ | 44/50 [11:01<01:15, 12.54s/it]
 90%|█████████ | 45/50 [11:17<01:07, 13.47s/it]
 92%|█████████▏| 46/50 [11:32<00:56, 14.10s/it]
 94%|█████████▍| 47/50 [11:48<00:43, 14.57s/it]
 96%|█████████▌| 48/50 [12:04<00:29, 14.90s/it]
 98%|█████████▊| 49/50 [12:19<00:15, 15.18s/it]
+Evaluation results saved to /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/vbench2_eval/step_100/vbench2_results/Thermotics/results_2026-06-22-21:06:46_eval_results.json
+done

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/debug-internal.log

@@ -0,0 +1,19 @@
+{"time":"2026-06-29T18:40:45.440011198Z","level":"INFO","msg":"wandb-core"}
+{"time":"2026-06-29T18:40:45.441261839Z","level":"INFO","msg":"stream: starting","core version":"0.28.0"}
+{"time":"2026-06-29T18:40:45.725249083Z","level":"INFO","msg":"stream: created new stream","id":"rgt7q8d2"}
+{"time":"2026-06-29T18:40:45.725342359Z","level":"INFO","msg":"handler: started"}
+{"time":"2026-06-29T18:40:45.728343397Z","level":"INFO","msg":"stream: started"}
+{"time":"2026-06-29T18:40:45.728433146Z","level":"INFO","msg":"writer: started","stream_id":"rgt7q8d2"}
+{"time":"2026-06-29T18:40:45.728436334Z","level":"INFO","msg":"sender: started"}
+{"time":"2026-06-29T18:40:46.455781437Z","level":"INFO","msg":"filestream: sending request","total_files":1,"console_offset":5453,"console_lines":1}
+{"time":"2026-06-29T18:40:47.00928851Z","level":"INFO","msg":"filestream: request sent","status":"200 OK"}
+{"time":"2026-06-29T18:41:01.456147642Z","level":"INFO","msg":"filestream: sending request","total_files":2,"events_offset":572,"events_lines":2,"console_offset":5454,"console_lines":132,"uploaded_len":2}
+{"time":"2026-06-29T18:41:01.839422055Z","level":"INFO","msg":"filestream: request sent","status":"200 OK"}
+{"time":"2026-06-29T18:41:16.456404094Z","level":"INFO","msg":"filestream: sending request","total_files":1,"events_offset":574,"events_lines":2}
+{"time":"2026-06-29T18:41:16.848783239Z","level":"INFO","msg":"filestream: request sent","status":"200 OK"}
+{"time":"2026-06-29T18:41:31.455823407Z","level":"INFO","msg":"filestream: sending request","total_files":1,"events_offset":576,"events_lines":2}
+{"time":"2026-06-29T18:41:31.724666236Z","level":"INFO","msg":"filestream: request sent","status":"200 OK"}
+{"time":"2026-06-29T18:41:46.456607312Z","level":"INFO","msg":"filestream: sending request","total_files":1,"events_offset":578,"events_lines":2}
+{"time":"2026-06-29T18:41:46.694582586Z","level":"INFO","msg":"filestream: request sent","status":"200 OK"}
+{"time":"2026-06-29T18:42:01.456681904Z","level":"INFO","msg":"filestream: sending request","total_files":1,"events_offset":580,"events_lines":2}
+{"time":"2026-06-29T18:42:01.767313215Z","level":"INFO","msg":"filestream: request sent","status":"200 OK"}

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/debug.log

@@ -0,0 +1,20 @@
+2026-06-29 18:40:45,206 INFO    MainThread:83423 [wandb_setup.py:_flush():81] Current SDK version is 0.28.0
+2026-06-29 18:40:45,207 INFO    MainThread:83423 [wandb_setup.py:_flush():81] Configure stats pid to 83423
+2026-06-29 18:40:45,207 INFO    MainThread:83423 [wandb_setup.py:_flush():81] Loading settings from environment variables
+2026-06-29 18:40:45,207 INFO    MainThread:83423 [wandb_init.py:setup_run_log_directory():725] Logging user logs to outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260629_184045-rgt7q8d2/logs/debug.log
+2026-06-29 18:40:45,207 INFO    MainThread:83423 [wandb_init.py:setup_run_log_directory():726] Logging internal logs to outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260629_184045-rgt7q8d2/logs/debug-internal.log
+2026-06-29 18:40:45,207 INFO    MainThread:83423 [wandb_init.py:init():768] calling init triggers
+2026-06-29 18:40:45,207 INFO    MainThread:83423 [wandb_init.py:init():773] wandb.init called with sweep_config: {}
+config: {'pretrained_model_root': './ckpts', 'pretrained_transformer_version': '480p_t2v', 'post_train_type': 'grpo', 'learning_rate': 1e-05, 'weight_decay': 0.01, 'max_steps': 10000, 'warmup_steps': 500, 'gradient_accumulation_steps': 1, 'max_grad_norm': 1.0, 'use_muon': True, 'num_train_timesteps': 1000, 'train_timestep_shift': 3.0, 'validation_timestep_shift': 5.0, 'snr_type': 'lognorm', 'task_type': 't2v', 'i2v_prob': 0.3, 'enable_fsdp': True, 'enable_gradient_checkpointing': True, 'sp_size': 2, 'dp_replicate': 1, 'batch_size': 2, 'num_workers': 4, 'prefetch_factor': 2, 'num_generations': 4, 'use_same_noise': False, 'output_dir': './outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3', 'checkpoints_directory': 's3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3', 'save_interval': 25, 'timer_interval': 0, 'auto_resume_from_timer': True, 'log_interval': 1, 'enable_wandb': True, 'project_name': 'SAGE-GRPO', 'run_name': 'yw006-4-sage-qwen35-9b-token-credit-video-adv-window3', 'wandb_key': None, 'wandb_username': None, 'wandb_base_url': 'https://api.wandb.ai', 'wandb_entity': None, 'wandb_silent': True, 'dtype': 'bf16', 'master_weight_type': 'fp32', 'seed': 42, 'validation_global_seed': 930, 'validation_interval': 0, 'validation_prompts': None, 'validate_video_length': 121, 'validation_aspect_ratio': '16:9', 'validation_fixed_size': '480x864', 'validation_num_inference_steps': 40, 'validation_negative_prompt': '', 'validation_enable_sr': False, 'validation_prompt_rewrite': False, 'validate_at_step0': False, 'resume_from_checkpoint': None, 'use_lora': False, 'lora_r': 8, 'lora_alpha': 16, 'lora_dropout': 0.0, 'lora_target_modules': None, 'pretrained_lora_path': None, 'train_video_csv': ['/nfs/ywang29/Reward_finetuning/SAGE-GRPO/assets/demo_train_decomposed_binary5.csv'], 'valid_video_csv': 'assets/demo_sample.csv', 'reward_model': 'snapvideo_vlm', 'reward_config': None, 'reward_checkpoint_mode': 'v3', 'reward_weights': '{"VQ":0.5,"MQ":0.5,"TA":1.0}', 'reward_gradient_credit': False, 'reward_credit_num_frames': 0, 'reward_credit_max_questions': 0, 'reward_credit_metric': 'weighted', 'token_credit_source_model': None, 'token_credit_advantage_source': 'token', 'token_credit_remote_url': None, 'token_credit_vlm_model_path': None, 'token_credit_vlm_model_family': None, 'token_credit_vlm_prompt_template': None, 'remote_reward_url': None, 'vlm_reward_model_path': '/nfs/ywang29/ckpts/Qwen3.5-9B', 'vlm_reward_model_family': 'qwen3_5', 'vlm_reward_prompt_template': None, 'vlm_reward_num_frames': 20, 'vlm_reward_max_pixels': 50176, 'vlm_reward_max_new_tokens': 128, 'vlm_reward_batch_size': 1, 'vlm_reward_score_type': 'token_credit', 'vlm_reward_missing_score': 0.0, 'vlm_reward_video_sample_type': 'uniform', 'vlm_reward_token_credit_max_questions': 5, 'token_credit_loss_mode': 'video_adv', 'token_credit_ratio_mode': 'weighted_logprob', 'token_credit_shuffle_weights': False, 'token_credit_spatial_granularity': 'window', 'token_credit_spatial_bins': 16, 'token_credit_spatial_window': 3, 'token_credit_chunk_size': 1, 'enable_grpo_policy_update': True, 'grpo_sampling_steps': 20, 'sample_n_frames': 81, 'eta': 0.5, 'cfg_scale': 6.0, 'neg_prompt': None, 'data_type': 'video', 'model_type': '16164', 'video_fps': 24, 'vae_name': '16164-32c-hy20250605', 'vae_tiling': False, 'infer_flow_shift_video': 5.0, 'video_bucket_hw_base_size': 480, 'video_bucket_hw_bucket_stride': 32, 'multitask_mask_training_type': 'concat', 'vision_num_semantic_tokens': 729, 'vision_states_dim': 1152, 'sde_type': 'sage_grpo', 'mini_batch_size_per_update': 1, 'clip_range': 0.0001, 'adv_clip_max': 5.0, 'timestep_fraction': 1.0, 'random_train_timestep_count': False, 'random_train_timestep_min': 1, 'random_train_timestep_max': 20, 'use_grad_balancing': True, 'enable_timestep_permutation': True, 'reference_mode_offload': True, 'kl_weight': 1e-05, 'kl_coef': 1e-07, 'kl_min_coef': 1e-07, 'use_moving_KL': True, 'update_ref_model_step': 10, 'use_dual_kl': True, 'dual_kl_moving_weight': 1.0, 'dual_kl_step_weight': 0.1, 'kl_compute_mode': 'rollout_phase', 'glyph_byT5_v2': True, 'byt5_max_length': 256, 'video_uncond_p_byt5': 0.0, 'precision': 'bf16', 'vae_precision': 'fp16', 'val_disable_autocast': False, 'debug_grad_flow': False, 'debug_train_diagnostics_interval': 1, '_wandb': {}}
+2026-06-29 18:40:45,207 INFO    MainThread:83423 [wandb_init.py:init():816] starting backend
+2026-06-29 18:40:45,429 INFO    MainThread:83423 [wandb_init.py:init():831] sending inform_init request
+2026-06-29 18:40:45,729 INFO    MainThread:83423 [wandb_init.py:init():836] backend started and connected
+2026-06-29 18:40:45,731 INFO    MainThread:83423 [wandb_init.py:init():906] updated telemetry
+2026-06-29 18:40:45,761 INFO    MainThread:83423 [wandb_init.py:init():929] communicating run to backend with 90.0 second timeout
+2026-06-29 18:40:46,248 INFO    MainThread:83423 [wandb_init.py:init():969] run resumed
+2026-06-29 18:40:46,249 INFO    MainThread:83423 [wandb_init.py:init():974] starting run threads in backend
+2026-06-29 18:40:46,450 INFO    MainThread:83423 [wandb_run.py:_console_start():2523] atexit reg
+2026-06-29 18:40:46,451 INFO    MainThread:83423 [wandb_run.py:_redirect():2373] redirect: wrap_raw
+2026-06-29 18:40:46,451 INFO    MainThread:83423 [wandb_run.py:_redirect():2442] Wrapping output streams.
+2026-06-29 18:40:46,451 INFO    MainThread:83423 [wandb_run.py:_redirect():2465] Redirects installed.
+2026-06-29 18:40:46,452 INFO    MainThread:83423 [wandb_init.py:init():1012] run started, returning control to user process

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260619_013012-rgt7q8d2/files/requirements.txt

@@ -0,0 +1,284 @@
+awscli==1.45.25
+boto3==1.43.25
+botocore==1.43.25
+certifi==2026.5.20
+charset-normalizer==3.4.7
+colorama==0.4.6
+docutils==0.19
+idna==3.18
+jmespath==1.1.0
+pip==26.1.2
+pyasn1==0.6.3
+python-dateutil==2.9.0.post0
+PyYAML==6.0.3
+requests==2.34.2
+rsa==4.7.2
+s3transfer==0.18.0
+six==1.17.0
+urllib3==2.7.0
+py-cpuinfo==9.0.0
+hjson==3.1.0
+easydict==1.13
+braceexpand==0.1.7
+antlr4-python3-runtime==4.9.3
+websocket-client==1.9.0
+webdataset==1.0.2
+triton==3.6.0
+torchcodec==0.14.0+cu130
+tifffile==2026.6.1
+tenacity==9.1.4
+smmap==5.0.3
+sentry-sdk==2.63.0
+sentencepiece==0.2.1
+regex==2026.5.9
+pyarrow==24.0.0
+pyaml==26.2.1
+propcache==0.5.2
+opencv-python-headless==4.13.0.92
+omegaconf==2.3.1
+nvidia-nvtx==13.0.85
+ImageIO==2.37.3
+nvidia-nvshmem-cu13==3.4.5
+nvidia-nvjitlink==13.0.88
+av==17.1.0
+nvidia-nccl-cu13==2.28.9
+nvidia-cufile==1.15.1.6
+imageio-ffmpeg==0.6.0
+nvidia-cuda-runtime==13.0.96
+nvidia-cuda-nvrtc==13.0.88
+attrs==26.1.0
+nvidia-cuda-cupti==13.0.85
+nvidia-cublas==13.1.0.3
+multidict==6.7.1
+msgpack==1.2.1
+Mako==1.3.12
+loguru==0.7.3
+lmdb==2.2.1
+lightning-utilities==0.15.3
+lazy-loader==0.5
+greenlet==3.5.2
+future==1.0.0
+ftfy==6.3.1
+frozenlist==1.8.0
+frozendict==2.4.7
+docstring_parser==0.18.0
+decord==0.6.0
+cuda-pathfinder==1.5.5
+aiohappyeyeballs==2.6.2
+yarl==1.24.2
+typed-argument-parser==1.12.0
+SQLAlchemy==2.0.51
+scikit-image==0.26.0
+qwen-vl-utils==0.0.14
+typeguard==4.5.2
+nvidia-cusparse==12.6.3.3
+nvidia-cufft==12.0.0.61
+nvidia-cudnn-cu13==9.15.1.9
+gitdb==4.0.12
+ffmpeg-python==0.2.0
+einx==0.4.3
+cupy-cuda12x==14.1.1
+cuda-bindings==13.0.3
+cattrs==26.1.0
+aiosignal==1.4.0
+tokenizers==0.22.2
+termcolor==3.3.0
+nvidia-cusolver==12.0.4.66
+mediapy==1.2.6
+ipdb==0.13.13
+GitPython==3.1.50
+diffusers==0.38.0
+coremltools==9.0
+alembic==1.18.4
+aiohttp==3.14.1
+wandb==0.27.2
+torch==2.10.0+cu130
+dataset==2.0.0
+dashscope==1.25.23
+torchvision==0.25.0+cu130
+torchmetrics==1.9.0
+sniffio==1.3.1
+torchdata==0.8.0
+tqdm==4.67.3
+torchaudio==2.10.0+cu130
+deepspeed==0.19.2
+timm==1.0.27
+pytorch-lightning==2.6.5
+peft==0.19.1
+lpips==0.1.4
+lightning==2.6.5
+flash_attn==2.8.3+cu130torch2.10
+xxhash==3.7.0
+shortuuid==1.0.13
+setuptools==82.0.0
+safetensors==0.7.0
+pillow==12.1.1
+numpy==2.4.2
+jiter==0.15.0
+fsspec==2026.2.0
+distro==1.9.0
+dill==0.4.1
+tyro==1.0.14
+tiktoken==0.13.0
+multiprocess==0.70.19
+modelscope==1.36.0
+fire==0.7.0
+openai==2.8.0
+datasets==4.8.4
+trl==0.8.6
+angelslim==0.2.2
+huggingface_hub==1.20.1
+transformers==5.12.1
+black==26.5.1
+pathspec==1.1.1
+pytokens==0.4.1
+importlib_metadata==9.0.0
+ml_dtypes==0.5.4
+onnx==1.21.0
+onnx-ir==0.2.1
+onnxscript==0.7.0
+transformer_engine==2.9.0+70f53666
+zipp==4.1.0
+nvidia-mathdx==25.6.0
+einops==0.8.2
+absl-py==2.4.0
+accelerate==1.14.0
+annotated-doc==0.0.4
+annotated-types==0.7.0
+anyio==4.13.0
+apsw==3.53.2.0
+apswutils==0.1.2
+beartype==0.22.9
+beautifulsoup4==4.15.0
+blis==1.2.1
+catalogue==2.0.10
+cloudpathlib==0.24.0
+cloudpickle==3.1.2
+confection==0.1.5
+contourpy==1.3.3
+cycler==0.12.1
+cymem==2.0.13
+fastai==2.8.7
+fastcore==1.13.3
+fastdownload==0.0.7
+fastlite==0.2.4
+fastprogress==1.1.6
+fasttransform==0.0.2
+fonttools==4.63.0
+grpcio==1.81.1
+h11==0.16.0
+hf-xet==1.5.1
+httpcore==1.0.9
+httptools==0.8.0
+httpx==0.28.1
+itsdangerous==2.2.0
+joblib==1.5.3
+kiwisolver==1.5.0
+langcodes==3.5.1
+Markdown==3.10.2
+markdown-it-py==4.2.0
+matplotlib==3.10.9
+mdurl==0.1.2
+mpmath==1.3.0
+murmurhash==1.0.15
+mypy_extensions==1.1.0
+narwhals==2.22.1
+networkx==3.6.1
+nvidia-curand==10.4.0.35
+nvidia-cusparselt-cu13==0.8.0
+oauthlib==3.3.1
+pandas==3.0.3
+plum-dispatch==2.9.0
+preshed==3.0.13
+protobuf==7.35.0
+pydantic==2.13.4
+pydantic_core==2.46.4
+pyparsing==3.3.2
+pyre-extensions==0.0.32
+python-dotenv==1.2.2
+python-fasthtml==0.14.2
+python-multipart==0.0.32
+rich==15.0.0
+s3torchconnector==1.5.0
+s3torchconnectorclient==1.5.0
+scikit-learn==1.9.0
+shellingham==1.5.4
+smart_open==7.6.1
+soupsieve==2.8.4
+spacy==3.8.7
+spacy-legacy==3.0.12
+spacy-loggers==1.0.5
+srsly==2.5.3
+starlette==1.3.0
+sympy==1.14.0
+tabulate==0.10.0
+tensorboard==2.20.0
+tensorboard-data-server==0.7.2
+thinc==8.3.4
+threadpoolctl==3.6.0
+torchtnt==0.2.4
+typer==0.25.1
+typer-slim==0.24.0
+typing-inspect==0.9.0
+typing-inspection==0.4.2
+uvicorn==0.49.0
+uvloop==0.22.1
+wasabi==1.1.3
+watchfiles==1.2.0
+weasel==0.4.3
+websockets==16.0
+Werkzeug==3.1.8
+wrapt==2.2.1
+asttokens==3.0.1
+cffi==2.0.0
+click==8.4.1
+comm==0.2.3
+cryptography==48.0.1
+Cython==3.2.5
+debugpy==1.8.21
+decorator==5.3.1
+executing==2.2.1
+filelock==3.29.3
+h5py==3.16.0
+intel-cmplr-lib-ur==2026.0.0
+intel-openmp==2026.0.0
+ipykernel==7.3.0
+ipython==9.14.1
+ipython_pygments_lexers==1.1.1
+jedi==0.20.0
+Jinja2==3.1.6
+jupyter_client==8.9.1
+jupyter_core==5.9.1
+MarkupSafe==3.0.3
+matplotlib-inline==0.2.2
+mkl==2026.0.0
+mkl-include==2026.0.0
+mpi4py==4.1.2
+nest-asyncio2==1.7.2
+ninja==1.13.0
+onemkl-license==2026.0.0
+opencv-python==4.11.0.86
+packaging==26.2
+parso==0.8.7
+pexpect==4.9.0
+platformdirs==4.10.0
+prompt_toolkit==3.0.52
+psutil==7.2.2
+ptyprocess==0.7.0
+pure_eval==0.2.3
+pybind11==3.0.4
+pycparser==3.0
+Pygments==2.20.0
+pyOpenSSL==26.2.0
+pyzmq==27.1.0
+scipy==1.17.1
+stack-data==0.6.3
+tbb==2023.0.0
+tcmlib==1.5.0
+tornado==6.5.7
+traitlets==5.15.1
+typing_extensions==4.15.0
+tzdata==2026.2
+umf==1.1.0
+wheel==0.47.0
+wcwidth==0.8.1

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260619_013012-rgt7q8d2/files/wandb-metadata.json

@@ -0,0 +1,183 @@
+{
+  "os": "Linux-6.1.163-186.299.amzn2023.x86_64-x86_64-with-glibc2.35",
+  "python": "CPython 3.12.10",
+  "startedAt": "2026-06-19T01:30:12.429129Z",
+  "args": [
+    "--pretrained_model_root",
+    "./ckpts",
+    "--train_video_csv",
+    "/nfs/ywang29/Reward_finetuning/SAGE-GRPO/assets/demo_train_decomposed_binary5.csv",
+    "--reward_model",
+    "snapvideo_vlm",
+    "--vlm_reward_model_path",
+    "/nfs/ywang29/ckpts/Qwen3.5-9B",
+    "--vlm_reward_model_family",
+    "qwen3_5",
+    "--vlm_reward_num_frames",
+    "20",
+    "--vlm_reward_max_pixels",
+    "50176",
+    "--vlm_reward_max_new_tokens",
+    "128",
+    "--vlm_reward_batch_size",
+    "1",
+    "--vlm_reward_score_type",
+    "token_credit",
+    "--vlm_reward_token_credit_max_questions",
+    "5",
+    "--token_credit_loss_mode",
+    "video_adv",
+    "--token_credit_ratio_mode",
+    "weighted_logprob",
+    "--token_credit_spatial_granularity",
+    "window",
+    "--token_credit_spatial_bins",
+    "16",
+    "--token_credit_spatial_window",
+    "3",
+    "--token_credit_chunk_size",
+    "1",
+    "--learning_rate",
+    "1e-5",
+    "--batch_size",
+    "2",
+    "--num_generations",
+    "4",
+    "--max_steps",
+    "10000",
+    "--output_dir",
+    "./outputs",
+    "--enable_wandb",
+    "True",
+    "--project_name",
+    "SAGE-GRPO",
+    "--run_name",
+    "yw006-4-sage-qwen35-9b-token-credit-video-adv-window3",
+    "--enable_fsdp",
+    "--enable_gradient_checkpointing",
+    "--sp_size",
+    "2",
+    "--reward_checkpoint_mode",
+    "v3",
+    "--validation_interval",
+    "0",
+    "--validate_at_step0",
+    "False",
+    "--validate_video_length",
+    "121",
+    "--validation_timestep_shift",
+    "5.0",
+    "--use_grad_balancing",
+    "True",
+    "--enable_timestep_permutation",
+    "True",
+    "--sde_type",
+    "sage_grpo",
+    "--kl_weight",
+    "1e-5",
+    "--kl_coef",
+    "1e-7",
+    "--use_moving_KL",
+    "True",
+    "--update_ref_model_step",
+    "10",
+    "--use_dual_kl",
+    "True",
+    "--dual_kl_moving_weight",
+    "1.0",
+    "--dual_kl_step_weight",
+    "0.1",
+    "--reference_mode_offload",
+    "True",
+    "--save_interval",
+    "25",
+    "--timer_interval",
+    "0",
+    "--checkpoints_directory",
+    "s3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3"
+  ],
+  "program": "/nfs/ywang29/Reward_finetuning/SAGE-GRPO/post_train.py",
+  "codePath": "post_train.py",
+  "codePathLocal": "post_train.py",
+  "git": {
+    "remote": "https://github.com/Tencent-Hunyuan/SAGE-GRPO.git",
+    "commit": "779776e42732e0262b3f2844dd206d9d8f043b2f"
+  },
+  "root": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb",
+  "host": "ywang29-hunyuan-sage-qwen35-9b-tcva-w3-worker-0",
+  "username": "yi-fan-wang1216",
+  "executable": "/usr/local/bin/python",
+  "cpu_count": 48,
+  "cpu_count_logical": 96,
+  "gpu": "NVIDIA A100-SXM4-80GB",
+  "gpu_count": 8,
+  "disk": {
+    "/": {
+      "total": "7516777873408",
+      "used": "454212927488"
+    }
+  },
+  "memory": {
+    "total": "1204539260928"
+  },
+  "gpu_nvidia": [
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-a46a9236-2f10-8dc6-a65a-3acdc400ac53"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-35a5b590-e4bb-5a60-c86e-be2ca8e30206"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-4699434d-8ca7-e16d-2126-ceed42408c3f"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-4be9caf6-8ba6-807b-758c-ba5e7a168c54"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-21bce20d-0690-445f-66e6-8f847e18c310"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-f9535ed2-1580-393b-3675-70fdbea1e453"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-c87db964-ba6d-cfff-e0c6-aa375f8ba349"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-0f3ed4a8-a0d0-63f3-7235-7ae048428d18"
+    }
+  ],
+  "cudaVersion": "13.0",
+  "writerId": "40amcdecq7vryezfaa3ad7uw3lf5slgv"
+}

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260619_013012-rgt7q8d2/logs/debug.log

@@ -0,0 +1,19 @@
+2026-06-19 01:30:12,450 INFO    MainThread:2055 [wandb_setup.py:_flush():81] Current SDK version is 0.27.2
+2026-06-19 01:30:12,451 INFO    MainThread:2055 [wandb_setup.py:_flush():81] Configure stats pid to 2055
+2026-06-19 01:30:12,451 INFO    MainThread:2055 [wandb_setup.py:_flush():81] Loading settings from environment variables
+2026-06-19 01:30:12,451 INFO    MainThread:2055 [wandb_init.py:setup_run_log_directory():723] Logging user logs to outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260619_013012-rgt7q8d2/logs/debug.log
+2026-06-19 01:30:12,452 INFO    MainThread:2055 [wandb_init.py:setup_run_log_directory():724] Logging internal logs to outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260619_013012-rgt7q8d2/logs/debug-internal.log
+2026-06-19 01:30:12,452 INFO    MainThread:2055 [wandb_init.py:init():766] calling init triggers
+2026-06-19 01:30:12,452 INFO    MainThread:2055 [wandb_init.py:init():771] wandb.init called with sweep_config: {}
+config: {'pretrained_model_root': './ckpts', 'pretrained_transformer_version': '480p_t2v', 'post_train_type': 'grpo', 'learning_rate': 1e-05, 'weight_decay': 0.01, 'max_steps': 10000, 'warmup_steps': 500, 'gradient_accumulation_steps': 1, 'max_grad_norm': 1.0, 'use_muon': True, 'num_train_timesteps': 1000, 'train_timestep_shift': 3.0, 'validation_timestep_shift': 5.0, 'snr_type': 'lognorm', 'task_type': 't2v', 'i2v_prob': 0.3, 'enable_fsdp': True, 'enable_gradient_checkpointing': True, 'sp_size': 2, 'dp_replicate': 1, 'batch_size': 2, 'num_workers': 4, 'prefetch_factor': 2, 'num_generations': 4, 'use_same_noise': False, 'output_dir': './outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3', 'checkpoints_directory': 's3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3', 'save_interval': 25, 'timer_interval': 0, 'auto_resume_from_timer': True, 'log_interval': 1, 'enable_wandb': True, 'project_name': 'SAGE-GRPO', 'run_name': 'yw006-4-sage-qwen35-9b-token-credit-video-adv-window3', 'wandb_key': None, 'wandb_username': None, 'wandb_base_url': 'https://api.wandb.ai', 'wandb_entity': None, 'wandb_silent': True, 'dtype': 'bf16', 'master_weight_type': 'fp32', 'seed': 42, 'validation_global_seed': 930, 'validation_interval': 0, 'validation_prompts': None, 'validate_video_length': 121, 'validation_aspect_ratio': '16:9', 'validation_fixed_size': '480x864', 'validation_num_inference_steps': 40, 'validation_negative_prompt': '', 'validation_enable_sr': False, 'validation_prompt_rewrite': False, 'validate_at_step0': False, 'resume_from_checkpoint': None, 'use_lora': False, 'lora_r': 8, 'lora_alpha': 16, 'lora_dropout': 0.0, 'lora_target_modules': None, 'pretrained_lora_path': None, 'train_video_csv': ['/nfs/ywang29/Reward_finetuning/SAGE-GRPO/assets/demo_train_decomposed_binary5.csv'], 'valid_video_csv': 'assets/demo_sample.csv', 'reward_model': 'snapvideo_vlm', 'reward_config': None, 'reward_checkpoint_mode': 'v3', 'reward_weights': '{"VQ":0.5,"MQ":0.5,"TA":1.0}', 'reward_gradient_credit': False, 'reward_credit_num_frames': 0, 'reward_credit_max_questions': 0, 'reward_credit_metric': 'weighted', 'token_credit_source_model': None, 'token_credit_advantage_source': 'token', 'token_credit_remote_url': None, 'remote_reward_url': None, 'vlm_reward_model_path': '/nfs/ywang29/ckpts/Qwen3.5-9B', 'vlm_reward_model_family': 'qwen3_5', 'vlm_reward_prompt_template': None, 'vlm_reward_num_frames': 20, 'vlm_reward_max_pixels': 50176, 'vlm_reward_max_new_tokens': 128, 'vlm_reward_batch_size': 1, 'vlm_reward_score_type': 'token_credit', 'vlm_reward_missing_score': 0.0, 'vlm_reward_video_sample_type': 'uniform', 'vlm_reward_token_credit_max_questions': 5, 'token_credit_loss_mode': 'video_adv', 'token_credit_ratio_mode': 'weighted_logprob', 'token_credit_shuffle_weights': False, 'token_credit_spatial_granularity': 'window', 'token_credit_spatial_bins': 16, 'token_credit_spatial_window': 3, 'token_credit_chunk_size': 1, 'enable_grpo_policy_update': True, 'grpo_sampling_steps': 20, 'sample_n_frames': 81, 'eta': 0.5, 'cfg_scale': 6.0, 'neg_prompt': None, 'data_type': 'video', 'model_type': '16164', 'video_fps': 24, 'vae_name': '16164-32c-hy20250605', 'vae_tiling': False, 'infer_flow_shift_video': 5.0, 'video_bucket_hw_base_size': 480, 'video_bucket_hw_bucket_stride': 32, 'multitask_mask_training_type': 'concat', 'vision_num_semantic_tokens': 729, 'vision_states_dim': 1152, 'sde_type': 'sage_grpo', 'mini_batch_size_per_update': 1, 'clip_range': 0.0001, 'adv_clip_max': 5.0, 'timestep_fraction': 1.0, 'random_train_timestep_count': False, 'random_train_timestep_min': 1, 'random_train_timestep_max': 20, 'use_grad_balancing': True, 'enable_timestep_permutation': True, 'reference_mode_offload': True, 'kl_weight': 1e-05, 'kl_coef': 1e-07, 'kl_min_coef': 1e-07, 'use_moving_KL': True, 'update_ref_model_step': 10, 'use_dual_kl': True, 'dual_kl_moving_weight': 1.0, 'dual_kl_step_weight': 0.1, 'kl_compute_mode': 'rollout_phase', 'glyph_byT5_v2': True, 'byt5_max_length': 256, 'video_uncond_p_byt5': 0.0, 'precision': 'bf16', 'vae_precision': 'fp16', 'val_disable_autocast': False, 'debug_grad_flow': False, 'debug_train_diagnostics_interval': 1, '_wandb': {}}
+2026-06-19 01:30:12,452 INFO    MainThread:2055 [wandb_init.py:init():814] starting backend
+2026-06-19 01:30:12,676 INFO    MainThread:2055 [wandb_init.py:init():829] sending inform_init request
+2026-06-19 01:30:12,932 INFO    MainThread:2055 [wandb_init.py:init():834] backend started and connected
+2026-06-19 01:30:12,935 INFO    MainThread:2055 [wandb_init.py:init():904] updated telemetry
+2026-06-19 01:30:12,966 INFO    MainThread:2055 [wandb_init.py:init():927] communicating run to backend with 90.0 second timeout
+2026-06-19 01:30:13,420 INFO    MainThread:2055 [wandb_init.py:init():972] starting run threads in backend
+2026-06-19 01:30:13,612 INFO    MainThread:2055 [wandb_run.py:_console_start():2523] atexit reg
+2026-06-19 01:30:13,612 INFO    MainThread:2055 [wandb_run.py:_redirect():2373] redirect: wrap_raw
+2026-06-19 01:30:13,612 INFO    MainThread:2055 [wandb_run.py:_redirect():2442] Wrapping output streams.
+2026-06-19 01:30:13,612 INFO    MainThread:2055 [wandb_run.py:_redirect():2465] Redirects installed.
+2026-06-19 01:30:13,614 INFO    MainThread:2055 [wandb_init.py:init():1010] run started, returning control to user process

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260629_013009-rgt7q8d2/files/output.log

@@ -0,0 +1,133 @@
+2026-06-29 01:30:10.649 | INFO     | __main__:print_training_configuration:970 - SYSTEM & HARDWARE CONFIGURATION
+2026-06-29 01:30:10.649 | INFO     | __main__:print_training_configuration:971 - --------------------------------------------------------------------------------
+2026-06-29 01:30:10.650 | INFO     | __main__:print_training_configuration:972 - Number of GPUs                                   : 64
+2026-06-29 01:30:10.650 | INFO     | __main__:print_training_configuration:973 - Local rank                                       : 0
+2026-06-29 01:30:10.651 | INFO     | __main__:print_training_configuration:974 - Global rank                                      : 0
+2026-06-29 01:30:10.651 | INFO     | __main__:print_training_configuration:975 - Device                                           : cuda:0
+2026-06-29 01:30:10.651 | INFO     | __main__:print_training_configuration:976 - Master weight dtype                              : torch.float32
+2026-06-29 01:30:10.652 | INFO     | __main__:print_training_configuration:978 - --------------------------------------------------------------------------------
+2026-06-29 01:30:10.652 | INFO     | __main__:print_training_configuration:979 - MODEL PARAMETERS & ARCHITECTURE
+2026-06-29 01:30:10.653 | INFO     | __main__:print_training_configuration:980 - --------------------------------------------------------------------------------
+2026-06-29 01:30:10.664 | INFO     | __main__:print_training_configuration:983 - Number of total                                  : 8,326,608,160
+2026-06-29 01:30:10.664 | INFO     | __main__:print_training_configuration:984 - Total trainable parameters                        : 8,326,608,160
+2026-06-29 01:30:10.665 | INFO     | __main__:print_training_configuration:985 - DIT model root                                   : ./ckpts
+2026-06-29 01:30:10.665 | INFO     | __main__:print_training_configuration:986 - Transformer version                              : 480p_t2v
+2026-06-29 01:30:10.665 | INFO     | __main__:print_training_configuration:987 - Attention mode                                   : flash
+2026-06-29 01:30:10.666 | INFO     | __main__:print_training_configuration:988 - Gradient checkpointing                           : True
+2026-06-29 01:30:10.666 | INFO     | __main__:print_training_configuration:989 - FSDP enabled                                     : True
+2026-06-29 01:30:10.667 | INFO     | __main__:print_training_configuration:990 - LoRA enabled                                     : False
+2026-06-29 01:30:10.667 | INFO     | __main__:print_training_configuration:991 - Glyph byT5 v2                                    : True
+2026-06-29 01:30:10.668 | INFO     | __main__:print_training_configuration:992 - ByT5 max length                                  : 256
+2026-06-29 01:30:10.668 | INFO     | __main__:print_training_configuration:994 - --------------------------------------------------------------------------------
+2026-06-29 01:30:10.668 | INFO     | __main__:print_training_configuration:995 - DISTRIBUTED TRAINING CONFIGURATION
+2026-06-29 01:30:10.669 | INFO     | __main__:print_training_configuration:996 - --------------------------------------------------------------------------------
+2026-06-29 01:30:10.669 | INFO     | __main__:print_training_configuration:997 - World size                                       : 64
+2026-06-29 01:30:10.670 | INFO     | __main__:print_training_configuration:998 - DP degree                                        : 32
+2026-06-29 01:30:10.670 | INFO     | __main__:print_training_configuration:999 - DP rank                                          : 0
+2026-06-29 01:30:10.670 | INFO     | __main__:print_training_configuration:1000 - SP size                                          : 2
+2026-06-29 01:30:10.671 | INFO     | __main__:print_training_configuration:1001 - SP rank                                          : 0
+2026-06-29 01:30:10.671 | INFO     | __main__:print_training_configuration:1003 - --------------------------------------------------------------------------------
+2026-06-29 01:30:10.672 | INFO     | __main__:print_training_configuration:1004 - BATCH SIZE & DATA FLOW CONFIGURATION
+2026-06-29 01:30:10.672 | INFO     | __main__:print_training_configuration:1005 - --------------------------------------------------------------------------------
+2026-06-29 01:30:10.672 | INFO     | __main__:print_training_configuration:1006 - Per GPU batch size                               : 2
+2026-06-29 01:30:10.673 | INFO     | __main__:print_training_configuration:1007 - DP degree batch size                             : 64
+2026-06-29 01:30:10.673 | INFO     | __main__:print_training_configuration:1008 - Gradient accumulation steps                      : 1
+2026-06-29 01:30:10.674 | INFO     | __main__:print_training_configuration:1009 - Num generations                                  : 4
+2026-06-29 01:30:10.674 | INFO     | __main__:print_training_configuration:1010 - Use same noise                                   : False
+2026-06-29 01:30:10.674 | INFO     | __main__:print_training_configuration:1011 - Num workers                                      : 4
+2026-06-29 01:30:10.675 | INFO     | __main__:print_training_configuration:1012 - Prefetch factor                                  : 2
+2026-06-29 01:30:10.675 | INFO     | __main__:print_training_configuration:1014 - --------------------------------------------------------------------------------
+2026-06-29 01:30:10.675 | INFO     | __main__:print_training_configuration:1015 - OPTIMIZATION CONFIGURATION
+2026-06-29 01:30:10.676 | INFO     | __main__:print_training_configuration:1016 - --------------------------------------------------------------------------------
+2026-06-29 01:30:10.676 | INFO     | __main__:print_training_configuration:1017 - Optimizer                                        : muon
+2026-06-29 01:30:10.677 | INFO     | __main__:print_training_configuration:1018 - Learning rate                                    : 1e-05
+2026-06-29 01:30:10.677 | INFO     | __main__:print_training_configuration:1019 - Weight decay                                     : 0.01
+2026-06-29 01:30:10.677 | INFO     | __main__:print_training_configuration:1020 - Max gradient norm                                : 1.0
+2026-06-29 01:30:10.678 | INFO     | __main__:print_training_configuration:1021 - LR scheduler                                     : constant
+2026-06-29 01:30:10.678 | INFO     | __main__:print_training_configuration:1022 - Warmup steps                                     : 500
+2026-06-29 01:30:10.679 | INFO     | __main__:print_training_configuration:1023 - Max training steps                               : 10000
+2026-06-29 01:30:10.679 | INFO     | __main__:print_training_configuration:1025 - --------------------------------------------------------------------------------
+2026-06-29 01:30:10.679 | INFO     | __main__:print_training_configuration:1026 - INFERENCE & VALIDATION CONFIGURATION
+2026-06-29 01:30:10.680 | INFO     | __main__:print_training_configuration:1027 - --------------------------------------------------------------------------------
+2026-06-29 01:30:10.680 | INFO     | __main__:print_training_configuration:1028 - Validation interval                              : 0
+2026-06-29 01:30:10.681 | INFO     | __main__:print_training_configuration:1029 - Validate at step 0                               : False
+2026-06-29 01:30:10.681 | INFO     | __main__:print_training_configuration:1030 - Validation video length                          : 121
+2026-06-29 01:30:10.681 | INFO     | __main__:print_training_configuration:1031 - Validation global seed                           : 930
+2026-06-29 01:30:10.682 | INFO     | __main__:print_training_configuration:1032 - Validation inference steps                        : 40
+2026-06-29 01:30:10.682 | INFO     | __main__:print_training_configuration:1033 - Validation aspect ratio                          : 16:9
+2026-06-29 01:30:10.683 | INFO     | __main__:print_training_configuration:1034 - Validation fixed size                            : 480x864
+2026-06-29 01:30:10.683 | INFO     | __main__:print_training_configuration:1035 - Validation prompt rewrite                         : False
+2026-06-29 01:30:10.683 | INFO     | __main__:print_training_configuration:1036 - Validation flow shift                             : 5.0
+2026-06-29 01:30:10.684 | INFO     | __main__:print_training_configuration:1037 - Sampling CFG scale                                : 6.0
+2026-06-29 01:30:10.684 | INFO     | __main__:print_training_configuration:1038 - Sampling eta                                      : 0.5
+2026-06-29 01:30:10.684 | INFO     | __main__:print_training_configuration:1039 - Sampling steps (GRPO rollout)                     : 20
+2026-06-29 01:30:10.685 | INFO     | __main__:print_training_configuration:1040 - Token credit loss mode                            : video_adv
+2026-06-29 01:30:10.685 | INFO     | __main__:print_training_configuration:1041 - Token credit ratio mode                           : weighted_logprob
+2026-06-29 01:30:10.686 | INFO     | __main__:print_training_configuration:1042 - Token credit shuffle weights                      : False
+2026-06-29 01:30:10.686 | INFO     | __main__:print_training_configuration:1043 - Token credit spatial granularity                  : window
+2026-06-29 01:30:10.686 | INFO     | __main__:print_training_configuration:1044 - Token credit spatial bins                         : 16
+2026-06-29 01:30:10.687 | INFO     | __main__:print_training_configuration:1045 - Token credit spatial window                       : 3
+2026-06-29 01:30:10.687 | INFO     | __main__:print_training_configuration:1046 - Token credit chunk size                           : 1
+2026-06-29 01:30:10.688 | INFO     | __main__:print_training_configuration:1047 - Token credit source model                         : None
+2026-06-29 01:30:10.688 | INFO     | __main__:print_training_configuration:1048 - Token credit advantage source                     : token
+2026-06-29 01:30:10.688 | INFO     | __main__:print_training_configuration:1049 - Token credit VLM model path                       : None
+2026-06-29 01:30:10.689 | INFO     | __main__:print_training_configuration:1050 - Token credit VLM model family                     : None
+2026-06-29 01:30:10.689 | INFO     | __main__:print_training_configuration:1051 - Random train timestep count                       : False
+2026-06-29 01:30:10.690 | INFO     | __main__:print_training_configuration:1052 - Random train timestep range                       : 1-20
+2026-06-29 01:30:10.690 | INFO     | __main__:print_training_configuration:1056 - SDE type                                          : sage_grpo
+2026-06-29 01:30:10.690 | INFO     | __main__:print_training_configuration:1058 - --------------------------------------------------------------------------------
+2026-06-29 01:30:10.691 | INFO     | __main__:print_training_configuration:1059 - GRPO & KL CONFIGURATION
+2026-06-29 01:30:10.691 | INFO     | __main__:print_training_configuration:1060 - --------------------------------------------------------------------------------
+2026-06-29 01:30:10.691 | INFO     | __main__:print_training_configuration:1061 - Reward model                                     : snapvideo_vlm
+2026-06-29 01:30:10.692 | INFO     | __main__:print_training_configuration:1062 - Reward gradient credit                          : False
+2026-06-29 01:30:10.692 | INFO     | __main__:print_training_configuration:1063 - Reward credit frames                            : 0
+2026-06-29 01:30:10.693 | INFO     | __main__:print_training_configuration:1064 - Reward credit max questions                     : 0
+2026-06-29 01:30:10.693 | INFO     | __main__:print_training_configuration:1065 - Reward credit metric                            : weighted
+2026-06-29 01:30:10.694 | INFO     | __main__:print_training_configuration:1067 - VLM reward model path                            : /nfs/ywang29/ckpts/Qwen3.5-9B
+2026-06-29 01:30:10.694 | INFO     | __main__:print_training_configuration:1068 - VLM reward model family                          : qwen3_5
+2026-06-29 01:30:10.694 | INFO     | __main__:print_training_configuration:1069 - VLM reward frames                                : 20
+2026-06-29 01:30:10.695 | INFO     | __main__:print_training_configuration:1070 - VLM reward max pixels                            : 50176
+2026-06-29 01:30:10.695 | INFO     | __main__:print_training_configuration:1071 - VLM reward score type                            : token_credit
+2026-06-29 01:30:10.695 | INFO     | __main__:print_training_configuration:1072 - VLM reward max new tokens                        : 128
+2026-06-29 01:30:10.696 | INFO     | __main__:print_training_configuration:1073 - VLM reward batch size                            : 1
+2026-06-29 01:30:10.696 | INFO     | __main__:print_training_configuration:1074 - VLM reward missing score                         : 0.0
+2026-06-29 01:30:10.697 | INFO     | __main__:print_training_configuration:1077 - KL weight                                        : 1e-05
+2026-06-29 01:30:10.697 | INFO     | __main__:print_training_configuration:1078 - KL coef                                          : 1e-07
+2026-06-29 01:30:10.697 | INFO     | __main__:print_training_configuration:1079 - Use moving KL                                    : True
+2026-06-29 01:30:10.698 | INFO     | __main__:print_training_configuration:1080 - Update ref model step                            : 10
+2026-06-29 01:30:10.698 | INFO     | __main__:print_training_configuration:1081 - Use dual KL                                      : True
+2026-06-29 01:30:10.699 | INFO     | __main__:print_training_configuration:1082 - Dual KL moving weight                            : 1.0
+2026-06-29 01:30:10.699 | INFO     | __main__:print_training_configuration:1083 - Dual KL step weight                              : 0.1
+2026-06-29 01:30:10.700 | INFO     | __main__:print_training_configuration:1084 - KL compute mode                                  : rollout_phase
+2026-06-29 01:30:10.700 | INFO     | __main__:print_training_configuration:1085 - Use grad balancing                               : True
+2026-06-29 01:30:10.701 | INFO     | __main__:print_training_configuration:1086 - Enable timestep permutation                      : True
+2026-06-29 01:30:10.701 | INFO     | __main__:print_training_configuration:1087 - Reference mode offload                           : True
+2026-06-29 01:30:10.701 | INFO     | __main__:print_training_configuration:1089 - --------------------------------------------------------------------------------
+2026-06-29 01:30:10.702 | INFO     | __main__:print_training_configuration:1090 - LOGGING & CHECKPOINTING CONFIGURATION
+2026-06-29 01:30:10.702 | INFO     | __main__:print_training_configuration:1091 - --------------------------------------------------------------------------------
+2026-06-29 01:30:10.703 | INFO     | __main__:print_training_configuration:1092 - Output directory                                 : ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3
+2026-06-29 01:30:10.703 | INFO     | __main__:print_training_configuration:1093 - Checkpoints directory                            : s3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3
+2026-06-29 01:30:10.703 | INFO     | __main__:print_training_configuration:1094 - Checkpoint interval                              : 25
+2026-06-29 01:30:10.704 | INFO     | __main__:print_training_configuration:1095 - Timer checkpoint interval                        : 0
+2026-06-29 01:30:10.704 | INFO     | __main__:print_training_configuration:1096 - Auto resume from timer/checkpoint                : True
+2026-06-29 01:30:10.705 | INFO     | __main__:print_training_configuration:1097 - Timer checkpoint directory                       : s3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/timer
+2026-06-29 01:30:10.705 | INFO     | __main__:print_training_configuration:1098 - Log interval                                     : 1
+2026-06-29 01:30:10.705 | INFO     | __main__:print_training_configuration:1099 - Resume checkpoint                                : None
+2026-06-29 01:30:10.706 | INFO     | __main__:print_training_configuration:1100 - Train CSV                                         : ['/nfs/ywang29/Reward_finetuning/SAGE-GRPO/assets/demo_train_decomposed_binary5.csv']
+2026-06-29 01:30:10.706 | INFO     | __main__:print_training_configuration:1101 - Valid CSV                                         : assets/demo_sample.csv
+2026-06-29 01:30:10.707 | INFO     | __main__:print_training_configuration:1102 - Seed                                             : 42
+2026-06-29 01:30:10.707 | INFO     | __main__:print_training_configuration:1103 - WandB enabled                                    : True
+2026-06-29 01:30:10.707 | INFO     | __main__:print_training_configuration:1105 - WandB project                                    : SAGE-GRPO
+2026-06-29 01:30:10.708 | INFO     | __main__:print_training_configuration:1106 - WandB run name                                   : yw006-4-sage-qwen35-9b-token-credit-video-adv-window3
+2026-06-29 01:30:10.708 | INFO     | __main__:__init__:958 - [Effective Config] world_size=64, rank=0, local_rank=0, sp_size=2, dp_size=32, dp_rank=0, sp_rank=0, batch_size=2, num_generations=4, grad_accum=1
+2026-06-29 01:30:10.709 | INFO     | __main__:_build_video_dataset:2088 - [Effective GRPO Grouping] video_batch_size=2, dp_replicas=32, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, product=64
+2026-06-29 01:30:11.375 | INFO     | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1
+2026-06-29 01:30:11.376 | INFO     | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done
+2026-06-29 01:30:11.377 | INFO     | __main__:_train_grpo:2171 - Starting GRPO training...
+2026-06-29 01:30:11.377 | INFO     | __main__:_train_grpo:2172 - Max steps: 10000
+2026-06-29 01:30:11.378 | INFO     | __main__:_train_grpo:2173 - Batch size: 2
+2026-06-29 01:30:11.378 | INFO     | __main__:_train_grpo:2174 - Learning rate: 1e-05
+2026-06-29 01:30:12.274 | INFO     | __main__:_resolve_resume_checkpoint:1860 - No timer checkpoint found at s3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/timer; auto-resuming from latest regular checkpoint: s3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/checkpoint-150
+2026-06-29 01:30:12.405 | INFO     | __main__:load_checkpoint:1992 - Loading checkpoint from s3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/checkpoint-150
+/usr/local/lib/python3.12/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)

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260629_013009-rgt7q8d2/files/requirements.txt

@@ -0,0 +1,287 @@
+awscli==1.45.30
+boto3==1.43.30
+botocore==1.43.30
+certifi==2026.5.20
+charset-normalizer==3.4.7
+colorama==0.4.6
+docutils==0.19
+idna==3.18
+jmespath==1.1.0
+pip==26.1.2
+pyasn1==0.6.3
+python-dateutil==2.9.0.post0
+PyYAML==6.0.3
+requests==2.34.2
+rsa==4.7.2
+s3transfer==0.18.0
+six==1.17.0
+urllib3==2.7.0
+py-cpuinfo==9.0.0
+hjson==3.1.0
+easydict==1.13
+braceexpand==0.1.7
+antlr4-python3-runtime==4.9.3
+websocket-client==1.9.0
+webdataset==1.0.2
+triton==3.6.0
+torchcodec==0.14.0+cu130
+tifffile==2026.6.1
+tenacity==9.1.4
+smmap==5.0.3
+sentry-sdk==2.63.0
+sentencepiece==0.2.1
+regex==2026.6.28
+pyarrow==24.0.0
+pyaml==26.2.1
+propcache==0.5.2
+opencv-python-headless==4.13.0.92
+omegaconf==2.3.1
+nvidia-nvtx==13.0.85
+ImageIO==2.37.3
+nvidia-nvshmem-cu13==3.4.5
+nvidia-nvjitlink==13.0.88
+av==17.1.0
+nvidia-nccl-cu13==2.28.9
+nvidia-cufile==1.15.1.6
+imageio-ffmpeg==0.6.0
+nvidia-cuda-runtime==13.0.96
+nvidia-cuda-nvrtc==13.0.88
+attrs==26.1.0
+nvidia-cuda-cupti==13.0.85
+nvidia-cublas==13.1.0.3
+multidict==6.7.1
+msgpack==1.2.1
+Mako==1.3.12
+loguru==0.7.3
+lmdb==2.2.1
+lightning-utilities==0.15.3
+lazy-loader==0.5
+greenlet==3.5.3
+future==1.0.0
+ftfy==6.3.1
+frozenlist==1.8.0
+frozendict==2.4.7
+docstring_parser==0.18.0
+decord==0.6.0
+cuda-pathfinder==1.5.5
+aiohappyeyeballs==2.6.2
+yarl==1.24.2
+typed-argument-parser==1.12.0
+SQLAlchemy==2.0.51
+scikit-image==0.26.0
+qwen-vl-utils==0.0.14
+typeguard==4.5.2
+nvidia-cusparse==12.6.3.3
+nvidia-cufft==12.0.0.61
+nvidia-cudnn-cu13==9.15.1.9
+gitdb==4.0.12
+ffmpeg-python==0.2.0
+einx==0.4.3
+cupy-cuda12x==14.1.1
+cuda-bindings==13.0.3
+cattrs==26.1.0
+aiosignal==1.4.0
+tokenizers==0.22.2
+termcolor==3.3.0
+nvidia-cusolver==12.0.4.66
+mediapy==1.2.6
+ipdb==0.13.13
+GitPython==3.1.50
+diffusers==0.38.0
+coremltools==9.0
+alembic==1.18.5
+aiohttp==3.14.1
+wandb==0.28.0
+torch==2.10.0+cu130
+dataset==2.0.0
+dashscope==1.26.0
+torchvision==0.25.0+cu130
+torchmetrics==1.9.0
+sniffio==1.3.1
+torchdata==0.8.0
+tqdm==4.67.3
+torchaudio==2.10.0+cu130
+deepspeed==0.19.2
+timm==1.0.27
+pytorch-lightning==2.6.5
+peft==0.19.1
+lpips==0.1.4
+lightning==2.6.5
+flash_attn==2.8.3+cu130torch2.10
+xxhash==3.8.0
+shortuuid==1.0.13
+setuptools==82.0.0
+safetensors==0.7.0
+pillow==12.1.1
+numpy==2.4.2
+jiter==0.15.0
+fsspec==2026.2.0
+distro==1.9.0
+dill==0.4.1
+tyro==1.0.15
+tiktoken==0.13.0
+multiprocess==0.70.19
+modelscope==1.36.0
+fire==0.7.0
+openai==2.8.0
+datasets==4.8.4
+trl==0.8.6
+angelslim==0.2.2
+huggingface_hub==1.21.0
+transformers==5.12.1
+black==26.5.1
+pathspec==1.1.1
+pytokens==0.4.1
+importlib_metadata==9.0.0
+ml_dtypes==0.5.4
+onnx==1.22.0
+onnx-ir==0.2.1
+onnxscript==0.7.0
+transformer_engine==2.9.0+70f53666
+zipp==4.1.0
+nvidia-mathdx==25.6.0
+einops==0.8.2
+absl-py==2.4.0
+accelerate==1.14.0
+annotated-doc==0.0.4
+annotated-types==0.7.0
+anyio==4.14.0
+apsw==3.53.2.0
+apswutils==0.1.2
+beartype==0.22.9
+beautifulsoup4==4.15.0
+blis==1.2.1
+catalogue==2.0.10
+cloudpathlib==0.24.0
+cloudpickle==3.1.2
+confection==0.1.5
+contourpy==1.3.3
+cycler==0.12.1
+cymem==2.0.13
+fastai==2.8.7
+fastcore==1.13.6
+fastdownload==0.0.7
+fastlite==0.2.4
+fastprogress==1.1.6
+fasttransform==0.0.2
+fonttools==4.63.0
+grpcio==1.81.1
+h11==0.16.0
+hf-xet==1.5.1
+httpcore==1.0.9
+httpcore2==2.4.0
+httptools==0.8.0
+httpx==0.28.1
+httpx2==2.4.0
+itsdangerous==2.2.0
+joblib==1.5.3
+kiwisolver==1.5.0
+langcodes==3.5.1
+Markdown==3.10.2
+markdown-it-py==4.2.0
+matplotlib==3.11.0
+mdurl==0.1.2
+mpmath==1.3.0
+murmurhash==1.0.15
+mypy_extensions==1.1.0
+narwhals==2.22.1
+networkx==3.6.1
+nvidia-curand==10.4.0.35
+nvidia-cusparselt-cu13==0.8.0
+oauthlib==3.3.1
+pandas==3.0.3
+plum-dispatch==2.9.0
+preshed==3.0.13
+protobuf==7.35.1
+pydantic==2.13.4
+pydantic_core==2.46.4
+pyparsing==3.3.2
+pyre-extensions==0.0.32
+python-dotenv==1.2.2
+python-fasthtml==0.14.3
+python-multipart==0.0.32
+rich==15.0.0
+s3torchconnector==1.5.0
+s3torchconnectorclient==1.5.0
+scikit-learn==1.9.0
+shellingham==1.5.4
+smart_open==7.6.1
+soupsieve==2.8.4
+spacy==3.8.7
+spacy-legacy==3.0.12
+spacy-loggers==1.0.5
+srsly==2.5.3
+starlette==1.3.1
+sympy==1.14.0
+tabulate==0.10.0
+tensorboard==2.20.0
+tensorboard-data-server==0.7.2
+thinc==8.3.4
+threadpoolctl==3.6.0
+torchtnt==0.2.4
+truststore==0.10.4
+typer==0.25.1
+typer-slim==0.24.0
+typing-inspect==0.9.0
+typing-inspection==0.4.2
+uvicorn==0.49.0
+uvloop==0.22.1
+wasabi==1.1.3
+watchfiles==1.2.0
+weasel==0.4.3
+websockets==16.0
+Werkzeug==3.1.8
+wrapt==2.2.2
+asttokens==3.0.1
+cffi==2.0.0
+click==8.4.1
+comm==0.2.3
+cryptography==49.0.0
+Cython==3.2.5
+debugpy==1.8.21
+decorator==5.3.1
+executing==2.2.1
+filelock==3.29.4
+h5py==3.16.0
+intel-cmplr-lib-ur==2026.0.0
+intel-openmp==2026.0.0
+ipykernel==7.3.0
+ipython==9.14.1
+ipython_pygments_lexers==1.1.1
+jedi==0.20.0
+Jinja2==3.1.6
+jupyter_client==8.9.1
+jupyter_core==5.9.1
+MarkupSafe==3.0.3
+matplotlib-inline==0.2.2
+mkl==2026.0.0
+mkl-include==2026.0.0
+mpi4py==4.1.2
+nest-asyncio2==1.7.2
+ninja==1.13.0
+onemkl-license==2026.0.0
+opencv-python==4.11.0.86
+packaging==26.2
+parso==0.8.7
+pexpect==4.9.0
+platformdirs==4.10.0
+prompt_toolkit==3.0.52
+psutil==7.2.2
+ptyprocess==0.7.0
+pure_eval==0.2.3
+pybind11==3.0.4
+pycparser==3.0
+Pygments==2.20.0
+pyOpenSSL==26.3.0
+pyzmq==27.1.0
+scipy==1.18.0
+stack-data==0.6.3
+tbb==2023.0.0
+tcmlib==1.5.0
+tornado==6.5.7
+traitlets==5.15.1
+typing_extensions==4.15.0
+tzdata==2026.2
+umf==1.1.0
+wheel==0.47.0
+wcwidth==0.8.1

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260629_013009-rgt7q8d2/files/wandb-metadata.json

@@ -0,0 +1,185 @@
+{
+  "os": "Linux-6.1.163-186.299.amzn2023.x86_64-x86_64-with-glibc2.35",
+  "python": "CPython 3.12.10",
+  "startedAt": "2026-06-29T01:30:09.016446Z",
+  "args": [
+    "--pretrained_model_root",
+    "./ckpts",
+    "--train_video_csv",
+    "/nfs/ywang29/Reward_finetuning/SAGE-GRPO/assets/demo_train_decomposed_binary5.csv",
+    "--reward_model",
+    "snapvideo_vlm",
+    "--vlm_reward_model_path",
+    "/nfs/ywang29/ckpts/Qwen3.5-9B",
+    "--vlm_reward_model_family",
+    "qwen3_5",
+    "--vlm_reward_num_frames",
+    "20",
+    "--vlm_reward_max_pixels",
+    "50176",
+    "--vlm_reward_max_new_tokens",
+    "128",
+    "--vlm_reward_batch_size",
+    "1",
+    "--vlm_reward_score_type",
+    "token_credit",
+    "--vlm_reward_token_credit_max_questions",
+    "5",
+    "--token_credit_loss_mode",
+    "video_adv",
+    "--token_credit_ratio_mode",
+    "weighted_logprob",
+    "--token_credit_spatial_granularity",
+    "window",
+    "--token_credit_spatial_bins",
+    "16",
+    "--token_credit_spatial_window",
+    "3",
+    "--token_credit_chunk_size",
+    "1",
+    "--learning_rate",
+    "1e-5",
+    "--batch_size",
+    "2",
+    "--num_generations",
+    "4",
+    "--max_steps",
+    "10000",
+    "--output_dir",
+    "./outputs",
+    "--enable_wandb",
+    "True",
+    "--project_name",
+    "SAGE-GRPO",
+    "--run_name",
+    "yw006-4-sage-qwen35-9b-token-credit-video-adv-window3",
+    "--enable_fsdp",
+    "--enable_gradient_checkpointing",
+    "--sp_size",
+    "2",
+    "--reward_checkpoint_mode",
+    "v3",
+    "--validation_interval",
+    "0",
+    "--validate_at_step0",
+    "False",
+    "--validate_video_length",
+    "121",
+    "--validation_timestep_shift",
+    "5.0",
+    "--use_grad_balancing",
+    "True",
+    "--enable_timestep_permutation",
+    "True",
+    "--sde_type",
+    "sage_grpo",
+    "--kl_weight",
+    "1e-5",
+    "--kl_coef",
+    "1e-7",
+    "--use_moving_KL",
+    "True",
+    "--update_ref_model_step",
+    "10",
+    "--use_dual_kl",
+    "True",
+    "--dual_kl_moving_weight",
+    "1.0",
+    "--dual_kl_step_weight",
+    "0.1",
+    "--reference_mode_offload",
+    "True",
+    "--save_interval",
+    "25",
+    "--timer_interval",
+    "0",
+    "--auto_resume_from_timer",
+    "True",
+    "--checkpoints_directory",
+    "s3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3"
+  ],
+  "program": "/nfs/ywang29/Reward_finetuning/SAGE-GRPO/post_train.py",
+  "codePath": "post_train.py",
+  "codePathLocal": "post_train.py",
+  "git": {
+    "remote": "https://github.com/Tencent-Hunyuan/SAGE-GRPO.git",
+    "commit": "779776e42732e0262b3f2844dd206d9d8f043b2f"
+  },
+  "root": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb",
+  "host": "ywang29-hunyuan-sage-qwen35-tcva-w3-worker-0",
+  "username": "yi-fan-wang1216",
+  "executable": "/usr/local/bin/python",
+  "cpu_count": 48,
+  "cpu_count_logical": 96,
+  "gpu": "NVIDIA A100-SXM4-80GB",
+  "gpu_count": 8,
+  "disk": {
+    "/": {
+      "total": "7516777873408",
+      "used": "619658489856"
+    }
+  },
+  "memory": {
+    "total": "1204539252736"
+  },
+  "gpu_nvidia": [
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-f6eafbbf-b72f-1725-61a0-5da7fda21fe2"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-8faea6eb-ec0a-0561-c1a7-608821eb24f5"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-6d3845d3-5be7-3130-93f6-aa8bc1fc8f9b"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-a80bc0c5-98ed-a833-ec83-51e290f6ac9c"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-a8dcbb50-ac55-1c6b-f4a3-66a2e17a2e9f"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-a6632210-2f96-54eb-7c24-6689d8c3ab90"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-e2324dd0-03ea-7053-4212-42facf1acdf7"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-20e27620-2f39-65f1-ff0c-215b32505ff5"
+    }
+  ],
+  "cudaVersion": "13.0",
+  "writerId": "gw7ezfrf1c1ofalwpf0n4wlm309go4at"
+}

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260629_013009-rgt7q8d2/logs/debug-internal.log

@@ -0,0 +1,25 @@
+{"time":"2026-06-29T01:30:09.280234912Z","level":"INFO","msg":"wandb-core"}
+{"time":"2026-06-29T01:30:09.28111396Z","level":"INFO","msg":"stream: starting","core version":"0.28.0"}
+{"time":"2026-06-29T01:30:09.913403741Z","level":"INFO","msg":"stream: created new stream","id":"rgt7q8d2"}
+{"time":"2026-06-29T01:30:09.913497531Z","level":"INFO","msg":"handler: started"}
+{"time":"2026-06-29T01:30:09.915836386Z","level":"INFO","msg":"stream: started"}
+{"time":"2026-06-29T01:30:09.915926757Z","level":"INFO","msg":"writer: started","stream_id":"rgt7q8d2"}
+{"time":"2026-06-29T01:30:09.915971076Z","level":"INFO","msg":"sender: started"}
+{"time":"2026-06-29T01:30:10.650717897Z","level":"INFO","msg":"filestream: sending request","total_files":1,"console_offset":0,"console_lines":2}
+{"time":"2026-06-29T01:30:10.880199143Z","level":"INFO","msg":"filestream: request sent","status":"200 OK"}
+{"time":"2026-06-29T01:30:25.651766565Z","level":"INFO","msg":"filestream: sending request","total_files":2,"events_offset":0,"events_lines":2,"console_offset":1,"console_lines":132,"uploaded_len":2}
+{"time":"2026-06-29T01:30:25.839685529Z","level":"INFO","msg":"filestream: request sent","status":"200 OK"}
+{"time":"2026-06-29T01:30:40.651324038Z","level":"INFO","msg":"filestream: sending request","total_files":1,"events_offset":2,"events_lines":2}
+{"time":"2026-06-29T01:30:40.818605206Z","level":"INFO","msg":"filestream: request sent","status":"200 OK"}
+{"time":"2026-06-29T01:30:55.651668998Z","level":"INFO","msg":"filestream: sending request","total_files":1,"events_offset":4,"events_lines":2}
+{"time":"2026-06-29T01:30:55.816615982Z","level":"INFO","msg":"filestream: request sent","status":"200 OK"}
+{"time":"2026-06-29T01:31:10.650842784Z","level":"INFO","msg":"filestream: sending request","total_files":1,"events_offset":6,"events_lines":2}
+{"time":"2026-06-29T01:31:10.846224204Z","level":"INFO","msg":"filestream: request sent","status":"200 OK"}
+{"time":"2026-06-29T01:31:25.65170482Z","level":"INFO","msg":"filestream: sending request","total_files":1,"events_offset":8,"events_lines":2}
+{"time":"2026-06-29T01:31:25.822361878Z","level":"INFO","msg":"filestream: request sent","status":"200 OK"}
+{"time":"2026-06-29T01:31:40.651772262Z","level":"INFO","msg":"filestream: sending request","total_files":1,"events_offset":10,"events_lines":2}
+{"time":"2026-06-29T01:31:40.83317116Z","level":"INFO","msg":"filestream: request sent","status":"200 OK"}
+{"time":"2026-06-29T01:31:55.650942111Z","level":"INFO","msg":"filestream: sending request","total_files":1,"events_offset":12,"events_lines":2}
+{"time":"2026-06-29T01:31:55.820758367Z","level":"INFO","msg":"filestream: request sent","status":"200 OK"}
+{"time":"2026-06-29T01:32:10.651595098Z","level":"INFO","msg":"filestream: sending request","total_files":1,"events_offset":14,"events_lines":2}
+{"time":"2026-06-29T01:32:10.828691468Z","level":"INFO","msg":"filestream: request sent","status":"200 OK"}

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260629_013009-rgt7q8d2/logs/debug.log

@@ -0,0 +1,19 @@
+2026-06-29 01:30:09,046 INFO    MainThread:1802 [wandb_setup.py:_flush():81] Current SDK version is 0.28.0
+2026-06-29 01:30:09,047 INFO    MainThread:1802 [wandb_setup.py:_flush():81] Configure stats pid to 1802
+2026-06-29 01:30:09,047 INFO    MainThread:1802 [wandb_setup.py:_flush():81] Loading settings from environment variables
+2026-06-29 01:30:09,047 INFO    MainThread:1802 [wandb_init.py:setup_run_log_directory():725] Logging user logs to outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260629_013009-rgt7q8d2/logs/debug.log
+2026-06-29 01:30:09,047 INFO    MainThread:1802 [wandb_init.py:setup_run_log_directory():726] Logging internal logs to outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260629_013009-rgt7q8d2/logs/debug-internal.log
+2026-06-29 01:30:09,048 INFO    MainThread:1802 [wandb_init.py:init():768] calling init triggers
+2026-06-29 01:30:09,048 INFO    MainThread:1802 [wandb_init.py:init():773] wandb.init called with sweep_config: {}
+config: {'pretrained_model_root': './ckpts', 'pretrained_transformer_version': '480p_t2v', 'post_train_type': 'grpo', 'learning_rate': 1e-05, 'weight_decay': 0.01, 'max_steps': 10000, 'warmup_steps': 500, 'gradient_accumulation_steps': 1, 'max_grad_norm': 1.0, 'use_muon': True, 'num_train_timesteps': 1000, 'train_timestep_shift': 3.0, 'validation_timestep_shift': 5.0, 'snr_type': 'lognorm', 'task_type': 't2v', 'i2v_prob': 0.3, 'enable_fsdp': True, 'enable_gradient_checkpointing': True, 'sp_size': 2, 'dp_replicate': 1, 'batch_size': 2, 'num_workers': 4, 'prefetch_factor': 2, 'num_generations': 4, 'use_same_noise': False, 'output_dir': './outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3', 'checkpoints_directory': 's3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3', 'save_interval': 25, 'timer_interval': 0, 'auto_resume_from_timer': True, 'log_interval': 1, 'enable_wandb': True, 'project_name': 'SAGE-GRPO', 'run_name': 'yw006-4-sage-qwen35-9b-token-credit-video-adv-window3', 'wandb_key': None, 'wandb_username': None, 'wandb_base_url': 'https://api.wandb.ai', 'wandb_entity': None, 'wandb_silent': True, 'dtype': 'bf16', 'master_weight_type': 'fp32', 'seed': 42, 'validation_global_seed': 930, 'validation_interval': 0, 'validation_prompts': None, 'validate_video_length': 121, 'validation_aspect_ratio': '16:9', 'validation_fixed_size': '480x864', 'validation_num_inference_steps': 40, 'validation_negative_prompt': '', 'validation_enable_sr': False, 'validation_prompt_rewrite': False, 'validate_at_step0': False, 'resume_from_checkpoint': None, 'use_lora': False, 'lora_r': 8, 'lora_alpha': 16, 'lora_dropout': 0.0, 'lora_target_modules': None, 'pretrained_lora_path': None, 'train_video_csv': ['/nfs/ywang29/Reward_finetuning/SAGE-GRPO/assets/demo_train_decomposed_binary5.csv'], 'valid_video_csv': 'assets/demo_sample.csv', 'reward_model': 'snapvideo_vlm', 'reward_config': None, 'reward_checkpoint_mode': 'v3', 'reward_weights': '{"VQ":0.5,"MQ":0.5,"TA":1.0}', 'reward_gradient_credit': False, 'reward_credit_num_frames': 0, 'reward_credit_max_questions': 0, 'reward_credit_metric': 'weighted', 'token_credit_source_model': None, 'token_credit_advantage_source': 'token', 'token_credit_remote_url': None, 'token_credit_vlm_model_path': None, 'token_credit_vlm_model_family': None, 'token_credit_vlm_prompt_template': None, 'remote_reward_url': None, 'vlm_reward_model_path': '/nfs/ywang29/ckpts/Qwen3.5-9B', 'vlm_reward_model_family': 'qwen3_5', 'vlm_reward_prompt_template': None, 'vlm_reward_num_frames': 20, 'vlm_reward_max_pixels': 50176, 'vlm_reward_max_new_tokens': 128, 'vlm_reward_batch_size': 1, 'vlm_reward_score_type': 'token_credit', 'vlm_reward_missing_score': 0.0, 'vlm_reward_video_sample_type': 'uniform', 'vlm_reward_token_credit_max_questions': 5, 'token_credit_loss_mode': 'video_adv', 'token_credit_ratio_mode': 'weighted_logprob', 'token_credit_shuffle_weights': False, 'token_credit_spatial_granularity': 'window', 'token_credit_spatial_bins': 16, 'token_credit_spatial_window': 3, 'token_credit_chunk_size': 1, 'enable_grpo_policy_update': True, 'grpo_sampling_steps': 20, 'sample_n_frames': 81, 'eta': 0.5, 'cfg_scale': 6.0, 'neg_prompt': None, 'data_type': 'video', 'model_type': '16164', 'video_fps': 24, 'vae_name': '16164-32c-hy20250605', 'vae_tiling': False, 'infer_flow_shift_video': 5.0, 'video_bucket_hw_base_size': 480, 'video_bucket_hw_bucket_stride': 32, 'multitask_mask_training_type': 'concat', 'vision_num_semantic_tokens': 729, 'vision_states_dim': 1152, 'sde_type': 'sage_grpo', 'mini_batch_size_per_update': 1, 'clip_range': 0.0001, 'adv_clip_max': 5.0, 'timestep_fraction': 1.0, 'random_train_timestep_count': False, 'random_train_timestep_min': 1, 'random_train_timestep_max': 20, 'use_grad_balancing': True, 'enable_timestep_permutation': True, 'reference_mode_offload': True, 'kl_weight': 1e-05, 'kl_coef': 1e-07, 'kl_min_coef': 1e-07, 'use_moving_KL': True, 'update_ref_model_step': 10, 'use_dual_kl': True, 'dual_kl_moving_weight': 1.0, 'dual_kl_step_weight': 0.1, 'kl_compute_mode': 'rollout_phase', 'glyph_byT5_v2': True, 'byt5_max_length': 256, 'video_uncond_p_byt5': 0.0, 'precision': 'bf16', 'vae_precision': 'fp16', 'val_disable_autocast': False, 'debug_grad_flow': False, 'debug_train_diagnostics_interval': 1, '_wandb': {}}
+2026-06-29 01:30:09,048 INFO    MainThread:1802 [wandb_init.py:init():816] starting backend
+2026-06-29 01:30:09,269 INFO    MainThread:1802 [wandb_init.py:init():831] sending inform_init request
+2026-06-29 01:30:09,916 INFO    MainThread:1802 [wandb_init.py:init():836] backend started and connected
+2026-06-29 01:30:10,039 INFO    MainThread:1802 [wandb_init.py:init():906] updated telemetry
+2026-06-29 01:30:10,085 INFO    MainThread:1802 [wandb_init.py:init():929] communicating run to backend with 90.0 second timeout
+2026-06-29 01:30:10,446 INFO    MainThread:1802 [wandb_init.py:init():974] starting run threads in backend
+2026-06-29 01:30:10,647 INFO    MainThread:1802 [wandb_run.py:_console_start():2523] atexit reg
+2026-06-29 01:30:10,647 INFO    MainThread:1802 [wandb_run.py:_redirect():2373] redirect: wrap_raw
+2026-06-29 01:30:10,647 INFO    MainThread:1802 [wandb_run.py:_redirect():2442] Wrapping output streams.
+2026-06-29 01:30:10,647 INFO    MainThread:1802 [wandb_run.py:_redirect():2465] Redirects installed.
+2026-06-29 01:30:10,649 INFO    MainThread:1802 [wandb_init.py:init():1012] run started, returning control to user process

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260629_020012-rgt7q8d2/files/output.log

@@ -0,0 +1,133 @@
+2026-06-29 02:00:13.304 | INFO     | __main__:print_training_configuration:970 - SYSTEM & HARDWARE CONFIGURATION
+2026-06-29 02:00:13.304 | INFO     | __main__:print_training_configuration:971 - --------------------------------------------------------------------------------
+2026-06-29 02:00:13.305 | INFO     | __main__:print_training_configuration:972 - Number of GPUs                                   : 64
+2026-06-29 02:00:13.305 | INFO     | __main__:print_training_configuration:973 - Local rank                                       : 0
+2026-06-29 02:00:13.306 | INFO     | __main__:print_training_configuration:974 - Global rank                                      : 0
+2026-06-29 02:00:13.306 | INFO     | __main__:print_training_configuration:975 - Device                                           : cuda:0
+2026-06-29 02:00:13.306 | INFO     | __main__:print_training_configuration:976 - Master weight dtype                              : torch.float32
+2026-06-29 02:00:13.307 | INFO     | __main__:print_training_configuration:978 - --------------------------------------------------------------------------------
+2026-06-29 02:00:13.307 | INFO     | __main__:print_training_configuration:979 - MODEL PARAMETERS & ARCHITECTURE
+2026-06-29 02:00:13.308 | INFO     | __main__:print_training_configuration:980 - --------------------------------------------------------------------------------
+2026-06-29 02:00:13.319 | INFO     | __main__:print_training_configuration:983 - Number of total                                  : 8,326,608,160
+2026-06-29 02:00:13.320 | INFO     | __main__:print_training_configuration:984 - Total trainable parameters                        : 8,326,608,160
+2026-06-29 02:00:13.320 | INFO     | __main__:print_training_configuration:985 - DIT model root                                   : ./ckpts
+2026-06-29 02:00:13.321 | INFO     | __main__:print_training_configuration:986 - Transformer version                              : 480p_t2v
+2026-06-29 02:00:13.321 | INFO     | __main__:print_training_configuration:987 - Attention mode                                   : flash
+2026-06-29 02:00:13.322 | INFO     | __main__:print_training_configuration:988 - Gradient checkpointing                           : True
+2026-06-29 02:00:13.322 | INFO     | __main__:print_training_configuration:989 - FSDP enabled                                     : True
+2026-06-29 02:00:13.323 | INFO     | __main__:print_training_configuration:990 - LoRA enabled                                     : False
+2026-06-29 02:00:13.323 | INFO     | __main__:print_training_configuration:991 - Glyph byT5 v2                                    : True
+2026-06-29 02:00:13.324 | INFO     | __main__:print_training_configuration:992 - ByT5 max length                                  : 256
+2026-06-29 02:00:13.324 | INFO     | __main__:print_training_configuration:994 - --------------------------------------------------------------------------------
+2026-06-29 02:00:13.324 | INFO     | __main__:print_training_configuration:995 - DISTRIBUTED TRAINING CONFIGURATION
+2026-06-29 02:00:13.325 | INFO     | __main__:print_training_configuration:996 - --------------------------------------------------------------------------------
+2026-06-29 02:00:13.325 | INFO     | __main__:print_training_configuration:997 - World size                                       : 64
+2026-06-29 02:00:13.326 | INFO     | __main__:print_training_configuration:998 - DP degree                                        : 32
+2026-06-29 02:00:13.326 | INFO     | __main__:print_training_configuration:999 - DP rank                                          : 0
+2026-06-29 02:00:13.327 | INFO     | __main__:print_training_configuration:1000 - SP size                                          : 2
+2026-06-29 02:00:13.327 | INFO     | __main__:print_training_configuration:1001 - SP rank                                          : 0
+2026-06-29 02:00:13.327 | INFO     | __main__:print_training_configuration:1003 - --------------------------------------------------------------------------------
+2026-06-29 02:00:13.328 | INFO     | __main__:print_training_configuration:1004 - BATCH SIZE & DATA FLOW CONFIGURATION
+2026-06-29 02:00:13.329 | INFO     | __main__:print_training_configuration:1005 - --------------------------------------------------------------------------------
+2026-06-29 02:00:13.329 | INFO     | __main__:print_training_configuration:1006 - Per GPU batch size                               : 2
+2026-06-29 02:00:13.330 | INFO     | __main__:print_training_configuration:1007 - DP degree batch size                             : 64
+2026-06-29 02:00:13.330 | INFO     | __main__:print_training_configuration:1008 - Gradient accumulation steps                      : 1
+2026-06-29 02:00:13.330 | INFO     | __main__:print_training_configuration:1009 - Num generations                                  : 4
+2026-06-29 02:00:13.331 | INFO     | __main__:print_training_configuration:1010 - Use same noise                                   : False
+2026-06-29 02:00:13.331 | INFO     | __main__:print_training_configuration:1011 - Num workers                                      : 4
+2026-06-29 02:00:13.332 | INFO     | __main__:print_training_configuration:1012 - Prefetch factor                                  : 2
+2026-06-29 02:00:13.332 | INFO     | __main__:print_training_configuration:1014 - --------------------------------------------------------------------------------
+2026-06-29 02:00:13.333 | INFO     | __main__:print_training_configuration:1015 - OPTIMIZATION CONFIGURATION
+2026-06-29 02:00:13.333 | INFO     | __main__:print_training_configuration:1016 - --------------------------------------------------------------------------------
+2026-06-29 02:00:13.333 | INFO     | __main__:print_training_configuration:1017 - Optimizer                                        : muon
+2026-06-29 02:00:13.334 | INFO     | __main__:print_training_configuration:1018 - Learning rate                                    : 1e-05
+2026-06-29 02:00:13.334 | INFO     | __main__:print_training_configuration:1019 - Weight decay                                     : 0.01
+2026-06-29 02:00:13.334 | INFO     | __main__:print_training_configuration:1020 - Max gradient norm                                : 1.0
+2026-06-29 02:00:13.335 | INFO     | __main__:print_training_configuration:1021 - LR scheduler                                     : constant
+2026-06-29 02:00:13.335 | INFO     | __main__:print_training_configuration:1022 - Warmup steps                                     : 500
+2026-06-29 02:00:13.336 | INFO     | __main__:print_training_configuration:1023 - Max training steps                               : 10000
+2026-06-29 02:00:13.336 | INFO     | __main__:print_training_configuration:1025 - --------------------------------------------------------------------------------
+2026-06-29 02:00:13.337 | INFO     | __main__:print_training_configuration:1026 - INFERENCE & VALIDATION CONFIGURATION
+2026-06-29 02:00:13.337 | INFO     | __main__:print_training_configuration:1027 - --------------------------------------------------------------------------------
+2026-06-29 02:00:13.338 | INFO     | __main__:print_training_configuration:1028 - Validation interval                              : 0
+2026-06-29 02:00:13.338 | INFO     | __main__:print_training_configuration:1029 - Validate at step 0                               : False
+2026-06-29 02:00:13.338 | INFO     | __main__:print_training_configuration:1030 - Validation video length                          : 121
+2026-06-29 02:00:13.339 | INFO     | __main__:print_training_configuration:1031 - Validation global seed                           : 930
+2026-06-29 02:00:13.339 | INFO     | __main__:print_training_configuration:1032 - Validation inference steps                        : 40
+2026-06-29 02:00:13.340 | INFO     | __main__:print_training_configuration:1033 - Validation aspect ratio                          : 16:9
+2026-06-29 02:00:13.340 | INFO     | __main__:print_training_configuration:1034 - Validation fixed size                            : 480x864
+2026-06-29 02:00:13.341 | INFO     | __main__:print_training_configuration:1035 - Validation prompt rewrite                         : False
+2026-06-29 02:00:13.341 | INFO     | __main__:print_training_configuration:1036 - Validation flow shift                             : 5.0
+2026-06-29 02:00:13.342 | INFO     | __main__:print_training_configuration:1037 - Sampling CFG scale                                : 6.0
+2026-06-29 02:00:13.342 | INFO     | __main__:print_training_configuration:1038 - Sampling eta                                      : 0.5
+2026-06-29 02:00:13.343 | INFO     | __main__:print_training_configuration:1039 - Sampling steps (GRPO rollout)                     : 20
+2026-06-29 02:00:13.343 | INFO     | __main__:print_training_configuration:1040 - Token credit loss mode                            : video_adv
+2026-06-29 02:00:13.344 | INFO     | __main__:print_training_configuration:1041 - Token credit ratio mode                           : weighted_logprob
+2026-06-29 02:00:13.344 | INFO     | __main__:print_training_configuration:1042 - Token credit shuffle weights                      : False
+2026-06-29 02:00:13.345 | INFO     | __main__:print_training_configuration:1043 - Token credit spatial granularity                  : window
+2026-06-29 02:00:13.345 | INFO     | __main__:print_training_configuration:1044 - Token credit spatial bins                         : 16
+2026-06-29 02:00:13.345 | INFO     | __main__:print_training_configuration:1045 - Token credit spatial window                       : 3
+2026-06-29 02:00:13.346 | INFO     | __main__:print_training_configuration:1046 - Token credit chunk size                           : 1
+2026-06-29 02:00:13.346 | INFO     | __main__:print_training_configuration:1047 - Token credit source model                         : None
+2026-06-29 02:00:13.346 | INFO     | __main__:print_training_configuration:1048 - Token credit advantage source                     : token
+2026-06-29 02:00:13.347 | INFO     | __main__:print_training_configuration:1049 - Token credit VLM model path                       : None
+2026-06-29 02:00:13.348 | INFO     | __main__:print_training_configuration:1050 - Token credit VLM model family                     : None
+2026-06-29 02:00:13.348 | INFO     | __main__:print_training_configuration:1051 - Random train timestep count                       : False
+2026-06-29 02:00:13.348 | INFO     | __main__:print_training_configuration:1052 - Random train timestep range                       : 1-20
+2026-06-29 02:00:13.349 | INFO     | __main__:print_training_configuration:1056 - SDE type                                          : sage_grpo
+2026-06-29 02:00:13.349 | INFO     | __main__:print_training_configuration:1058 - --------------------------------------------------------------------------------
+2026-06-29 02:00:13.350 | INFO     | __main__:print_training_configuration:1059 - GRPO & KL CONFIGURATION
+2026-06-29 02:00:13.350 | INFO     | __main__:print_training_configuration:1060 - --------------------------------------------------------------------------------
+2026-06-29 02:00:13.351 | INFO     | __main__:print_training_configuration:1061 - Reward model                                     : snapvideo_vlm
+2026-06-29 02:00:13.351 | INFO     | __main__:print_training_configuration:1062 - Reward gradient credit                          : False
+2026-06-29 02:00:13.351 | INFO     | __main__:print_training_configuration:1063 - Reward credit frames                            : 0
+2026-06-29 02:00:13.352 | INFO     | __main__:print_training_configuration:1064 - Reward credit max questions                     : 0
+2026-06-29 02:00:13.352 | INFO     | __main__:print_training_configuration:1065 - Reward credit metric                            : weighted
+2026-06-29 02:00:13.353 | INFO     | __main__:print_training_configuration:1067 - VLM reward model path                            : /nfs/ywang29/ckpts/Qwen3.5-9B
+2026-06-29 02:00:13.353 | INFO     | __main__:print_training_configuration:1068 - VLM reward model family                          : qwen3_5
+2026-06-29 02:00:13.354 | INFO     | __main__:print_training_configuration:1069 - VLM reward frames                                : 20
+2026-06-29 02:00:13.354 | INFO     | __main__:print_training_configuration:1070 - VLM reward max pixels                            : 50176
+2026-06-29 02:00:13.355 | INFO     | __main__:print_training_configuration:1071 - VLM reward score type                            : token_credit
+2026-06-29 02:00:13.355 | INFO     | __main__:print_training_configuration:1072 - VLM reward max new tokens                        : 128
+2026-06-29 02:00:13.355 | INFO     | __main__:print_training_configuration:1073 - VLM reward batch size                            : 1
+2026-06-29 02:00:13.356 | INFO     | __main__:print_training_configuration:1074 - VLM reward missing score                         : 0.0
+2026-06-29 02:00:13.356 | INFO     | __main__:print_training_configuration:1077 - KL weight                                        : 1e-05
+2026-06-29 02:00:13.357 | INFO     | __main__:print_training_configuration:1078 - KL coef                                          : 1e-07
+2026-06-29 02:00:13.357 | INFO     | __main__:print_training_configuration:1079 - Use moving KL                                    : True
+2026-06-29 02:00:13.358 | INFO     | __main__:print_training_configuration:1080 - Update ref model step                            : 10
+2026-06-29 02:00:13.358 | INFO     | __main__:print_training_configuration:1081 - Use dual KL                                      : True
+2026-06-29 02:00:13.359 | INFO     | __main__:print_training_configuration:1082 - Dual KL moving weight                            : 1.0
+2026-06-29 02:00:13.359 | INFO     | __main__:print_training_configuration:1083 - Dual KL step weight                              : 0.1
+2026-06-29 02:00:13.360 | INFO     | __main__:print_training_configuration:1084 - KL compute mode                                  : rollout_phase
+2026-06-29 02:00:13.360 | INFO     | __main__:print_training_configuration:1085 - Use grad balancing                               : True
+2026-06-29 02:00:13.360 | INFO     | __main__:print_training_configuration:1086 - Enable timestep permutation                      : True
+2026-06-29 02:00:13.361 | INFO     | __main__:print_training_configuration:1087 - Reference mode offload                           : True
+2026-06-29 02:00:13.361 | INFO     | __main__:print_training_configuration:1089 - --------------------------------------------------------------------------------
+2026-06-29 02:00:13.362 | INFO     | __main__:print_training_configuration:1090 - LOGGING & CHECKPOINTING CONFIGURATION
+2026-06-29 02:00:13.362 | INFO     | __main__:print_training_configuration:1091 - --------------------------------------------------------------------------------
+2026-06-29 02:00:13.363 | INFO     | __main__:print_training_configuration:1092 - Output directory                                 : ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3
+2026-06-29 02:00:13.363 | INFO     | __main__:print_training_configuration:1093 - Checkpoints directory                            : s3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3
+2026-06-29 02:00:13.363 | INFO     | __main__:print_training_configuration:1094 - Checkpoint interval                              : 25
+2026-06-29 02:00:13.364 | INFO     | __main__:print_training_configuration:1095 - Timer checkpoint interval                        : 0
+2026-06-29 02:00:13.364 | INFO     | __main__:print_training_configuration:1096 - Auto resume from timer/checkpoint                : True
+2026-06-29 02:00:13.365 | INFO     | __main__:print_training_configuration:1097 - Timer checkpoint directory                       : s3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/timer
+2026-06-29 02:00:13.365 | INFO     | __main__:print_training_configuration:1098 - Log interval                                     : 1
+2026-06-29 02:00:13.366 | INFO     | __main__:print_training_configuration:1099 - Resume checkpoint                                : None
+2026-06-29 02:00:13.366 | INFO     | __main__:print_training_configuration:1100 - Train CSV                                         : ['/nfs/ywang29/Reward_finetuning/SAGE-GRPO/assets/demo_train_decomposed_binary5.csv']
+2026-06-29 02:00:13.367 | INFO     | __main__:print_training_configuration:1101 - Valid CSV                                         : assets/demo_sample.csv
+2026-06-29 02:00:13.367 | INFO     | __main__:print_training_configuration:1102 - Seed                                             : 42
+2026-06-29 02:00:13.368 | INFO     | __main__:print_training_configuration:1103 - WandB enabled                                    : True
+2026-06-29 02:00:13.368 | INFO     | __main__:print_training_configuration:1105 - WandB project                                    : SAGE-GRPO
+2026-06-29 02:00:13.369 | INFO     | __main__:print_training_configuration:1106 - WandB run name                                   : yw006-4-sage-qwen35-9b-token-credit-video-adv-window3
+2026-06-29 02:00:13.369 | INFO     | __main__:__init__:958 - [Effective Config] world_size=64, rank=0, local_rank=0, sp_size=2, dp_size=32, dp_rank=0, sp_rank=0, batch_size=2, num_generations=4, grad_accum=1
+2026-06-29 02:00:13.369 | INFO     | __main__:_build_video_dataset:2088 - [Effective GRPO Grouping] video_batch_size=2, dp_replicas=32, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, product=64
+2026-06-29 02:00:14.039 | INFO     | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1
+2026-06-29 02:00:14.040 | INFO     | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done
+2026-06-29 02:00:14.041 | INFO     | __main__:_train_grpo:2171 - Starting GRPO training...
+2026-06-29 02:00:14.042 | INFO     | __main__:_train_grpo:2172 - Max steps: 10000
+2026-06-29 02:00:14.042 | INFO     | __main__:_train_grpo:2173 - Batch size: 2
+2026-06-29 02:00:14.043 | INFO     | __main__:_train_grpo:2174 - Learning rate: 1e-05
+2026-06-29 02:00:14.971 | INFO     | __main__:_resolve_resume_checkpoint:1860 - No timer checkpoint found at s3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/timer; auto-resuming from latest regular checkpoint: s3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/checkpoint-150
+2026-06-29 02:00:15.117 | INFO     | __main__:load_checkpoint:1992 - Loading checkpoint from s3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/checkpoint-150
+/usr/local/lib/python3.12/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)

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260629_020012-rgt7q8d2/files/requirements.txt

@@ -0,0 +1,287 @@
+awscli==1.45.30
+boto3==1.43.30
+botocore==1.43.30
+certifi==2026.5.20
+charset-normalizer==3.4.7
+colorama==0.4.6
+docutils==0.19
+idna==3.18
+jmespath==1.1.0
+pip==26.1.2
+pyasn1==0.6.3
+python-dateutil==2.9.0.post0
+PyYAML==6.0.3
+requests==2.34.2
+rsa==4.7.2
+s3transfer==0.18.0
+six==1.17.0
+urllib3==2.7.0
+py-cpuinfo==9.0.0
+hjson==3.1.0
+easydict==1.13
+braceexpand==0.1.7
+antlr4-python3-runtime==4.9.3
+websocket-client==1.9.0
+webdataset==1.0.2
+triton==3.6.0
+torchcodec==0.14.0+cu130
+tifffile==2026.6.1
+tenacity==9.1.4
+smmap==5.0.3
+sentry-sdk==2.63.0
+sentencepiece==0.2.1
+regex==2026.6.28
+pyarrow==24.0.0
+pyaml==26.2.1
+propcache==0.5.2
+opencv-python-headless==4.13.0.92
+omegaconf==2.3.1
+nvidia-nvtx==13.0.85
+ImageIO==2.37.3
+nvidia-nvshmem-cu13==3.4.5
+nvidia-nvjitlink==13.0.88
+av==17.1.0
+nvidia-nccl-cu13==2.28.9
+nvidia-cufile==1.15.1.6
+imageio-ffmpeg==0.6.0
+nvidia-cuda-runtime==13.0.96
+nvidia-cuda-nvrtc==13.0.88
+attrs==26.1.0
+nvidia-cuda-cupti==13.0.85
+nvidia-cublas==13.1.0.3
+multidict==6.7.1
+msgpack==1.2.1
+Mako==1.3.12
+loguru==0.7.3
+lmdb==2.2.1
+lightning-utilities==0.15.3
+lazy-loader==0.5
+greenlet==3.5.3
+future==1.0.0
+ftfy==6.3.1
+frozenlist==1.8.0
+frozendict==2.4.7
+docstring_parser==0.18.0
+decord==0.6.0
+cuda-pathfinder==1.5.5
+aiohappyeyeballs==2.6.2
+yarl==1.24.2
+typed-argument-parser==1.12.0
+SQLAlchemy==2.0.51
+scikit-image==0.26.0
+qwen-vl-utils==0.0.14
+typeguard==4.5.2
+nvidia-cusparse==12.6.3.3
+nvidia-cufft==12.0.0.61
+nvidia-cudnn-cu13==9.15.1.9
+gitdb==4.0.12
+ffmpeg-python==0.2.0
+einx==0.4.3
+cupy-cuda12x==14.1.1
+cuda-bindings==13.0.3
+cattrs==26.1.0
+aiosignal==1.4.0
+tokenizers==0.22.2
+termcolor==3.3.0
+nvidia-cusolver==12.0.4.66
+mediapy==1.2.6
+ipdb==0.13.13
+GitPython==3.1.50
+diffusers==0.38.0
+coremltools==9.0
+alembic==1.18.5
+aiohttp==3.14.1
+wandb==0.28.0
+torch==2.10.0+cu130
+dataset==2.0.0
+dashscope==1.26.0
+torchvision==0.25.0+cu130
+torchmetrics==1.9.0
+sniffio==1.3.1
+torchdata==0.8.0
+tqdm==4.67.3
+torchaudio==2.10.0+cu130
+deepspeed==0.19.2
+timm==1.0.27
+pytorch-lightning==2.6.5
+peft==0.19.1
+lpips==0.1.4
+lightning==2.6.5
+flash_attn==2.8.3+cu130torch2.10
+xxhash==3.8.0
+shortuuid==1.0.13
+setuptools==82.0.0
+safetensors==0.7.0
+pillow==12.1.1
+numpy==2.4.2
+jiter==0.15.0
+fsspec==2026.2.0
+distro==1.9.0
+dill==0.4.1
+tyro==1.0.15
+tiktoken==0.13.0
+multiprocess==0.70.19
+modelscope==1.36.0
+fire==0.7.0
+openai==2.8.0
+datasets==4.8.4
+trl==0.8.6
+angelslim==0.2.2
+huggingface_hub==1.21.0
+transformers==5.12.1
+black==26.5.1
+pathspec==1.1.1
+pytokens==0.4.1
+importlib_metadata==9.0.0
+ml_dtypes==0.5.4
+onnx==1.22.0
+onnx-ir==0.2.1
+onnxscript==0.7.0
+transformer_engine==2.9.0+70f53666
+zipp==4.1.0
+nvidia-mathdx==25.6.0
+einops==0.8.2
+absl-py==2.4.0
+accelerate==1.14.0
+annotated-doc==0.0.4
+annotated-types==0.7.0
+anyio==4.14.0
+apsw==3.53.2.0
+apswutils==0.1.2
+beartype==0.22.9
+beautifulsoup4==4.15.0
+blis==1.2.1
+catalogue==2.0.10
+cloudpathlib==0.24.0
+cloudpickle==3.1.2
+confection==0.1.5
+contourpy==1.3.3
+cycler==0.12.1
+cymem==2.0.13
+fastai==2.8.7
+fastcore==1.13.6
+fastdownload==0.0.7
+fastlite==0.2.4
+fastprogress==1.1.6
+fasttransform==0.0.2
+fonttools==4.63.0
+grpcio==1.81.1
+h11==0.16.0
+hf-xet==1.5.1
+httpcore==1.0.9
+httpcore2==2.4.0
+httptools==0.8.0
+httpx==0.28.1
+httpx2==2.4.0
+itsdangerous==2.2.0
+joblib==1.5.3
+kiwisolver==1.5.0
+langcodes==3.5.1
+Markdown==3.10.2
+markdown-it-py==4.2.0
+matplotlib==3.11.0
+mdurl==0.1.2
+mpmath==1.3.0
+murmurhash==1.0.15
+mypy_extensions==1.1.0
+narwhals==2.22.1
+networkx==3.6.1
+nvidia-curand==10.4.0.35
+nvidia-cusparselt-cu13==0.8.0
+oauthlib==3.3.1
+pandas==3.0.3
+plum-dispatch==2.9.0
+preshed==3.0.13
+protobuf==7.35.1
+pydantic==2.13.4
+pydantic_core==2.46.4
+pyparsing==3.3.2
+pyre-extensions==0.0.32
+python-dotenv==1.2.2
+python-fasthtml==0.14.3
+python-multipart==0.0.32
+rich==15.0.0
+s3torchconnector==1.5.0
+s3torchconnectorclient==1.5.0
+scikit-learn==1.9.0
+shellingham==1.5.4
+smart_open==7.6.1
+soupsieve==2.8.4
+spacy==3.8.7
+spacy-legacy==3.0.12
+spacy-loggers==1.0.5
+srsly==2.5.3
+starlette==1.3.1
+sympy==1.14.0
+tabulate==0.10.0
+tensorboard==2.20.0
+tensorboard-data-server==0.7.2
+thinc==8.3.4
+threadpoolctl==3.6.0
+torchtnt==0.2.4
+truststore==0.10.4
+typer==0.25.1
+typer-slim==0.24.0
+typing-inspect==0.9.0
+typing-inspection==0.4.2
+uvicorn==0.49.0
+uvloop==0.22.1
+wasabi==1.1.3
+watchfiles==1.2.0
+weasel==0.4.3
+websockets==16.0
+Werkzeug==3.1.8
+wrapt==2.2.2
+asttokens==3.0.1
+cffi==2.0.0
+click==8.4.1
+comm==0.2.3
+cryptography==49.0.0
+Cython==3.2.5
+debugpy==1.8.21
+decorator==5.3.1
+executing==2.2.1
+filelock==3.29.4
+h5py==3.16.0
+intel-cmplr-lib-ur==2026.0.0
+intel-openmp==2026.0.0
+ipykernel==7.3.0
+ipython==9.14.1
+ipython_pygments_lexers==1.1.1
+jedi==0.20.0
+Jinja2==3.1.6
+jupyter_client==8.9.1
+jupyter_core==5.9.1
+MarkupSafe==3.0.3
+matplotlib-inline==0.2.2
+mkl==2026.0.0
+mkl-include==2026.0.0
+mpi4py==4.1.2
+nest-asyncio2==1.7.2
+ninja==1.13.0
+onemkl-license==2026.0.0
+opencv-python==4.11.0.86
+packaging==26.2
+parso==0.8.7
+pexpect==4.9.0
+platformdirs==4.10.0
+prompt_toolkit==3.0.52
+psutil==7.2.2
+ptyprocess==0.7.0
+pure_eval==0.2.3
+pybind11==3.0.4
+pycparser==3.0
+Pygments==2.20.0
+pyOpenSSL==26.3.0
+pyzmq==27.1.0
+scipy==1.18.0
+stack-data==0.6.3
+tbb==2023.0.0
+tcmlib==1.5.0
+tornado==6.5.7
+traitlets==5.15.1
+typing_extensions==4.15.0
+tzdata==2026.2
+umf==1.1.0
+wheel==0.47.0
+wcwidth==0.8.1

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260629_020012-rgt7q8d2/files/wandb-metadata.json

@@ -0,0 +1,185 @@
+{
+  "os": "Linux-6.1.163-186.299.amzn2023.x86_64-x86_64-with-glibc2.35",
+  "python": "CPython 3.12.10",
+  "startedAt": "2026-06-29T02:00:12.081548Z",
+  "args": [
+    "--pretrained_model_root",
+    "./ckpts",
+    "--train_video_csv",
+    "/nfs/ywang29/Reward_finetuning/SAGE-GRPO/assets/demo_train_decomposed_binary5.csv",
+    "--reward_model",
+    "snapvideo_vlm",
+    "--vlm_reward_model_path",
+    "/nfs/ywang29/ckpts/Qwen3.5-9B",
+    "--vlm_reward_model_family",
+    "qwen3_5",
+    "--vlm_reward_num_frames",
+    "20",
+    "--vlm_reward_max_pixels",
+    "50176",
+    "--vlm_reward_max_new_tokens",
+    "128",
+    "--vlm_reward_batch_size",
+    "1",
+    "--vlm_reward_score_type",
+    "token_credit",
+    "--vlm_reward_token_credit_max_questions",
+    "5",
+    "--token_credit_loss_mode",
+    "video_adv",
+    "--token_credit_ratio_mode",
+    "weighted_logprob",
+    "--token_credit_spatial_granularity",
+    "window",
+    "--token_credit_spatial_bins",
+    "16",
+    "--token_credit_spatial_window",
+    "3",
+    "--token_credit_chunk_size",
+    "1",
+    "--learning_rate",
+    "1e-5",
+    "--batch_size",
+    "2",
+    "--num_generations",
+    "4",
+    "--max_steps",
+    "10000",
+    "--output_dir",
+    "./outputs",
+    "--enable_wandb",
+    "True",
+    "--project_name",
+    "SAGE-GRPO",
+    "--run_name",
+    "yw006-4-sage-qwen35-9b-token-credit-video-adv-window3",
+    "--enable_fsdp",
+    "--enable_gradient_checkpointing",
+    "--sp_size",
+    "2",
+    "--reward_checkpoint_mode",
+    "v3",
+    "--validation_interval",
+    "0",
+    "--validate_at_step0",
+    "False",
+    "--validate_video_length",
+    "121",
+    "--validation_timestep_shift",
+    "5.0",
+    "--use_grad_balancing",
+    "True",
+    "--enable_timestep_permutation",
+    "True",
+    "--sde_type",
+    "sage_grpo",
+    "--kl_weight",
+    "1e-5",
+    "--kl_coef",
+    "1e-7",
+    "--use_moving_KL",
+    "True",
+    "--update_ref_model_step",
+    "10",
+    "--use_dual_kl",
+    "True",
+    "--dual_kl_moving_weight",
+    "1.0",
+    "--dual_kl_step_weight",
+    "0.1",
+    "--reference_mode_offload",
+    "True",
+    "--save_interval",
+    "25",
+    "--timer_interval",
+    "0",
+    "--auto_resume_from_timer",
+    "True",
+    "--checkpoints_directory",
+    "s3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3"
+  ],
+  "program": "/nfs/ywang29/Reward_finetuning/SAGE-GRPO/post_train.py",
+  "codePath": "post_train.py",
+  "codePathLocal": "post_train.py",
+  "git": {
+    "remote": "https://github.com/Tencent-Hunyuan/SAGE-GRPO.git",
+    "commit": "779776e42732e0262b3f2844dd206d9d8f043b2f"
+  },
+  "root": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb",
+  "host": "ywang29-hunyuan-sage-qwen35-tcva-w3-worker-0",
+  "username": "yi-fan-wang1216",
+  "executable": "/usr/local/bin/python",
+  "cpu_count": 48,
+  "cpu_count_logical": 96,
+  "gpu": "NVIDIA A100-SXM4-80GB",
+  "gpu_count": 8,
+  "disk": {
+    "/": {
+      "total": "7516777873408",
+      "used": "619660148736"
+    }
+  },
+  "memory": {
+    "total": "1204539252736"
+  },
+  "gpu_nvidia": [
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-f6eafbbf-b72f-1725-61a0-5da7fda21fe2"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-8faea6eb-ec0a-0561-c1a7-608821eb24f5"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-6d3845d3-5be7-3130-93f6-aa8bc1fc8f9b"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-a80bc0c5-98ed-a833-ec83-51e290f6ac9c"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-a8dcbb50-ac55-1c6b-f4a3-66a2e17a2e9f"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-a6632210-2f96-54eb-7c24-6689d8c3ab90"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-e2324dd0-03ea-7053-4212-42facf1acdf7"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-20e27620-2f39-65f1-ff0c-215b32505ff5"
+    }
+  ],
+  "cudaVersion": "13.0",
+  "writerId": "8tsdd85v87gov2byexwfb4ogc8n07hdx"
+}

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260629_020012-rgt7q8d2/logs/debug-internal.log

@@ -0,0 +1,11 @@
+{"time":"2026-06-29T02:00:12.332949814Z","level":"INFO","msg":"wandb-core"}
+{"time":"2026-06-29T02:00:12.334181034Z","level":"INFO","msg":"stream: starting","core version":"0.28.0"}
+{"time":"2026-06-29T02:00:12.650629116Z","level":"INFO","msg":"stream: created new stream","id":"rgt7q8d2"}
+{"time":"2026-06-29T02:00:12.650750746Z","level":"INFO","msg":"handler: started"}
+{"time":"2026-06-29T02:00:12.653611717Z","level":"INFO","msg":"stream: started"}
+{"time":"2026-06-29T02:00:12.65363787Z","level":"INFO","msg":"writer: started","stream_id":"rgt7q8d2"}
+{"time":"2026-06-29T02:00:12.65366017Z","level":"INFO","msg":"sender: started"}
+{"time":"2026-06-29T02:00:13.304991122Z","level":"INFO","msg":"filestream: sending request","total_files":1,"console_offset":133,"console_lines":1}
+{"time":"2026-06-29T02:00:13.78763685Z","level":"INFO","msg":"filestream: request sent","status":"200 OK"}
+{"time":"2026-06-29T02:00:28.305937275Z","level":"INFO","msg":"filestream: sending request","total_files":2,"events_offset":16,"events_lines":2,"console_offset":133,"console_lines":133,"uploaded_len":2}
+{"time":"2026-06-29T02:00:28.520823732Z","level":"INFO","msg":"filestream: request sent","status":"200 OK"}

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260629_020012-rgt7q8d2/logs/debug.log

@@ -0,0 +1,20 @@
+2026-06-29 02:00:12,099 INFO    MainThread:3640 [wandb_setup.py:_flush():81] Current SDK version is 0.28.0
+2026-06-29 02:00:12,100 INFO    MainThread:3640 [wandb_setup.py:_flush():81] Configure stats pid to 3640
+2026-06-29 02:00:12,100 INFO    MainThread:3640 [wandb_setup.py:_flush():81] Loading settings from environment variables
+2026-06-29 02:00:12,100 INFO    MainThread:3640 [wandb_init.py:setup_run_log_directory():725] Logging user logs to outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260629_020012-rgt7q8d2/logs/debug.log
+2026-06-29 02:00:12,101 INFO    MainThread:3640 [wandb_init.py:setup_run_log_directory():726] Logging internal logs to outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260629_020012-rgt7q8d2/logs/debug-internal.log
+2026-06-29 02:00:12,101 INFO    MainThread:3640 [wandb_init.py:init():768] calling init triggers
+2026-06-29 02:00:12,101 INFO    MainThread:3640 [wandb_init.py:init():773] wandb.init called with sweep_config: {}
+config: {'pretrained_model_root': './ckpts', 'pretrained_transformer_version': '480p_t2v', 'post_train_type': 'grpo', 'learning_rate': 1e-05, 'weight_decay': 0.01, 'max_steps': 10000, 'warmup_steps': 500, 'gradient_accumulation_steps': 1, 'max_grad_norm': 1.0, 'use_muon': True, 'num_train_timesteps': 1000, 'train_timestep_shift': 3.0, 'validation_timestep_shift': 5.0, 'snr_type': 'lognorm', 'task_type': 't2v', 'i2v_prob': 0.3, 'enable_fsdp': True, 'enable_gradient_checkpointing': True, 'sp_size': 2, 'dp_replicate': 1, 'batch_size': 2, 'num_workers': 4, 'prefetch_factor': 2, 'num_generations': 4, 'use_same_noise': False, 'output_dir': './outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3', 'checkpoints_directory': 's3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3', 'save_interval': 25, 'timer_interval': 0, 'auto_resume_from_timer': True, 'log_interval': 1, 'enable_wandb': True, 'project_name': 'SAGE-GRPO', 'run_name': 'yw006-4-sage-qwen35-9b-token-credit-video-adv-window3', 'wandb_key': None, 'wandb_username': None, 'wandb_base_url': 'https://api.wandb.ai', 'wandb_entity': None, 'wandb_silent': True, 'dtype': 'bf16', 'master_weight_type': 'fp32', 'seed': 42, 'validation_global_seed': 930, 'validation_interval': 0, 'validation_prompts': None, 'validate_video_length': 121, 'validation_aspect_ratio': '16:9', 'validation_fixed_size': '480x864', 'validation_num_inference_steps': 40, 'validation_negative_prompt': '', 'validation_enable_sr': False, 'validation_prompt_rewrite': False, 'validate_at_step0': False, 'resume_from_checkpoint': None, 'use_lora': False, 'lora_r': 8, 'lora_alpha': 16, 'lora_dropout': 0.0, 'lora_target_modules': None, 'pretrained_lora_path': None, 'train_video_csv': ['/nfs/ywang29/Reward_finetuning/SAGE-GRPO/assets/demo_train_decomposed_binary5.csv'], 'valid_video_csv': 'assets/demo_sample.csv', 'reward_model': 'snapvideo_vlm', 'reward_config': None, 'reward_checkpoint_mode': 'v3', 'reward_weights': '{"VQ":0.5,"MQ":0.5,"TA":1.0}', 'reward_gradient_credit': False, 'reward_credit_num_frames': 0, 'reward_credit_max_questions': 0, 'reward_credit_metric': 'weighted', 'token_credit_source_model': None, 'token_credit_advantage_source': 'token', 'token_credit_remote_url': None, 'token_credit_vlm_model_path': None, 'token_credit_vlm_model_family': None, 'token_credit_vlm_prompt_template': None, 'remote_reward_url': None, 'vlm_reward_model_path': '/nfs/ywang29/ckpts/Qwen3.5-9B', 'vlm_reward_model_family': 'qwen3_5', 'vlm_reward_prompt_template': None, 'vlm_reward_num_frames': 20, 'vlm_reward_max_pixels': 50176, 'vlm_reward_max_new_tokens': 128, 'vlm_reward_batch_size': 1, 'vlm_reward_score_type': 'token_credit', 'vlm_reward_missing_score': 0.0, 'vlm_reward_video_sample_type': 'uniform', 'vlm_reward_token_credit_max_questions': 5, 'token_credit_loss_mode': 'video_adv', 'token_credit_ratio_mode': 'weighted_logprob', 'token_credit_shuffle_weights': False, 'token_credit_spatial_granularity': 'window', 'token_credit_spatial_bins': 16, 'token_credit_spatial_window': 3, 'token_credit_chunk_size': 1, 'enable_grpo_policy_update': True, 'grpo_sampling_steps': 20, 'sample_n_frames': 81, 'eta': 0.5, 'cfg_scale': 6.0, 'neg_prompt': None, 'data_type': 'video', 'model_type': '16164', 'video_fps': 24, 'vae_name': '16164-32c-hy20250605', 'vae_tiling': False, 'infer_flow_shift_video': 5.0, 'video_bucket_hw_base_size': 480, 'video_bucket_hw_bucket_stride': 32, 'multitask_mask_training_type': 'concat', 'vision_num_semantic_tokens': 729, 'vision_states_dim': 1152, 'sde_type': 'sage_grpo', 'mini_batch_size_per_update': 1, 'clip_range': 0.0001, 'adv_clip_max': 5.0, 'timestep_fraction': 1.0, 'random_train_timestep_count': False, 'random_train_timestep_min': 1, 'random_train_timestep_max': 20, 'use_grad_balancing': True, 'enable_timestep_permutation': True, 'reference_mode_offload': True, 'kl_weight': 1e-05, 'kl_coef': 1e-07, 'kl_min_coef': 1e-07, 'use_moving_KL': True, 'update_ref_model_step': 10, 'use_dual_kl': True, 'dual_kl_moving_weight': 1.0, 'dual_kl_step_weight': 0.1, 'kl_compute_mode': 'rollout_phase', 'glyph_byT5_v2': True, 'byt5_max_length': 256, 'video_uncond_p_byt5': 0.0, 'precision': 'bf16', 'vae_precision': 'fp16', 'val_disable_autocast': False, 'debug_grad_flow': False, 'debug_train_diagnostics_interval': 1, '_wandb': {}}
+2026-06-29 02:00:12,101 INFO    MainThread:3640 [wandb_init.py:init():816] starting backend
+2026-06-29 02:00:12,322 INFO    MainThread:3640 [wandb_init.py:init():831] sending inform_init request
+2026-06-29 02:00:12,654 INFO    MainThread:3640 [wandb_init.py:init():836] backend started and connected
+2026-06-29 02:00:12,657 INFO    MainThread:3640 [wandb_init.py:init():906] updated telemetry
+2026-06-29 02:00:12,688 INFO    MainThread:3640 [wandb_init.py:init():929] communicating run to backend with 90.0 second timeout
+2026-06-29 02:00:13,102 INFO    MainThread:3640 [wandb_init.py:init():969] run resumed
+2026-06-29 02:00:13,103 INFO    MainThread:3640 [wandb_init.py:init():974] starting run threads in backend
+2026-06-29 02:00:13,302 INFO    MainThread:3640 [wandb_run.py:_console_start():2523] atexit reg
+2026-06-29 02:00:13,302 INFO    MainThread:3640 [wandb_run.py:_redirect():2373] redirect: wrap_raw
+2026-06-29 02:00:13,302 INFO    MainThread:3640 [wandb_run.py:_redirect():2442] Wrapping output streams.
+2026-06-29 02:00:13,302 INFO    MainThread:3640 [wandb_run.py:_redirect():2465] Redirects installed.
+2026-06-29 02:00:13,303 INFO    MainThread:3640 [wandb_init.py:init():1012] run started, returning control to user process

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260629_022719-rgt7q8d2/files/output.log

@@ -0,0 +1,133 @@
+2026-06-29 02:27:20.882 | INFO     | __main__:print_training_configuration:970 - SYSTEM & HARDWARE CONFIGURATION
+2026-06-29 02:27:20.883 | INFO     | __main__:print_training_configuration:971 - --------------------------------------------------------------------------------
+2026-06-29 02:27:20.883 | INFO     | __main__:print_training_configuration:972 - Number of GPUs                                   : 64
+2026-06-29 02:27:20.884 | INFO     | __main__:print_training_configuration:973 - Local rank                                       : 0
+2026-06-29 02:27:20.884 | INFO     | __main__:print_training_configuration:974 - Global rank                                      : 0
+2026-06-29 02:27:20.885 | INFO     | __main__:print_training_configuration:975 - Device                                           : cuda:0
+2026-06-29 02:27:20.885 | INFO     | __main__:print_training_configuration:976 - Master weight dtype                              : torch.float32
+2026-06-29 02:27:20.886 | INFO     | __main__:print_training_configuration:978 - --------------------------------------------------------------------------------
+2026-06-29 02:27:20.886 | INFO     | __main__:print_training_configuration:979 - MODEL PARAMETERS & ARCHITECTURE
+2026-06-29 02:27:20.887 | INFO     | __main__:print_training_configuration:980 - --------------------------------------------------------------------------------
+2026-06-29 02:27:20.898 | INFO     | __main__:print_training_configuration:983 - Number of total                                  : 8,326,608,160
+2026-06-29 02:27:20.899 | INFO     | __main__:print_training_configuration:984 - Total trainable parameters                        : 8,326,608,160
+2026-06-29 02:27:20.899 | INFO     | __main__:print_training_configuration:985 - DIT model root                                   : ./ckpts
+2026-06-29 02:27:20.899 | INFO     | __main__:print_training_configuration:986 - Transformer version                              : 480p_t2v
+2026-06-29 02:27:20.900 | INFO     | __main__:print_training_configuration:987 - Attention mode                                   : flash
+2026-06-29 02:27:20.900 | INFO     | __main__:print_training_configuration:988 - Gradient checkpointing                           : True
+2026-06-29 02:27:20.901 | INFO     | __main__:print_training_configuration:989 - FSDP enabled                                     : True
+2026-06-29 02:27:20.901 | INFO     | __main__:print_training_configuration:990 - LoRA enabled                                     : False
+2026-06-29 02:27:20.901 | INFO     | __main__:print_training_configuration:991 - Glyph byT5 v2                                    : True
+2026-06-29 02:27:20.902 | INFO     | __main__:print_training_configuration:992 - ByT5 max length                                  : 256
+2026-06-29 02:27:20.902 | INFO     | __main__:print_training_configuration:994 - --------------------------------------------------------------------------------
+2026-06-29 02:27:20.903 | INFO     | __main__:print_training_configuration:995 - DISTRIBUTED TRAINING CONFIGURATION
+2026-06-29 02:27:20.903 | INFO     | __main__:print_training_configuration:996 - --------------------------------------------------------------------------------
+2026-06-29 02:27:20.904 | INFO     | __main__:print_training_configuration:997 - World size                                       : 64
+2026-06-29 02:27:20.904 | INFO     | __main__:print_training_configuration:998 - DP degree                                        : 32
+2026-06-29 02:27:20.905 | INFO     | __main__:print_training_configuration:999 - DP rank                                          : 0
+2026-06-29 02:27:20.905 | INFO     | __main__:print_training_configuration:1000 - SP size                                          : 2
+2026-06-29 02:27:20.905 | INFO     | __main__:print_training_configuration:1001 - SP rank                                          : 0
+2026-06-29 02:27:20.906 | INFO     | __main__:print_training_configuration:1003 - --------------------------------------------------------------------------------
+2026-06-29 02:27:20.906 | INFO     | __main__:print_training_configuration:1004 - BATCH SIZE & DATA FLOW CONFIGURATION
+2026-06-29 02:27:20.907 | INFO     | __main__:print_training_configuration:1005 - --------------------------------------------------------------------------------
+2026-06-29 02:27:20.907 | INFO     | __main__:print_training_configuration:1006 - Per GPU batch size                               : 2
+2026-06-29 02:27:20.908 | INFO     | __main__:print_training_configuration:1007 - DP degree batch size                             : 64
+2026-06-29 02:27:20.908 | INFO     | __main__:print_training_configuration:1008 - Gradient accumulation steps                      : 1
+2026-06-29 02:27:20.908 | INFO     | __main__:print_training_configuration:1009 - Num generations                                  : 4
+2026-06-29 02:27:20.909 | INFO     | __main__:print_training_configuration:1010 - Use same noise                                   : False
+2026-06-29 02:27:20.909 | INFO     | __main__:print_training_configuration:1011 - Num workers                                      : 4
+2026-06-29 02:27:20.909 | INFO     | __main__:print_training_configuration:1012 - Prefetch factor                                  : 2
+2026-06-29 02:27:20.910 | INFO     | __main__:print_training_configuration:1014 - --------------------------------------------------------------------------------
+2026-06-29 02:27:20.910 | INFO     | __main__:print_training_configuration:1015 - OPTIMIZATION CONFIGURATION
+2026-06-29 02:27:20.911 | INFO     | __main__:print_training_configuration:1016 - --------------------------------------------------------------------------------
+2026-06-29 02:27:20.911 | INFO     | __main__:print_training_configuration:1017 - Optimizer                                        : muon
+2026-06-29 02:27:20.912 | INFO     | __main__:print_training_configuration:1018 - Learning rate                                    : 1e-05
+2026-06-29 02:27:20.912 | INFO     | __main__:print_training_configuration:1019 - Weight decay                                     : 0.01
+2026-06-29 02:27:20.912 | INFO     | __main__:print_training_configuration:1020 - Max gradient norm                                : 1.0
+2026-06-29 02:27:20.913 | INFO     | __main__:print_training_configuration:1021 - LR scheduler                                     : constant
+2026-06-29 02:27:20.913 | INFO     | __main__:print_training_configuration:1022 - Warmup steps                                     : 500
+2026-06-29 02:27:20.914 | INFO     | __main__:print_training_configuration:1023 - Max training steps                               : 10000
+2026-06-29 02:27:20.914 | INFO     | __main__:print_training_configuration:1025 - --------------------------------------------------------------------------------
+2026-06-29 02:27:20.915 | INFO     | __main__:print_training_configuration:1026 - INFERENCE & VALIDATION CONFIGURATION
+2026-06-29 02:27:20.915 | INFO     | __main__:print_training_configuration:1027 - --------------------------------------------------------------------------------
+2026-06-29 02:27:20.915 | INFO     | __main__:print_training_configuration:1028 - Validation interval                              : 0
+2026-06-29 02:27:20.916 | INFO     | __main__:print_training_configuration:1029 - Validate at step 0                               : False
+2026-06-29 02:27:20.916 | INFO     | __main__:print_training_configuration:1030 - Validation video length                          : 121
+2026-06-29 02:27:20.916 | INFO     | __main__:print_training_configuration:1031 - Validation global seed                           : 930
+2026-06-29 02:27:20.917 | INFO     | __main__:print_training_configuration:1032 - Validation inference steps                        : 40
+2026-06-29 02:27:20.917 | INFO     | __main__:print_training_configuration:1033 - Validation aspect ratio                          : 16:9
+2026-06-29 02:27:20.918 | INFO     | __main__:print_training_configuration:1034 - Validation fixed size                            : 480x864
+2026-06-29 02:27:20.918 | INFO     | __main__:print_training_configuration:1035 - Validation prompt rewrite                         : False
+2026-06-29 02:27:20.919 | INFO     | __main__:print_training_configuration:1036 - Validation flow shift                             : 5.0
+2026-06-29 02:27:20.919 | INFO     | __main__:print_training_configuration:1037 - Sampling CFG scale                                : 6.0
+2026-06-29 02:27:20.920 | INFO     | __main__:print_training_configuration:1038 - Sampling eta                                      : 0.5
+2026-06-29 02:27:20.920 | INFO     | __main__:print_training_configuration:1039 - Sampling steps (GRPO rollout)                     : 20
+2026-06-29 02:27:20.920 | INFO     | __main__:print_training_configuration:1040 - Token credit loss mode                            : video_adv
+2026-06-29 02:27:20.921 | INFO     | __main__:print_training_configuration:1041 - Token credit ratio mode                           : weighted_logprob
+2026-06-29 02:27:20.921 | INFO     | __main__:print_training_configuration:1042 - Token credit shuffle weights                      : False
+2026-06-29 02:27:20.922 | INFO     | __main__:print_training_configuration:1043 - Token credit spatial granularity                  : window
+2026-06-29 02:27:20.922 | INFO     | __main__:print_training_configuration:1044 - Token credit spatial bins                         : 16
+2026-06-29 02:27:20.922 | INFO     | __main__:print_training_configuration:1045 - Token credit spatial window                       : 3
+2026-06-29 02:27:20.923 | INFO     | __main__:print_training_configuration:1046 - Token credit chunk size                           : 1
+2026-06-29 02:27:20.923 | INFO     | __main__:print_training_configuration:1047 - Token credit source model                         : None
+2026-06-29 02:27:20.924 | INFO     | __main__:print_training_configuration:1048 - Token credit advantage source                     : token
+2026-06-29 02:27:20.924 | INFO     | __main__:print_training_configuration:1049 - Token credit VLM model path                       : None
+2026-06-29 02:27:20.924 | INFO     | __main__:print_training_configuration:1050 - Token credit VLM model family                     : None
+2026-06-29 02:27:20.925 | INFO     | __main__:print_training_configuration:1051 - Random train timestep count                       : False
+2026-06-29 02:27:20.925 | INFO     | __main__:print_training_configuration:1052 - Random train timestep range                       : 1-20
+2026-06-29 02:27:20.926 | INFO     | __main__:print_training_configuration:1056 - SDE type                                          : sage_grpo
+2026-06-29 02:27:20.926 | INFO     | __main__:print_training_configuration:1058 - --------------------------------------------------------------------------------
+2026-06-29 02:27:20.926 | INFO     | __main__:print_training_configuration:1059 - GRPO & KL CONFIGURATION
+2026-06-29 02:27:20.927 | INFO     | __main__:print_training_configuration:1060 - --------------------------------------------------------------------------------
+2026-06-29 02:27:20.927 | INFO     | __main__:print_training_configuration:1061 - Reward model                                     : snapvideo_vlm
+2026-06-29 02:27:20.928 | INFO     | __main__:print_training_configuration:1062 - Reward gradient credit                          : False
+2026-06-29 02:27:20.928 | INFO     | __main__:print_training_configuration:1063 - Reward credit frames                            : 0
+2026-06-29 02:27:20.928 | INFO     | __main__:print_training_configuration:1064 - Reward credit max questions                     : 0
+2026-06-29 02:27:20.929 | INFO     | __main__:print_training_configuration:1065 - Reward credit metric                            : weighted
+2026-06-29 02:27:20.929 | INFO     | __main__:print_training_configuration:1067 - VLM reward model path                            : /nfs/ywang29/ckpts/Qwen3.5-9B
+2026-06-29 02:27:20.930 | INFO     | __main__:print_training_configuration:1068 - VLM reward model family                          : qwen3_5
+2026-06-29 02:27:20.930 | INFO     | __main__:print_training_configuration:1069 - VLM reward frames                                : 20
+2026-06-29 02:27:20.931 | INFO     | __main__:print_training_configuration:1070 - VLM reward max pixels                            : 50176
+2026-06-29 02:27:20.931 | INFO     | __main__:print_training_configuration:1071 - VLM reward score type                            : token_credit
+2026-06-29 02:27:20.931 | INFO     | __main__:print_training_configuration:1072 - VLM reward max new tokens                        : 128
+2026-06-29 02:27:20.932 | INFO     | __main__:print_training_configuration:1073 - VLM reward batch size                            : 1
+2026-06-29 02:27:20.932 | INFO     | __main__:print_training_configuration:1074 - VLM reward missing score                         : 0.0
+2026-06-29 02:27:20.933 | INFO     | __main__:print_training_configuration:1077 - KL weight                                        : 1e-05
+2026-06-29 02:27:20.933 | INFO     | __main__:print_training_configuration:1078 - KL coef                                          : 1e-07
+2026-06-29 02:27:20.933 | INFO     | __main__:print_training_configuration:1079 - Use moving KL                                    : True
+2026-06-29 02:27:20.934 | INFO     | __main__:print_training_configuration:1080 - Update ref model step                            : 10
+2026-06-29 02:27:20.934 | INFO     | __main__:print_training_configuration:1081 - Use dual KL                                      : True
+2026-06-29 02:27:20.935 | INFO     | __main__:print_training_configuration:1082 - Dual KL moving weight                            : 1.0
+2026-06-29 02:27:20.935 | INFO     | __main__:print_training_configuration:1083 - Dual KL step weight                              : 0.1
+2026-06-29 02:27:20.935 | INFO     | __main__:print_training_configuration:1084 - KL compute mode                                  : rollout_phase
+2026-06-29 02:27:20.936 | INFO     | __main__:print_training_configuration:1085 - Use grad balancing                               : True
+2026-06-29 02:27:20.936 | INFO     | __main__:print_training_configuration:1086 - Enable timestep permutation                      : True
+2026-06-29 02:27:20.937 | INFO     | __main__:print_training_configuration:1087 - Reference mode offload                           : True
+2026-06-29 02:27:20.937 | INFO     | __main__:print_training_configuration:1089 - --------------------------------------------------------------------------------
+2026-06-29 02:27:20.937 | INFO     | __main__:print_training_configuration:1090 - LOGGING & CHECKPOINTING CONFIGURATION
+2026-06-29 02:27:20.938 | INFO     | __main__:print_training_configuration:1091 - --------------------------------------------------------------------------------
+2026-06-29 02:27:20.938 | INFO     | __main__:print_training_configuration:1092 - Output directory                                 : ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3
+2026-06-29 02:27:20.938 | INFO     | __main__:print_training_configuration:1093 - Checkpoints directory                            : s3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3
+2026-06-29 02:27:20.939 | INFO     | __main__:print_training_configuration:1094 - Checkpoint interval                              : 25
+2026-06-29 02:27:20.939 | INFO     | __main__:print_training_configuration:1095 - Timer checkpoint interval                        : 0
+2026-06-29 02:27:20.940 | INFO     | __main__:print_training_configuration:1096 - Auto resume from timer/checkpoint                : True
+2026-06-29 02:27:20.941 | INFO     | __main__:print_training_configuration:1097 - Timer checkpoint directory                       : s3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/timer
+2026-06-29 02:27:20.941 | INFO     | __main__:print_training_configuration:1098 - Log interval                                     : 1
+2026-06-29 02:27:20.941 | INFO     | __main__:print_training_configuration:1099 - Resume checkpoint                                : None
+2026-06-29 02:27:20.942 | INFO     | __main__:print_training_configuration:1100 - Train CSV                                         : ['/nfs/ywang29/Reward_finetuning/SAGE-GRPO/assets/demo_train_decomposed_binary5.csv']
+2026-06-29 02:27:20.942 | INFO     | __main__:print_training_configuration:1101 - Valid CSV                                         : assets/demo_sample.csv
+2026-06-29 02:27:20.943 | INFO     | __main__:print_training_configuration:1102 - Seed                                             : 42
+2026-06-29 02:27:20.943 | INFO     | __main__:print_training_configuration:1103 - WandB enabled                                    : True
+2026-06-29 02:27:20.944 | INFO     | __main__:print_training_configuration:1105 - WandB project                                    : SAGE-GRPO
+2026-06-29 02:27:20.944 | INFO     | __main__:print_training_configuration:1106 - WandB run name                                   : yw006-4-sage-qwen35-9b-token-credit-video-adv-window3
+2026-06-29 02:27:20.944 | INFO     | __main__:__init__:958 - [Effective Config] world_size=64, rank=0, local_rank=0, sp_size=2, dp_size=32, dp_rank=0, sp_rank=0, batch_size=2, num_generations=4, grad_accum=1
+2026-06-29 02:27:20.945 | INFO     | __main__:_build_video_dataset:2088 - [Effective GRPO Grouping] video_batch_size=2, dp_replicas=32, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, product=64
+2026-06-29 02:27:21.616 | INFO     | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1
+2026-06-29 02:27:21.617 | INFO     | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done
+2026-06-29 02:27:21.617 | INFO     | __main__:_train_grpo:2171 - Starting GRPO training...
+2026-06-29 02:27:21.618 | INFO     | __main__:_train_grpo:2172 - Max steps: 10000
+2026-06-29 02:27:21.618 | INFO     | __main__:_train_grpo:2173 - Batch size: 2
+2026-06-29 02:27:21.619 | INFO     | __main__:_train_grpo:2174 - Learning rate: 1e-05
+2026-06-29 02:27:22.593 | INFO     | __main__:_resolve_resume_checkpoint:1860 - No timer checkpoint found at s3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/timer; auto-resuming from latest regular checkpoint: s3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/checkpoint-150
+2026-06-29 02:27:22.736 | INFO     | __main__:load_checkpoint:1992 - Loading checkpoint from s3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/checkpoint-150
+/usr/local/lib/python3.12/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)

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260629_022719-rgt7q8d2/files/requirements.txt

@@ -0,0 +1,287 @@
+awscli==1.45.30
+boto3==1.43.30
+botocore==1.43.30
+certifi==2026.5.20
+charset-normalizer==3.4.7
+colorama==0.4.6
+docutils==0.19
+idna==3.18
+jmespath==1.1.0
+pip==26.1.2
+pyasn1==0.6.3
+python-dateutil==2.9.0.post0
+PyYAML==6.0.3
+requests==2.34.2
+rsa==4.7.2
+s3transfer==0.18.0
+six==1.17.0
+urllib3==2.7.0
+py-cpuinfo==9.0.0
+hjson==3.1.0
+easydict==1.13
+braceexpand==0.1.7
+antlr4-python3-runtime==4.9.3
+websocket-client==1.9.0
+webdataset==1.0.2
+triton==3.6.0
+torchcodec==0.14.0+cu130
+tifffile==2026.6.1
+tenacity==9.1.4
+smmap==5.0.3
+sentry-sdk==2.63.0
+sentencepiece==0.2.1
+regex==2026.6.28
+pyarrow==24.0.0
+pyaml==26.2.1
+propcache==0.5.2
+opencv-python-headless==4.13.0.92
+omegaconf==2.3.1
+nvidia-nvtx==13.0.85
+ImageIO==2.37.3
+nvidia-nvshmem-cu13==3.4.5
+nvidia-nvjitlink==13.0.88
+av==17.1.0
+nvidia-nccl-cu13==2.28.9
+nvidia-cufile==1.15.1.6
+imageio-ffmpeg==0.6.0
+nvidia-cuda-runtime==13.0.96
+nvidia-cuda-nvrtc==13.0.88
+attrs==26.1.0
+nvidia-cuda-cupti==13.0.85
+nvidia-cublas==13.1.0.3
+multidict==6.7.1
+msgpack==1.2.1
+Mako==1.3.12
+loguru==0.7.3
+lmdb==2.2.1
+lightning-utilities==0.15.3
+lazy-loader==0.5
+greenlet==3.5.3
+future==1.0.0
+ftfy==6.3.1
+frozenlist==1.8.0
+frozendict==2.4.7
+docstring_parser==0.18.0
+decord==0.6.0
+cuda-pathfinder==1.5.5
+aiohappyeyeballs==2.6.2
+yarl==1.24.2
+typed-argument-parser==1.12.0
+SQLAlchemy==2.0.51
+scikit-image==0.26.0
+qwen-vl-utils==0.0.14
+typeguard==4.5.2
+nvidia-cusparse==12.6.3.3
+nvidia-cufft==12.0.0.61
+nvidia-cudnn-cu13==9.15.1.9
+gitdb==4.0.12
+ffmpeg-python==0.2.0
+einx==0.4.3
+cupy-cuda12x==14.1.1
+cuda-bindings==13.0.3
+cattrs==26.1.0
+aiosignal==1.4.0
+tokenizers==0.22.2
+termcolor==3.3.0
+nvidia-cusolver==12.0.4.66
+mediapy==1.2.6
+ipdb==0.13.13
+GitPython==3.1.50
+diffusers==0.38.0
+coremltools==9.0
+alembic==1.18.5
+aiohttp==3.14.1
+wandb==0.28.0
+torch==2.10.0+cu130
+dataset==2.0.0
+dashscope==1.26.0
+torchvision==0.25.0+cu130
+torchmetrics==1.9.0
+sniffio==1.3.1
+torchdata==0.8.0
+tqdm==4.67.3
+torchaudio==2.10.0+cu130
+deepspeed==0.19.2
+timm==1.0.27
+pytorch-lightning==2.6.5
+peft==0.19.1
+lpips==0.1.4
+lightning==2.6.5
+flash_attn==2.8.3+cu130torch2.10
+xxhash==3.8.0
+shortuuid==1.0.13
+setuptools==82.0.0
+safetensors==0.7.0
+pillow==12.1.1
+numpy==2.4.2
+jiter==0.15.0
+fsspec==2026.2.0
+distro==1.9.0
+dill==0.4.1
+tyro==1.0.15
+tiktoken==0.13.0
+multiprocess==0.70.19
+modelscope==1.36.0
+fire==0.7.0
+openai==2.8.0
+datasets==4.8.4
+trl==0.8.6
+angelslim==0.2.2
+huggingface_hub==1.21.0
+transformers==5.12.1
+black==26.5.1
+pathspec==1.1.1
+pytokens==0.4.1
+importlib_metadata==9.0.0
+ml_dtypes==0.5.4
+onnx==1.22.0
+onnx-ir==0.2.1
+onnxscript==0.7.0
+transformer_engine==2.9.0+70f53666
+zipp==4.1.0
+nvidia-mathdx==25.6.0
+einops==0.8.2
+absl-py==2.4.0
+accelerate==1.14.0
+annotated-doc==0.0.4
+annotated-types==0.7.0
+anyio==4.14.0
+apsw==3.53.2.0
+apswutils==0.1.2
+beartype==0.22.9
+beautifulsoup4==4.15.0
+blis==1.2.1
+catalogue==2.0.10
+cloudpathlib==0.24.0
+cloudpickle==3.1.2
+confection==0.1.5
+contourpy==1.3.3
+cycler==0.12.1
+cymem==2.0.13
+fastai==2.8.7
+fastcore==1.13.6
+fastdownload==0.0.7
+fastlite==0.2.4
+fastprogress==1.1.6
+fasttransform==0.0.2
+fonttools==4.63.0
+grpcio==1.81.1
+h11==0.16.0
+hf-xet==1.5.1
+httpcore==1.0.9
+httpcore2==2.4.0
+httptools==0.8.0
+httpx==0.28.1
+httpx2==2.4.0
+itsdangerous==2.2.0
+joblib==1.5.3
+kiwisolver==1.5.0
+langcodes==3.5.1
+Markdown==3.10.2
+markdown-it-py==4.2.0
+matplotlib==3.11.0
+mdurl==0.1.2
+mpmath==1.3.0
+murmurhash==1.0.15
+mypy_extensions==1.1.0
+narwhals==2.22.1
+networkx==3.6.1
+nvidia-curand==10.4.0.35
+nvidia-cusparselt-cu13==0.8.0
+oauthlib==3.3.1
+pandas==3.0.3
+plum-dispatch==2.9.0
+preshed==3.0.13
+protobuf==7.35.1
+pydantic==2.13.4
+pydantic_core==2.46.4
+pyparsing==3.3.2
+pyre-extensions==0.0.32
+python-dotenv==1.2.2
+python-fasthtml==0.14.3
+python-multipart==0.0.32
+rich==15.0.0
+s3torchconnector==1.5.0
+s3torchconnectorclient==1.5.0
+scikit-learn==1.9.0
+shellingham==1.5.4
+smart_open==7.6.1
+soupsieve==2.8.4
+spacy==3.8.7
+spacy-legacy==3.0.12
+spacy-loggers==1.0.5
+srsly==2.5.3
+starlette==1.3.1
+sympy==1.14.0
+tabulate==0.10.0
+tensorboard==2.20.0
+tensorboard-data-server==0.7.2
+thinc==8.3.4
+threadpoolctl==3.6.0
+torchtnt==0.2.4
+truststore==0.10.4
+typer==0.25.1
+typer-slim==0.24.0
+typing-inspect==0.9.0
+typing-inspection==0.4.2
+uvicorn==0.49.0
+uvloop==0.22.1
+wasabi==1.1.3
+watchfiles==1.2.0
+weasel==0.4.3
+websockets==16.0
+Werkzeug==3.1.8
+wrapt==2.2.2
+asttokens==3.0.1
+cffi==2.0.0
+click==8.4.1
+comm==0.2.3
+cryptography==49.0.0
+Cython==3.2.5
+debugpy==1.8.21
+decorator==5.3.1
+executing==2.2.1
+filelock==3.29.4
+h5py==3.16.0
+intel-cmplr-lib-ur==2026.0.0
+intel-openmp==2026.0.0
+ipykernel==7.3.0
+ipython==9.14.1
+ipython_pygments_lexers==1.1.1
+jedi==0.20.0
+Jinja2==3.1.6
+jupyter_client==8.9.1
+jupyter_core==5.9.1
+MarkupSafe==3.0.3
+matplotlib-inline==0.2.2
+mkl==2026.0.0
+mkl-include==2026.0.0
+mpi4py==4.1.2
+nest-asyncio2==1.7.2
+ninja==1.13.0
+onemkl-license==2026.0.0
+opencv-python==4.11.0.86
+packaging==26.2
+parso==0.8.7
+pexpect==4.9.0
+platformdirs==4.10.0
+prompt_toolkit==3.0.52
+psutil==7.2.2
+ptyprocess==0.7.0
+pure_eval==0.2.3
+pybind11==3.0.4
+pycparser==3.0
+Pygments==2.20.0
+pyOpenSSL==26.3.0
+pyzmq==27.1.0
+scipy==1.18.0
+stack-data==0.6.3
+tbb==2023.0.0
+tcmlib==1.5.0
+tornado==6.5.7
+traitlets==5.15.1
+typing_extensions==4.15.0
+tzdata==2026.2
+umf==1.1.0
+wheel==0.47.0
+wcwidth==0.8.1

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260629_022719-rgt7q8d2/files/wandb-metadata.json

@@ -0,0 +1,185 @@
+{
+  "os": "Linux-6.1.163-186.299.amzn2023.x86_64-x86_64-with-glibc2.35",
+  "python": "CPython 3.12.10",
+  "startedAt": "2026-06-29T02:27:19.653235Z",
+  "args": [
+    "--pretrained_model_root",
+    "./ckpts",
+    "--train_video_csv",
+    "/nfs/ywang29/Reward_finetuning/SAGE-GRPO/assets/demo_train_decomposed_binary5.csv",
+    "--reward_model",
+    "snapvideo_vlm",
+    "--vlm_reward_model_path",
+    "/nfs/ywang29/ckpts/Qwen3.5-9B",
+    "--vlm_reward_model_family",
+    "qwen3_5",
+    "--vlm_reward_num_frames",
+    "20",
+    "--vlm_reward_max_pixels",
+    "50176",
+    "--vlm_reward_max_new_tokens",
+    "128",
+    "--vlm_reward_batch_size",
+    "1",
+    "--vlm_reward_score_type",
+    "token_credit",
+    "--vlm_reward_token_credit_max_questions",
+    "5",
+    "--token_credit_loss_mode",
+    "video_adv",
+    "--token_credit_ratio_mode",
+    "weighted_logprob",
+    "--token_credit_spatial_granularity",
+    "window",
+    "--token_credit_spatial_bins",
+    "16",
+    "--token_credit_spatial_window",
+    "3",
+    "--token_credit_chunk_size",
+    "1",
+    "--learning_rate",
+    "1e-5",
+    "--batch_size",
+    "2",
+    "--num_generations",
+    "4",
+    "--max_steps",
+    "10000",
+    "--output_dir",
+    "./outputs",
+    "--enable_wandb",
+    "True",
+    "--project_name",
+    "SAGE-GRPO",
+    "--run_name",
+    "yw006-4-sage-qwen35-9b-token-credit-video-adv-window3",
+    "--enable_fsdp",
+    "--enable_gradient_checkpointing",
+    "--sp_size",
+    "2",
+    "--reward_checkpoint_mode",
+    "v3",
+    "--validation_interval",
+    "0",
+    "--validate_at_step0",
+    "False",
+    "--validate_video_length",
+    "121",
+    "--validation_timestep_shift",
+    "5.0",
+    "--use_grad_balancing",
+    "True",
+    "--enable_timestep_permutation",
+    "True",
+    "--sde_type",
+    "sage_grpo",
+    "--kl_weight",
+    "1e-5",
+    "--kl_coef",
+    "1e-7",
+    "--use_moving_KL",
+    "True",
+    "--update_ref_model_step",
+    "10",
+    "--use_dual_kl",
+    "True",
+    "--dual_kl_moving_weight",
+    "1.0",
+    "--dual_kl_step_weight",
+    "0.1",
+    "--reference_mode_offload",
+    "True",
+    "--save_interval",
+    "25",
+    "--timer_interval",
+    "0",
+    "--auto_resume_from_timer",
+    "True",
+    "--checkpoints_directory",
+    "s3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3"
+  ],
+  "program": "/nfs/ywang29/Reward_finetuning/SAGE-GRPO/post_train.py",
+  "codePath": "post_train.py",
+  "codePathLocal": "post_train.py",
+  "git": {
+    "remote": "https://github.com/Tencent-Hunyuan/SAGE-GRPO.git",
+    "commit": "779776e42732e0262b3f2844dd206d9d8f043b2f"
+  },
+  "root": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb",
+  "host": "ywang29-hunyuan-sage-qwen35-tcva-w3-worker-0",
+  "username": "yi-fan-wang1216",
+  "executable": "/usr/local/bin/python",
+  "cpu_count": 48,
+  "cpu_count_logical": 96,
+  "gpu": "NVIDIA A100-SXM4-80GB",
+  "gpu_count": 8,
+  "disk": {
+    "/": {
+      "total": "7516777873408",
+      "used": "619661799424"
+    }
+  },
+  "memory": {
+    "total": "1204539252736"
+  },
+  "gpu_nvidia": [
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-f6eafbbf-b72f-1725-61a0-5da7fda21fe2"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-8faea6eb-ec0a-0561-c1a7-608821eb24f5"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-6d3845d3-5be7-3130-93f6-aa8bc1fc8f9b"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-a80bc0c5-98ed-a833-ec83-51e290f6ac9c"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-a8dcbb50-ac55-1c6b-f4a3-66a2e17a2e9f"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-a6632210-2f96-54eb-7c24-6689d8c3ab90"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-e2324dd0-03ea-7053-4212-42facf1acdf7"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-20e27620-2f39-65f1-ff0c-215b32505ff5"
+    }
+  ],
+  "cudaVersion": "13.0",
+  "writerId": "0u4g8qsrfznxf8wb91bx1uod0bac83d8"
+}

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260629_022719-rgt7q8d2/logs/debug-internal.log

@@ -0,0 +1,11 @@
+{"time":"2026-06-29T02:27:19.904585291Z","level":"INFO","msg":"wandb-core"}
+{"time":"2026-06-29T02:27:19.906362878Z","level":"INFO","msg":"stream: starting","core version":"0.28.0"}
+{"time":"2026-06-29T02:27:20.180379016Z","level":"INFO","msg":"stream: created new stream","id":"rgt7q8d2"}
+{"time":"2026-06-29T02:27:20.180497358Z","level":"INFO","msg":"handler: started"}
+{"time":"2026-06-29T02:27:20.183873081Z","level":"INFO","msg":"stream: started"}
+{"time":"2026-06-29T02:27:20.184018172Z","level":"INFO","msg":"writer: started","stream_id":"rgt7q8d2"}
+{"time":"2026-06-29T02:27:20.18403467Z","level":"INFO","msg":"sender: started"}
+{"time":"2026-06-29T02:27:20.884462636Z","level":"INFO","msg":"filestream: sending request","total_files":1,"console_offset":266,"console_lines":1}
+{"time":"2026-06-29T02:27:21.195900211Z","level":"INFO","msg":"filestream: request sent","status":"200 OK"}
+{"time":"2026-06-29T02:27:35.884778037Z","level":"INFO","msg":"filestream: sending request","total_files":2,"events_offset":18,"events_lines":2,"console_offset":267,"console_lines":132,"uploaded_len":2}
+{"time":"2026-06-29T02:27:36.107662068Z","level":"INFO","msg":"filestream: request sent","status":"200 OK"}

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260629_022719-rgt7q8d2/logs/debug.log

@@ -0,0 +1,20 @@
+2026-06-29 02:27:19,672 INFO    MainThread:5672 [wandb_setup.py:_flush():81] Current SDK version is 0.28.0
+2026-06-29 02:27:19,673 INFO    MainThread:5672 [wandb_setup.py:_flush():81] Configure stats pid to 5672
+2026-06-29 02:27:19,673 INFO    MainThread:5672 [wandb_setup.py:_flush():81] Loading settings from environment variables
+2026-06-29 02:27:19,673 INFO    MainThread:5672 [wandb_init.py:setup_run_log_directory():725] Logging user logs to outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260629_022719-rgt7q8d2/logs/debug.log
+2026-06-29 02:27:19,673 INFO    MainThread:5672 [wandb_init.py:setup_run_log_directory():726] Logging internal logs to outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260629_022719-rgt7q8d2/logs/debug-internal.log
+2026-06-29 02:27:19,673 INFO    MainThread:5672 [wandb_init.py:init():768] calling init triggers
+2026-06-29 02:27:19,673 INFO    MainThread:5672 [wandb_init.py:init():773] wandb.init called with sweep_config: {}
+config: {'pretrained_model_root': './ckpts', 'pretrained_transformer_version': '480p_t2v', 'post_train_type': 'grpo', 'learning_rate': 1e-05, 'weight_decay': 0.01, 'max_steps': 10000, 'warmup_steps': 500, 'gradient_accumulation_steps': 1, 'max_grad_norm': 1.0, 'use_muon': True, 'num_train_timesteps': 1000, 'train_timestep_shift': 3.0, 'validation_timestep_shift': 5.0, 'snr_type': 'lognorm', 'task_type': 't2v', 'i2v_prob': 0.3, 'enable_fsdp': True, 'enable_gradient_checkpointing': True, 'sp_size': 2, 'dp_replicate': 1, 'batch_size': 2, 'num_workers': 4, 'prefetch_factor': 2, 'num_generations': 4, 'use_same_noise': False, 'output_dir': './outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3', 'checkpoints_directory': 's3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3', 'save_interval': 25, 'timer_interval': 0, 'auto_resume_from_timer': True, 'log_interval': 1, 'enable_wandb': True, 'project_name': 'SAGE-GRPO', 'run_name': 'yw006-4-sage-qwen35-9b-token-credit-video-adv-window3', 'wandb_key': None, 'wandb_username': None, 'wandb_base_url': 'https://api.wandb.ai', 'wandb_entity': None, 'wandb_silent': True, 'dtype': 'bf16', 'master_weight_type': 'fp32', 'seed': 42, 'validation_global_seed': 930, 'validation_interval': 0, 'validation_prompts': None, 'validate_video_length': 121, 'validation_aspect_ratio': '16:9', 'validation_fixed_size': '480x864', 'validation_num_inference_steps': 40, 'validation_negative_prompt': '', 'validation_enable_sr': False, 'validation_prompt_rewrite': False, 'validate_at_step0': False, 'resume_from_checkpoint': None, 'use_lora': False, 'lora_r': 8, 'lora_alpha': 16, 'lora_dropout': 0.0, 'lora_target_modules': None, 'pretrained_lora_path': None, 'train_video_csv': ['/nfs/ywang29/Reward_finetuning/SAGE-GRPO/assets/demo_train_decomposed_binary5.csv'], 'valid_video_csv': 'assets/demo_sample.csv', 'reward_model': 'snapvideo_vlm', 'reward_config': None, 'reward_checkpoint_mode': 'v3', 'reward_weights': '{"VQ":0.5,"MQ":0.5,"TA":1.0}', 'reward_gradient_credit': False, 'reward_credit_num_frames': 0, 'reward_credit_max_questions': 0, 'reward_credit_metric': 'weighted', 'token_credit_source_model': None, 'token_credit_advantage_source': 'token', 'token_credit_remote_url': None, 'token_credit_vlm_model_path': None, 'token_credit_vlm_model_family': None, 'token_credit_vlm_prompt_template': None, 'remote_reward_url': None, 'vlm_reward_model_path': '/nfs/ywang29/ckpts/Qwen3.5-9B', 'vlm_reward_model_family': 'qwen3_5', 'vlm_reward_prompt_template': None, 'vlm_reward_num_frames': 20, 'vlm_reward_max_pixels': 50176, 'vlm_reward_max_new_tokens': 128, 'vlm_reward_batch_size': 1, 'vlm_reward_score_type': 'token_credit', 'vlm_reward_missing_score': 0.0, 'vlm_reward_video_sample_type': 'uniform', 'vlm_reward_token_credit_max_questions': 5, 'token_credit_loss_mode': 'video_adv', 'token_credit_ratio_mode': 'weighted_logprob', 'token_credit_shuffle_weights': False, 'token_credit_spatial_granularity': 'window', 'token_credit_spatial_bins': 16, 'token_credit_spatial_window': 3, 'token_credit_chunk_size': 1, 'enable_grpo_policy_update': True, 'grpo_sampling_steps': 20, 'sample_n_frames': 81, 'eta': 0.5, 'cfg_scale': 6.0, 'neg_prompt': None, 'data_type': 'video', 'model_type': '16164', 'video_fps': 24, 'vae_name': '16164-32c-hy20250605', 'vae_tiling': False, 'infer_flow_shift_video': 5.0, 'video_bucket_hw_base_size': 480, 'video_bucket_hw_bucket_stride': 32, 'multitask_mask_training_type': 'concat', 'vision_num_semantic_tokens': 729, 'vision_states_dim': 1152, 'sde_type': 'sage_grpo', 'mini_batch_size_per_update': 1, 'clip_range': 0.0001, 'adv_clip_max': 5.0, 'timestep_fraction': 1.0, 'random_train_timestep_count': False, 'random_train_timestep_min': 1, 'random_train_timestep_max': 20, 'use_grad_balancing': True, 'enable_timestep_permutation': True, 'reference_mode_offload': True, 'kl_weight': 1e-05, 'kl_coef': 1e-07, 'kl_min_coef': 1e-07, 'use_moving_KL': True, 'update_ref_model_step': 10, 'use_dual_kl': True, 'dual_kl_moving_weight': 1.0, 'dual_kl_step_weight': 0.1, 'kl_compute_mode': 'rollout_phase', 'glyph_byT5_v2': True, 'byt5_max_length': 256, 'video_uncond_p_byt5': 0.0, 'precision': 'bf16', 'vae_precision': 'fp16', 'val_disable_autocast': False, 'debug_grad_flow': False, 'debug_train_diagnostics_interval': 1, '_wandb': {}}
+2026-06-29 02:27:19,673 INFO    MainThread:5672 [wandb_init.py:init():816] starting backend
+2026-06-29 02:27:19,895 INFO    MainThread:5672 [wandb_init.py:init():831] sending inform_init request
+2026-06-29 02:27:20,184 INFO    MainThread:5672 [wandb_init.py:init():836] backend started and connected
+2026-06-29 02:27:20,187 INFO    MainThread:5672 [wandb_init.py:init():906] updated telemetry
+2026-06-29 02:27:20,219 INFO    MainThread:5672 [wandb_init.py:init():929] communicating run to backend with 90.0 second timeout
+2026-06-29 02:27:20,680 INFO    MainThread:5672 [wandb_init.py:init():969] run resumed
+2026-06-29 02:27:20,681 INFO    MainThread:5672 [wandb_init.py:init():974] starting run threads in backend
+2026-06-29 02:27:20,880 INFO    MainThread:5672 [wandb_run.py:_console_start():2523] atexit reg
+2026-06-29 02:27:20,880 INFO    MainThread:5672 [wandb_run.py:_redirect():2373] redirect: wrap_raw
+2026-06-29 02:27:20,881 INFO    MainThread:5672 [wandb_run.py:_redirect():2442] Wrapping output streams.
+2026-06-29 02:27:20,881 INFO    MainThread:5672 [wandb_run.py:_redirect():2465] Redirects installed.
+2026-06-29 02:27:20,882 INFO    MainThread:5672 [wandb_init.py:init():1012] run started, returning control to user process

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260629_024802-rgt7q8d2/files/output.log

@@ -0,0 +1,133 @@
+2026-06-29 02:48:04.032 | INFO     | __main__:print_training_configuration:970 - SYSTEM & HARDWARE CONFIGURATION
+2026-06-29 02:48:04.035 | INFO     | __main__:print_training_configuration:971 - --------------------------------------------------------------------------------
+2026-06-29 02:48:04.035 | INFO     | __main__:print_training_configuration:972 - Number of GPUs                                   : 64
+2026-06-29 02:48:04.036 | INFO     | __main__:print_training_configuration:973 - Local rank                                       : 0
+2026-06-29 02:48:04.036 | INFO     | __main__:print_training_configuration:974 - Global rank                                      : 0
+2026-06-29 02:48:04.037 | INFO     | __main__:print_training_configuration:975 - Device                                           : cuda:0
+2026-06-29 02:48:04.037 | INFO     | __main__:print_training_configuration:976 - Master weight dtype                              : torch.float32
+2026-06-29 02:48:04.038 | INFO     | __main__:print_training_configuration:978 - --------------------------------------------------------------------------------
+2026-06-29 02:48:04.038 | INFO     | __main__:print_training_configuration:979 - MODEL PARAMETERS & ARCHITECTURE
+2026-06-29 02:48:04.039 | INFO     | __main__:print_training_configuration:980 - --------------------------------------------------------------------------------
+2026-06-29 02:48:04.049 | INFO     | __main__:print_training_configuration:983 - Number of total                                  : 8,326,608,160
+2026-06-29 02:48:04.049 | INFO     | __main__:print_training_configuration:984 - Total trainable parameters                        : 8,326,608,160
+2026-06-29 02:48:04.050 | INFO     | __main__:print_training_configuration:985 - DIT model root                                   : ./ckpts
+2026-06-29 02:48:04.050 | INFO     | __main__:print_training_configuration:986 - Transformer version                              : 480p_t2v
+2026-06-29 02:48:04.051 | INFO     | __main__:print_training_configuration:987 - Attention mode                                   : flash
+2026-06-29 02:48:04.051 | INFO     | __main__:print_training_configuration:988 - Gradient checkpointing                           : True
+2026-06-29 02:48:04.052 | INFO     | __main__:print_training_configuration:989 - FSDP enabled                                     : True
+2026-06-29 02:48:04.052 | INFO     | __main__:print_training_configuration:990 - LoRA enabled                                     : False
+2026-06-29 02:48:04.052 | INFO     | __main__:print_training_configuration:991 - Glyph byT5 v2                                    : True
+2026-06-29 02:48:04.053 | INFO     | __main__:print_training_configuration:992 - ByT5 max length                                  : 256
+2026-06-29 02:48:04.053 | INFO     | __main__:print_training_configuration:994 - --------------------------------------------------------------------------------
+2026-06-29 02:48:04.054 | INFO     | __main__:print_training_configuration:995 - DISTRIBUTED TRAINING CONFIGURATION
+2026-06-29 02:48:04.054 | INFO     | __main__:print_training_configuration:996 - --------------------------------------------------------------------------------
+2026-06-29 02:48:04.055 | INFO     | __main__:print_training_configuration:997 - World size                                       : 64
+2026-06-29 02:48:04.055 | INFO     | __main__:print_training_configuration:998 - DP degree                                        : 32
+2026-06-29 02:48:04.055 | INFO     | __main__:print_training_configuration:999 - DP rank                                          : 0
+2026-06-29 02:48:04.056 | INFO     | __main__:print_training_configuration:1000 - SP size                                          : 2
+2026-06-29 02:48:04.056 | INFO     | __main__:print_training_configuration:1001 - SP rank                                          : 0
+2026-06-29 02:48:04.057 | INFO     | __main__:print_training_configuration:1003 - --------------------------------------------------------------------------------
+2026-06-29 02:48:04.057 | INFO     | __main__:print_training_configuration:1004 - BATCH SIZE & DATA FLOW CONFIGURATION
+2026-06-29 02:48:04.057 | INFO     | __main__:print_training_configuration:1005 - --------------------------------------------------------------------------------
+2026-06-29 02:48:04.058 | INFO     | __main__:print_training_configuration:1006 - Per GPU batch size                               : 2
+2026-06-29 02:48:04.058 | INFO     | __main__:print_training_configuration:1007 - DP degree batch size                             : 64
+2026-06-29 02:48:04.059 | INFO     | __main__:print_training_configuration:1008 - Gradient accumulation steps                      : 1
+2026-06-29 02:48:04.059 | INFO     | __main__:print_training_configuration:1009 - Num generations                                  : 4
+2026-06-29 02:48:04.059 | INFO     | __main__:print_training_configuration:1010 - Use same noise                                   : False
+2026-06-29 02:48:04.060 | INFO     | __main__:print_training_configuration:1011 - Num workers                                      : 4
+2026-06-29 02:48:04.060 | INFO     | __main__:print_training_configuration:1012 - Prefetch factor                                  : 2
+2026-06-29 02:48:04.061 | INFO     | __main__:print_training_configuration:1014 - --------------------------------------------------------------------------------
+2026-06-29 02:48:04.061 | INFO     | __main__:print_training_configuration:1015 - OPTIMIZATION CONFIGURATION
+2026-06-29 02:48:04.061 | INFO     | __main__:print_training_configuration:1016 - --------------------------------------------------------------------------------
+2026-06-29 02:48:04.062 | INFO     | __main__:print_training_configuration:1017 - Optimizer                                        : muon
+2026-06-29 02:48:04.062 | INFO     | __main__:print_training_configuration:1018 - Learning rate                                    : 1e-05
+2026-06-29 02:48:04.063 | INFO     | __main__:print_training_configuration:1019 - Weight decay                                     : 0.01
+2026-06-29 02:48:04.063 | INFO     | __main__:print_training_configuration:1020 - Max gradient norm                                : 1.0
+2026-06-29 02:48:04.064 | INFO     | __main__:print_training_configuration:1021 - LR scheduler                                     : constant
+2026-06-29 02:48:04.064 | INFO     | __main__:print_training_configuration:1022 - Warmup steps                                     : 500
+2026-06-29 02:48:04.064 | INFO     | __main__:print_training_configuration:1023 - Max training steps                               : 10000
+2026-06-29 02:48:04.065 | INFO     | __main__:print_training_configuration:1025 - --------------------------------------------------------------------------------
+2026-06-29 02:48:04.065 | INFO     | __main__:print_training_configuration:1026 - INFERENCE & VALIDATION CONFIGURATION
+2026-06-29 02:48:04.066 | INFO     | __main__:print_training_configuration:1027 - --------------------------------------------------------------------------------
+2026-06-29 02:48:04.066 | INFO     | __main__:print_training_configuration:1028 - Validation interval                              : 0
+2026-06-29 02:48:04.066 | INFO     | __main__:print_training_configuration:1029 - Validate at step 0                               : False
+2026-06-29 02:48:04.067 | INFO     | __main__:print_training_configuration:1030 - Validation video length                          : 121
+2026-06-29 02:48:04.067 | INFO     | __main__:print_training_configuration:1031 - Validation global seed                           : 930
+2026-06-29 02:48:04.068 | INFO     | __main__:print_training_configuration:1032 - Validation inference steps                        : 40
+2026-06-29 02:48:04.068 | INFO     | __main__:print_training_configuration:1033 - Validation aspect ratio                          : 16:9
+2026-06-29 02:48:04.068 | INFO     | __main__:print_training_configuration:1034 - Validation fixed size                            : 480x864
+2026-06-29 02:48:04.069 | INFO     | __main__:print_training_configuration:1035 - Validation prompt rewrite                         : False
+2026-06-29 02:48:04.069 | INFO     | __main__:print_training_configuration:1036 - Validation flow shift                             : 5.0
+2026-06-29 02:48:04.070 | INFO     | __main__:print_training_configuration:1037 - Sampling CFG scale                                : 6.0
+2026-06-29 02:48:04.070 | INFO     | __main__:print_training_configuration:1038 - Sampling eta                                      : 0.5
+2026-06-29 02:48:04.070 | INFO     | __main__:print_training_configuration:1039 - Sampling steps (GRPO rollout)                     : 20
+2026-06-29 02:48:04.071 | INFO     | __main__:print_training_configuration:1040 - Token credit loss mode                            : video_adv
+2026-06-29 02:48:04.071 | INFO     | __main__:print_training_configuration:1041 - Token credit ratio mode                           : weighted_logprob
+2026-06-29 02:48:04.071 | INFO     | __main__:print_training_configuration:1042 - Token credit shuffle weights                      : False
+2026-06-29 02:48:04.072 | INFO     | __main__:print_training_configuration:1043 - Token credit spatial granularity                  : window
+2026-06-29 02:48:04.072 | INFO     | __main__:print_training_configuration:1044 - Token credit spatial bins                         : 16
+2026-06-29 02:48:04.073 | INFO     | __main__:print_training_configuration:1045 - Token credit spatial window                       : 3
+2026-06-29 02:48:04.073 | INFO     | __main__:print_training_configuration:1046 - Token credit chunk size                           : 1
+2026-06-29 02:48:04.073 | INFO     | __main__:print_training_configuration:1047 - Token credit source model                         : None
+2026-06-29 02:48:04.074 | INFO     | __main__:print_training_configuration:1048 - Token credit advantage source                     : token
+2026-06-29 02:48:04.074 | INFO     | __main__:print_training_configuration:1049 - Token credit VLM model path                       : None
+2026-06-29 02:48:04.075 | INFO     | __main__:print_training_configuration:1050 - Token credit VLM model family                     : None
+2026-06-29 02:48:04.075 | INFO     | __main__:print_training_configuration:1051 - Random train timestep count                       : False
+2026-06-29 02:48:04.075 | INFO     | __main__:print_training_configuration:1052 - Random train timestep range                       : 1-20
+2026-06-29 02:48:04.076 | INFO     | __main__:print_training_configuration:1056 - SDE type                                          : sage_grpo
+2026-06-29 02:48:04.076 | INFO     | __main__:print_training_configuration:1058 - --------------------------------------------------------------------------------
+2026-06-29 02:48:04.077 | INFO     | __main__:print_training_configuration:1059 - GRPO & KL CONFIGURATION
+2026-06-29 02:48:04.077 | INFO     | __main__:print_training_configuration:1060 - --------------------------------------------------------------------------------
+2026-06-29 02:48:04.077 | INFO     | __main__:print_training_configuration:1061 - Reward model                                     : snapvideo_vlm
+2026-06-29 02:48:04.078 | INFO     | __main__:print_training_configuration:1062 - Reward gradient credit                          : False
+2026-06-29 02:48:04.078 | INFO     | __main__:print_training_configuration:1063 - Reward credit frames                            : 0
+2026-06-29 02:48:04.078 | INFO     | __main__:print_training_configuration:1064 - Reward credit max questions                     : 0
+2026-06-29 02:48:04.079 | INFO     | __main__:print_training_configuration:1065 - Reward credit metric                            : weighted
+2026-06-29 02:48:04.079 | INFO     | __main__:print_training_configuration:1067 - VLM reward model path                            : /nfs/ywang29/ckpts/Qwen3.5-9B
+2026-06-29 02:48:04.079 | INFO     | __main__:print_training_configuration:1068 - VLM reward model family                          : qwen3_5
+2026-06-29 02:48:04.080 | INFO     | __main__:print_training_configuration:1069 - VLM reward frames                                : 20
+2026-06-29 02:48:04.080 | INFO     | __main__:print_training_configuration:1070 - VLM reward max pixels                            : 50176
+2026-06-29 02:48:04.081 | INFO     | __main__:print_training_configuration:1071 - VLM reward score type                            : token_credit
+2026-06-29 02:48:04.081 | INFO     | __main__:print_training_configuration:1072 - VLM reward max new tokens                        : 128
+2026-06-29 02:48:04.081 | INFO     | __main__:print_training_configuration:1073 - VLM reward batch size                            : 1
+2026-06-29 02:48:04.082 | INFO     | __main__:print_training_configuration:1074 - VLM reward missing score                         : 0.0
+2026-06-29 02:48:04.082 | INFO     | __main__:print_training_configuration:1077 - KL weight                                        : 1e-05
+2026-06-29 02:48:04.083 | INFO     | __main__:print_training_configuration:1078 - KL coef                                          : 1e-07
+2026-06-29 02:48:04.083 | INFO     | __main__:print_training_configuration:1079 - Use moving KL                                    : True
+2026-06-29 02:48:04.083 | INFO     | __main__:print_training_configuration:1080 - Update ref model step                            : 10
+2026-06-29 02:48:04.084 | INFO     | __main__:print_training_configuration:1081 - Use dual KL                                      : True
+2026-06-29 02:48:04.084 | INFO     | __main__:print_training_configuration:1082 - Dual KL moving weight                            : 1.0
+2026-06-29 02:48:04.085 | INFO     | __main__:print_training_configuration:1083 - Dual KL step weight                              : 0.1
+2026-06-29 02:48:04.085 | INFO     | __main__:print_training_configuration:1084 - KL compute mode                                  : rollout_phase
+2026-06-29 02:48:04.085 | INFO     | __main__:print_training_configuration:1085 - Use grad balancing                               : True
+2026-06-29 02:48:04.086 | INFO     | __main__:print_training_configuration:1086 - Enable timestep permutation                      : True
+2026-06-29 02:48:04.086 | INFO     | __main__:print_training_configuration:1087 - Reference mode offload                           : True
+2026-06-29 02:48:04.087 | INFO     | __main__:print_training_configuration:1089 - --------------------------------------------------------------------------------
+2026-06-29 02:48:04.087 | INFO     | __main__:print_training_configuration:1090 - LOGGING & CHECKPOINTING CONFIGURATION
+2026-06-29 02:48:04.087 | INFO     | __main__:print_training_configuration:1091 - --------------------------------------------------------------------------------
+2026-06-29 02:48:04.088 | INFO     | __main__:print_training_configuration:1092 - Output directory                                 : ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3
+2026-06-29 02:48:04.088 | INFO     | __main__:print_training_configuration:1093 - Checkpoints directory                            : s3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3
+2026-06-29 02:48:04.089 | INFO     | __main__:print_training_configuration:1094 - Checkpoint interval                              : 25
+2026-06-29 02:48:04.089 | INFO     | __main__:print_training_configuration:1095 - Timer checkpoint interval                        : 0
+2026-06-29 02:48:04.089 | INFO     | __main__:print_training_configuration:1096 - Auto resume from timer/checkpoint                : True
+2026-06-29 02:48:04.090 | INFO     | __main__:print_training_configuration:1097 - Timer checkpoint directory                       : s3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/timer
+2026-06-29 02:48:04.090 | INFO     | __main__:print_training_configuration:1098 - Log interval                                     : 1
+2026-06-29 02:48:04.091 | INFO     | __main__:print_training_configuration:1099 - Resume checkpoint                                : None
+2026-06-29 02:48:04.091 | INFO     | __main__:print_training_configuration:1100 - Train CSV                                         : ['/nfs/ywang29/Reward_finetuning/SAGE-GRPO/assets/demo_train_decomposed_binary5.csv']
+2026-06-29 02:48:04.092 | INFO     | __main__:print_training_configuration:1101 - Valid CSV                                         : assets/demo_sample.csv
+2026-06-29 02:48:04.092 | INFO     | __main__:print_training_configuration:1102 - Seed                                             : 42
+2026-06-29 02:48:04.093 | INFO     | __main__:print_training_configuration:1103 - WandB enabled                                    : True
+2026-06-29 02:48:04.093 | INFO     | __main__:print_training_configuration:1105 - WandB project                                    : SAGE-GRPO
+2026-06-29 02:48:04.094 | INFO     | __main__:print_training_configuration:1106 - WandB run name                                   : yw006-4-sage-qwen35-9b-token-credit-video-adv-window3
+2026-06-29 02:48:04.094 | INFO     | __main__:__init__:958 - [Effective Config] world_size=64, rank=0, local_rank=0, sp_size=2, dp_size=32, dp_rank=0, sp_rank=0, batch_size=2, num_generations=4, grad_accum=1
+2026-06-29 02:48:04.095 | INFO     | __main__:_build_video_dataset:2088 - [Effective GRPO Grouping] video_batch_size=2, dp_replicas=32, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, product=64
+2026-06-29 02:48:04.758 | INFO     | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1
+2026-06-29 02:48:04.759 | INFO     | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done
+2026-06-29 02:48:04.759 | INFO     | __main__:_train_grpo:2171 - Starting GRPO training...
+2026-06-29 02:48:04.760 | INFO     | __main__:_train_grpo:2172 - Max steps: 10000
+2026-06-29 02:48:04.760 | INFO     | __main__:_train_grpo:2173 - Batch size: 2
+2026-06-29 02:48:04.761 | INFO     | __main__:_train_grpo:2174 - Learning rate: 1e-05
+2026-06-29 02:48:05.681 | INFO     | __main__:_resolve_resume_checkpoint:1860 - No timer checkpoint found at s3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/timer; auto-resuming from latest regular checkpoint: s3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/checkpoint-150
+2026-06-29 02:48:05.827 | INFO     | __main__:load_checkpoint:1992 - Loading checkpoint from s3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/checkpoint-150
+/usr/local/lib/python3.12/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)

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260629_024802-rgt7q8d2/files/requirements.txt

@@ -0,0 +1,287 @@
+awscli==1.45.30
+boto3==1.43.30
+botocore==1.43.30
+certifi==2026.5.20
+charset-normalizer==3.4.7
+colorama==0.4.6
+docutils==0.19
+idna==3.18
+jmespath==1.1.0
+pip==26.1.2
+pyasn1==0.6.3
+python-dateutil==2.9.0.post0
+PyYAML==6.0.3
+requests==2.34.2
+rsa==4.7.2
+s3transfer==0.18.0
+six==1.17.0
+urllib3==2.7.0
+py-cpuinfo==9.0.0
+hjson==3.1.0
+easydict==1.13
+braceexpand==0.1.7
+antlr4-python3-runtime==4.9.3
+websocket-client==1.9.0
+webdataset==1.0.2
+triton==3.6.0
+torchcodec==0.14.0+cu130
+tifffile==2026.6.1
+tenacity==9.1.4
+smmap==5.0.3
+sentry-sdk==2.63.0
+sentencepiece==0.2.1
+regex==2026.6.28
+pyarrow==24.0.0
+pyaml==26.2.1
+propcache==0.5.2
+opencv-python-headless==4.13.0.92
+omegaconf==2.3.1
+nvidia-nvtx==13.0.85
+ImageIO==2.37.3
+nvidia-nvshmem-cu13==3.4.5
+nvidia-nvjitlink==13.0.88
+av==17.1.0
+nvidia-nccl-cu13==2.28.9
+nvidia-cufile==1.15.1.6
+imageio-ffmpeg==0.6.0
+nvidia-cuda-runtime==13.0.96
+nvidia-cuda-nvrtc==13.0.88
+attrs==26.1.0
+nvidia-cuda-cupti==13.0.85
+nvidia-cublas==13.1.0.3
+multidict==6.7.1
+msgpack==1.2.1
+Mako==1.3.12
+loguru==0.7.3
+lmdb==2.2.1
+lightning-utilities==0.15.3
+lazy-loader==0.5
+greenlet==3.5.3
+future==1.0.0
+ftfy==6.3.1
+frozenlist==1.8.0
+frozendict==2.4.7
+docstring_parser==0.18.0
+decord==0.6.0
+cuda-pathfinder==1.5.5
+aiohappyeyeballs==2.6.2
+yarl==1.24.2
+typed-argument-parser==1.12.0
+SQLAlchemy==2.0.51
+scikit-image==0.26.0
+qwen-vl-utils==0.0.14
+typeguard==4.5.2
+nvidia-cusparse==12.6.3.3
+nvidia-cufft==12.0.0.61
+nvidia-cudnn-cu13==9.15.1.9
+gitdb==4.0.12
+ffmpeg-python==0.2.0
+einx==0.4.3
+cupy-cuda12x==14.1.1
+cuda-bindings==13.0.3
+cattrs==26.1.0
+aiosignal==1.4.0
+tokenizers==0.22.2
+termcolor==3.3.0
+nvidia-cusolver==12.0.4.66
+mediapy==1.2.6
+ipdb==0.13.13
+GitPython==3.1.50
+diffusers==0.38.0
+coremltools==9.0
+alembic==1.18.5
+aiohttp==3.14.1
+wandb==0.28.0
+torch==2.10.0+cu130
+dataset==2.0.0
+dashscope==1.26.0
+torchvision==0.25.0+cu130
+torchmetrics==1.9.0
+sniffio==1.3.1
+torchdata==0.8.0
+tqdm==4.67.3
+torchaudio==2.10.0+cu130
+deepspeed==0.19.2
+timm==1.0.27
+pytorch-lightning==2.6.5
+peft==0.19.1
+lpips==0.1.4
+lightning==2.6.5
+flash_attn==2.8.3+cu130torch2.10
+xxhash==3.8.0
+shortuuid==1.0.13
+setuptools==82.0.0
+safetensors==0.7.0
+pillow==12.1.1
+numpy==2.4.2
+jiter==0.15.0
+fsspec==2026.2.0
+distro==1.9.0
+dill==0.4.1
+tyro==1.0.15
+tiktoken==0.13.0
+multiprocess==0.70.19
+modelscope==1.36.0
+fire==0.7.0
+openai==2.8.0
+datasets==4.8.4
+trl==0.8.6
+angelslim==0.2.2
+huggingface_hub==1.21.0
+transformers==5.12.1
+black==26.5.1
+pathspec==1.1.1
+pytokens==0.4.1
+importlib_metadata==9.0.0
+ml_dtypes==0.5.4
+onnx==1.22.0
+onnx-ir==0.2.1
+onnxscript==0.7.0
+transformer_engine==2.9.0+70f53666
+zipp==4.1.0
+nvidia-mathdx==25.6.0
+einops==0.8.2
+absl-py==2.4.0
+accelerate==1.14.0
+annotated-doc==0.0.4
+annotated-types==0.7.0
+anyio==4.14.0
+apsw==3.53.2.0
+apswutils==0.1.2
+beartype==0.22.9
+beautifulsoup4==4.15.0
+blis==1.2.1
+catalogue==2.0.10
+cloudpathlib==0.24.0
+cloudpickle==3.1.2
+confection==0.1.5
+contourpy==1.3.3
+cycler==0.12.1
+cymem==2.0.13
+fastai==2.8.7
+fastcore==1.13.6
+fastdownload==0.0.7
+fastlite==0.2.4
+fastprogress==1.1.6
+fasttransform==0.0.2
+fonttools==4.63.0
+grpcio==1.81.1
+h11==0.16.0
+hf-xet==1.5.1
+httpcore==1.0.9
+httpcore2==2.4.0
+httptools==0.8.0
+httpx==0.28.1
+httpx2==2.4.0
+itsdangerous==2.2.0
+joblib==1.5.3
+kiwisolver==1.5.0
+langcodes==3.5.1
+Markdown==3.10.2
+markdown-it-py==4.2.0
+matplotlib==3.11.0
+mdurl==0.1.2
+mpmath==1.3.0
+murmurhash==1.0.15
+mypy_extensions==1.1.0
+narwhals==2.22.1
+networkx==3.6.1
+nvidia-curand==10.4.0.35
+nvidia-cusparselt-cu13==0.8.0
+oauthlib==3.3.1
+pandas==3.0.3
+plum-dispatch==2.9.0
+preshed==3.0.13
+protobuf==7.35.1
+pydantic==2.13.4
+pydantic_core==2.46.4
+pyparsing==3.3.2
+pyre-extensions==0.0.32
+python-dotenv==1.2.2
+python-fasthtml==0.14.3
+python-multipart==0.0.32
+rich==15.0.0
+s3torchconnector==1.5.0
+s3torchconnectorclient==1.5.0
+scikit-learn==1.9.0
+shellingham==1.5.4
+smart_open==7.6.1
+soupsieve==2.8.4
+spacy==3.8.7
+spacy-legacy==3.0.12
+spacy-loggers==1.0.5
+srsly==2.5.3
+starlette==1.3.1
+sympy==1.14.0
+tabulate==0.10.0
+tensorboard==2.20.0
+tensorboard-data-server==0.7.2
+thinc==8.3.4
+threadpoolctl==3.6.0
+torchtnt==0.2.4
+truststore==0.10.4
+typer==0.25.1
+typer-slim==0.24.0
+typing-inspect==0.9.0
+typing-inspection==0.4.2
+uvicorn==0.49.0
+uvloop==0.22.1
+wasabi==1.1.3
+watchfiles==1.2.0
+weasel==0.4.3
+websockets==16.0
+Werkzeug==3.1.8
+wrapt==2.2.2
+asttokens==3.0.1
+cffi==2.0.0
+click==8.4.1
+comm==0.2.3
+cryptography==49.0.0
+Cython==3.2.5
+debugpy==1.8.21
+decorator==5.3.1
+executing==2.2.1
+filelock==3.29.4
+h5py==3.16.0
+intel-cmplr-lib-ur==2026.0.0
+intel-openmp==2026.0.0
+ipykernel==7.3.0
+ipython==9.14.1
+ipython_pygments_lexers==1.1.1
+jedi==0.20.0
+Jinja2==3.1.6
+jupyter_client==8.9.1
+jupyter_core==5.9.1
+MarkupSafe==3.0.3
+matplotlib-inline==0.2.2
+mkl==2026.0.0
+mkl-include==2026.0.0
+mpi4py==4.1.2
+nest-asyncio2==1.7.2
+ninja==1.13.0
+onemkl-license==2026.0.0
+opencv-python==4.11.0.86
+packaging==26.2
+parso==0.8.7
+pexpect==4.9.0
+platformdirs==4.10.0
+prompt_toolkit==3.0.52
+psutil==7.2.2
+ptyprocess==0.7.0
+pure_eval==0.2.3
+pybind11==3.0.4
+pycparser==3.0
+Pygments==2.20.0
+pyOpenSSL==26.3.0
+pyzmq==27.1.0
+scipy==1.18.0
+stack-data==0.6.3
+tbb==2023.0.0
+tcmlib==1.5.0
+tornado==6.5.7
+traitlets==5.15.1
+typing_extensions==4.15.0
+tzdata==2026.2
+umf==1.1.0
+wheel==0.47.0
+wcwidth==0.8.1

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260629_024802-rgt7q8d2/files/wandb-metadata.json

@@ -0,0 +1,185 @@
+{
+  "os": "Linux-6.1.163-186.299.amzn2023.x86_64-x86_64-with-glibc2.35",
+  "python": "CPython 3.12.10",
+  "startedAt": "2026-06-29T02:48:02.600291Z",
+  "args": [
+    "--pretrained_model_root",
+    "./ckpts",
+    "--train_video_csv",
+    "/nfs/ywang29/Reward_finetuning/SAGE-GRPO/assets/demo_train_decomposed_binary5.csv",
+    "--reward_model",
+    "snapvideo_vlm",
+    "--vlm_reward_model_path",
+    "/nfs/ywang29/ckpts/Qwen3.5-9B",
+    "--vlm_reward_model_family",
+    "qwen3_5",
+    "--vlm_reward_num_frames",
+    "20",
+    "--vlm_reward_max_pixels",
+    "50176",
+    "--vlm_reward_max_new_tokens",
+    "128",
+    "--vlm_reward_batch_size",
+    "1",
+    "--vlm_reward_score_type",
+    "token_credit",
+    "--vlm_reward_token_credit_max_questions",
+    "5",
+    "--token_credit_loss_mode",
+    "video_adv",
+    "--token_credit_ratio_mode",
+    "weighted_logprob",
+    "--token_credit_spatial_granularity",
+    "window",
+    "--token_credit_spatial_bins",
+    "16",
+    "--token_credit_spatial_window",
+    "3",
+    "--token_credit_chunk_size",
+    "1",
+    "--learning_rate",
+    "1e-5",
+    "--batch_size",
+    "2",
+    "--num_generations",
+    "4",
+    "--max_steps",
+    "10000",
+    "--output_dir",
+    "./outputs",
+    "--enable_wandb",
+    "True",
+    "--project_name",
+    "SAGE-GRPO",
+    "--run_name",
+    "yw006-4-sage-qwen35-9b-token-credit-video-adv-window3",
+    "--enable_fsdp",
+    "--enable_gradient_checkpointing",
+    "--sp_size",
+    "2",
+    "--reward_checkpoint_mode",
+    "v3",
+    "--validation_interval",
+    "0",
+    "--validate_at_step0",
+    "False",
+    "--validate_video_length",
+    "121",
+    "--validation_timestep_shift",
+    "5.0",
+    "--use_grad_balancing",
+    "True",
+    "--enable_timestep_permutation",
+    "True",
+    "--sde_type",
+    "sage_grpo",
+    "--kl_weight",
+    "1e-5",
+    "--kl_coef",
+    "1e-7",
+    "--use_moving_KL",
+    "True",
+    "--update_ref_model_step",
+    "10",
+    "--use_dual_kl",
+    "True",
+    "--dual_kl_moving_weight",
+    "1.0",
+    "--dual_kl_step_weight",
+    "0.1",
+    "--reference_mode_offload",
+    "True",
+    "--save_interval",
+    "25",
+    "--timer_interval",
+    "0",
+    "--auto_resume_from_timer",
+    "True",
+    "--checkpoints_directory",
+    "s3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3"
+  ],
+  "program": "/nfs/ywang29/Reward_finetuning/SAGE-GRPO/post_train.py",
+  "codePath": "post_train.py",
+  "codePathLocal": "post_train.py",
+  "git": {
+    "remote": "https://github.com/Tencent-Hunyuan/SAGE-GRPO.git",
+    "commit": "779776e42732e0262b3f2844dd206d9d8f043b2f"
+  },
+  "root": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb",
+  "host": "ywang29-hunyuan-sage-qwen35-tcva-w3-worker-0",
+  "username": "yi-fan-wang1216",
+  "executable": "/usr/local/bin/python",
+  "cpu_count": 48,
+  "cpu_count_logical": 96,
+  "gpu": "NVIDIA A100-SXM4-80GB",
+  "gpu_count": 8,
+  "disk": {
+    "/": {
+      "total": "7516777873408",
+      "used": "619662102528"
+    }
+  },
+  "memory": {
+    "total": "1204539252736"
+  },
+  "gpu_nvidia": [
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-f6eafbbf-b72f-1725-61a0-5da7fda21fe2"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-8faea6eb-ec0a-0561-c1a7-608821eb24f5"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-6d3845d3-5be7-3130-93f6-aa8bc1fc8f9b"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-a80bc0c5-98ed-a833-ec83-51e290f6ac9c"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-a8dcbb50-ac55-1c6b-f4a3-66a2e17a2e9f"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-a6632210-2f96-54eb-7c24-6689d8c3ab90"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-e2324dd0-03ea-7053-4212-42facf1acdf7"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-20e27620-2f39-65f1-ff0c-215b32505ff5"
+    }
+  ],
+  "cudaVersion": "13.0",
+  "writerId": "r67343ly7cdn89a96rq36ht8y9oi83kq"
+}

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260629_024802-rgt7q8d2/logs/debug-internal.log

@@ -0,0 +1,15 @@
+{"time":"2026-06-29T02:48:02.852198932Z","level":"INFO","msg":"wandb-core"}
+{"time":"2026-06-29T02:48:02.855022661Z","level":"INFO","msg":"stream: starting","core version":"0.28.0"}
+{"time":"2026-06-29T02:48:03.311055668Z","level":"INFO","msg":"stream: created new stream","id":"rgt7q8d2"}
+{"time":"2026-06-29T02:48:03.311179032Z","level":"INFO","msg":"handler: started"}
+{"time":"2026-06-29T02:48:03.313936516Z","level":"INFO","msg":"stream: started"}
+{"time":"2026-06-29T02:48:03.313968038Z","level":"INFO","msg":"writer: started","stream_id":"rgt7q8d2"}
+{"time":"2026-06-29T02:48:03.314047859Z","level":"INFO","msg":"sender: started"}
+{"time":"2026-06-29T02:48:04.035828268Z","level":"INFO","msg":"filestream: sending request","total_files":1,"console_offset":399,"console_lines":1}
+{"time":"2026-06-29T02:48:04.552117748Z","level":"INFO","msg":"filestream: request sent","status":"200 OK"}
+{"time":"2026-06-29T02:48:19.037684411Z","level":"INFO","msg":"filestream: sending request","total_files":2,"events_offset":20,"events_lines":2,"console_offset":400,"console_lines":132,"uploaded_len":2}
+{"time":"2026-06-29T02:48:19.310946496Z","level":"INFO","msg":"filestream: request sent","status":"200 OK"}
+{"time":"2026-06-29T02:48:34.03620747Z","level":"INFO","msg":"filestream: sending request","total_files":1,"events_offset":22,"events_lines":2}
+{"time":"2026-06-29T02:48:34.209407654Z","level":"INFO","msg":"filestream: request sent","status":"200 OK"}
+{"time":"2026-06-29T02:48:49.03659551Z","level":"INFO","msg":"filestream: sending request","total_files":1,"events_offset":24,"events_lines":2}
+{"time":"2026-06-29T02:48:49.267042634Z","level":"INFO","msg":"filestream: request sent","status":"200 OK"}

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260629_024802-rgt7q8d2/logs/debug.log

@@ -0,0 +1,20 @@
+2026-06-29 02:48:02,621 INFO    MainThread:7550 [wandb_setup.py:_flush():81] Current SDK version is 0.28.0
+2026-06-29 02:48:02,621 INFO    MainThread:7550 [wandb_setup.py:_flush():81] Configure stats pid to 7550
+2026-06-29 02:48:02,621 INFO    MainThread:7550 [wandb_setup.py:_flush():81] Loading settings from environment variables
+2026-06-29 02:48:02,621 INFO    MainThread:7550 [wandb_init.py:setup_run_log_directory():725] Logging user logs to outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260629_024802-rgt7q8d2/logs/debug.log
+2026-06-29 02:48:02,621 INFO    MainThread:7550 [wandb_init.py:setup_run_log_directory():726] Logging internal logs to outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260629_024802-rgt7q8d2/logs/debug-internal.log
+2026-06-29 02:48:02,621 INFO    MainThread:7550 [wandb_init.py:init():768] calling init triggers
+2026-06-29 02:48:02,622 INFO    MainThread:7550 [wandb_init.py:init():773] wandb.init called with sweep_config: {}
+config: {'pretrained_model_root': './ckpts', 'pretrained_transformer_version': '480p_t2v', 'post_train_type': 'grpo', 'learning_rate': 1e-05, 'weight_decay': 0.01, 'max_steps': 10000, 'warmup_steps': 500, 'gradient_accumulation_steps': 1, 'max_grad_norm': 1.0, 'use_muon': True, 'num_train_timesteps': 1000, 'train_timestep_shift': 3.0, 'validation_timestep_shift': 5.0, 'snr_type': 'lognorm', 'task_type': 't2v', 'i2v_prob': 0.3, 'enable_fsdp': True, 'enable_gradient_checkpointing': True, 'sp_size': 2, 'dp_replicate': 1, 'batch_size': 2, 'num_workers': 4, 'prefetch_factor': 2, 'num_generations': 4, 'use_same_noise': False, 'output_dir': './outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3', 'checkpoints_directory': 's3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3', 'save_interval': 25, 'timer_interval': 0, 'auto_resume_from_timer': True, 'log_interval': 1, 'enable_wandb': True, 'project_name': 'SAGE-GRPO', 'run_name': 'yw006-4-sage-qwen35-9b-token-credit-video-adv-window3', 'wandb_key': None, 'wandb_username': None, 'wandb_base_url': 'https://api.wandb.ai', 'wandb_entity': None, 'wandb_silent': True, 'dtype': 'bf16', 'master_weight_type': 'fp32', 'seed': 42, 'validation_global_seed': 930, 'validation_interval': 0, 'validation_prompts': None, 'validate_video_length': 121, 'validation_aspect_ratio': '16:9', 'validation_fixed_size': '480x864', 'validation_num_inference_steps': 40, 'validation_negative_prompt': '', 'validation_enable_sr': False, 'validation_prompt_rewrite': False, 'validate_at_step0': False, 'resume_from_checkpoint': None, 'use_lora': False, 'lora_r': 8, 'lora_alpha': 16, 'lora_dropout': 0.0, 'lora_target_modules': None, 'pretrained_lora_path': None, 'train_video_csv': ['/nfs/ywang29/Reward_finetuning/SAGE-GRPO/assets/demo_train_decomposed_binary5.csv'], 'valid_video_csv': 'assets/demo_sample.csv', 'reward_model': 'snapvideo_vlm', 'reward_config': None, 'reward_checkpoint_mode': 'v3', 'reward_weights': '{"VQ":0.5,"MQ":0.5,"TA":1.0}', 'reward_gradient_credit': False, 'reward_credit_num_frames': 0, 'reward_credit_max_questions': 0, 'reward_credit_metric': 'weighted', 'token_credit_source_model': None, 'token_credit_advantage_source': 'token', 'token_credit_remote_url': None, 'token_credit_vlm_model_path': None, 'token_credit_vlm_model_family': None, 'token_credit_vlm_prompt_template': None, 'remote_reward_url': None, 'vlm_reward_model_path': '/nfs/ywang29/ckpts/Qwen3.5-9B', 'vlm_reward_model_family': 'qwen3_5', 'vlm_reward_prompt_template': None, 'vlm_reward_num_frames': 20, 'vlm_reward_max_pixels': 50176, 'vlm_reward_max_new_tokens': 128, 'vlm_reward_batch_size': 1, 'vlm_reward_score_type': 'token_credit', 'vlm_reward_missing_score': 0.0, 'vlm_reward_video_sample_type': 'uniform', 'vlm_reward_token_credit_max_questions': 5, 'token_credit_loss_mode': 'video_adv', 'token_credit_ratio_mode': 'weighted_logprob', 'token_credit_shuffle_weights': False, 'token_credit_spatial_granularity': 'window', 'token_credit_spatial_bins': 16, 'token_credit_spatial_window': 3, 'token_credit_chunk_size': 1, 'enable_grpo_policy_update': True, 'grpo_sampling_steps': 20, 'sample_n_frames': 81, 'eta': 0.5, 'cfg_scale': 6.0, 'neg_prompt': None, 'data_type': 'video', 'model_type': '16164', 'video_fps': 24, 'vae_name': '16164-32c-hy20250605', 'vae_tiling': False, 'infer_flow_shift_video': 5.0, 'video_bucket_hw_base_size': 480, 'video_bucket_hw_bucket_stride': 32, 'multitask_mask_training_type': 'concat', 'vision_num_semantic_tokens': 729, 'vision_states_dim': 1152, 'sde_type': 'sage_grpo', 'mini_batch_size_per_update': 1, 'clip_range': 0.0001, 'adv_clip_max': 5.0, 'timestep_fraction': 1.0, 'random_train_timestep_count': False, 'random_train_timestep_min': 1, 'random_train_timestep_max': 20, 'use_grad_balancing': True, 'enable_timestep_permutation': True, 'reference_mode_offload': True, 'kl_weight': 1e-05, 'kl_coef': 1e-07, 'kl_min_coef': 1e-07, 'use_moving_KL': True, 'update_ref_model_step': 10, 'use_dual_kl': True, 'dual_kl_moving_weight': 1.0, 'dual_kl_step_weight': 0.1, 'kl_compute_mode': 'rollout_phase', 'glyph_byT5_v2': True, 'byt5_max_length': 256, 'video_uncond_p_byt5': 0.0, 'precision': 'bf16', 'vae_precision': 'fp16', 'val_disable_autocast': False, 'debug_grad_flow': False, 'debug_train_diagnostics_interval': 1, '_wandb': {}}
+2026-06-29 02:48:02,622 INFO    MainThread:7550 [wandb_init.py:init():816] starting backend
+2026-06-29 02:48:02,842 INFO    MainThread:7550 [wandb_init.py:init():831] sending inform_init request
+2026-06-29 02:48:03,314 INFO    MainThread:7550 [wandb_init.py:init():836] backend started and connected
+2026-06-29 02:48:03,317 INFO    MainThread:7550 [wandb_init.py:init():906] updated telemetry
+2026-06-29 02:48:03,348 INFO    MainThread:7550 [wandb_init.py:init():929] communicating run to backend with 90.0 second timeout
+2026-06-29 02:48:03,833 INFO    MainThread:7550 [wandb_init.py:init():969] run resumed
+2026-06-29 02:48:03,834 INFO    MainThread:7550 [wandb_init.py:init():974] starting run threads in backend
+2026-06-29 02:48:04,030 INFO    MainThread:7550 [wandb_run.py:_console_start():2523] atexit reg
+2026-06-29 02:48:04,030 INFO    MainThread:7550 [wandb_run.py:_redirect():2373] redirect: wrap_raw
+2026-06-29 02:48:04,031 INFO    MainThread:7550 [wandb_run.py:_redirect():2442] Wrapping output streams.
+2026-06-29 02:48:04,031 INFO    MainThread:7550 [wandb_run.py:_redirect():2465] Redirects installed.
+2026-06-29 02:48:04,032 INFO    MainThread:7550 [wandb_init.py:init():1012] run started, returning control to user process

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260629_030914-rgt7q8d2/files/output.log

@@ -0,0 +1,133 @@
+2026-06-29 03:09:15.389 | INFO     | __main__:print_training_configuration:970 - SYSTEM & HARDWARE CONFIGURATION
+2026-06-29 03:09:15.389 | INFO     | __main__:print_training_configuration:971 - --------------------------------------------------------------------------------
+2026-06-29 03:09:15.390 | INFO     | __main__:print_training_configuration:972 - Number of GPUs                                   : 64
+2026-06-29 03:09:15.390 | INFO     | __main__:print_training_configuration:973 - Local rank                                       : 0
+2026-06-29 03:09:15.391 | INFO     | __main__:print_training_configuration:974 - Global rank                                      : 0
+2026-06-29 03:09:15.391 | INFO     | __main__:print_training_configuration:975 - Device                                           : cuda:0
+2026-06-29 03:09:15.392 | INFO     | __main__:print_training_configuration:976 - Master weight dtype                              : torch.float32
+2026-06-29 03:09:15.392 | INFO     | __main__:print_training_configuration:978 - --------------------------------------------------------------------------------
+2026-06-29 03:09:15.393 | INFO     | __main__:print_training_configuration:979 - MODEL PARAMETERS & ARCHITECTURE
+2026-06-29 03:09:15.393 | INFO     | __main__:print_training_configuration:980 - --------------------------------------------------------------------------------
+2026-06-29 03:09:15.404 | INFO     | __main__:print_training_configuration:983 - Number of total                                  : 8,326,608,160
+2026-06-29 03:09:15.405 | INFO     | __main__:print_training_configuration:984 - Total trainable parameters                        : 8,326,608,160
+2026-06-29 03:09:15.405 | INFO     | __main__:print_training_configuration:985 - DIT model root                                   : ./ckpts
+2026-06-29 03:09:15.406 | INFO     | __main__:print_training_configuration:986 - Transformer version                              : 480p_t2v
+2026-06-29 03:09:15.406 | INFO     | __main__:print_training_configuration:987 - Attention mode                                   : flash
+2026-06-29 03:09:15.407 | INFO     | __main__:print_training_configuration:988 - Gradient checkpointing                           : True
+2026-06-29 03:09:15.407 | INFO     | __main__:print_training_configuration:989 - FSDP enabled                                     : True
+2026-06-29 03:09:15.407 | INFO     | __main__:print_training_configuration:990 - LoRA enabled                                     : False
+2026-06-29 03:09:15.408 | INFO     | __main__:print_training_configuration:991 - Glyph byT5 v2                                    : True
+2026-06-29 03:09:15.408 | INFO     | __main__:print_training_configuration:992 - ByT5 max length                                  : 256
+2026-06-29 03:09:15.409 | INFO     | __main__:print_training_configuration:994 - --------------------------------------------------------------------------------
+2026-06-29 03:09:15.409 | INFO     | __main__:print_training_configuration:995 - DISTRIBUTED TRAINING CONFIGURATION
+2026-06-29 03:09:15.410 | INFO     | __main__:print_training_configuration:996 - --------------------------------------------------------------------------------
+2026-06-29 03:09:15.410 | INFO     | __main__:print_training_configuration:997 - World size                                       : 64
+2026-06-29 03:09:15.410 | INFO     | __main__:print_training_configuration:998 - DP degree                                        : 32
+2026-06-29 03:09:15.411 | INFO     | __main__:print_training_configuration:999 - DP rank                                          : 0
+2026-06-29 03:09:15.411 | INFO     | __main__:print_training_configuration:1000 - SP size                                          : 2
+2026-06-29 03:09:15.412 | INFO     | __main__:print_training_configuration:1001 - SP rank                                          : 0
+2026-06-29 03:09:15.412 | INFO     | __main__:print_training_configuration:1003 - --------------------------------------------------------------------------------
+2026-06-29 03:09:15.412 | INFO     | __main__:print_training_configuration:1004 - BATCH SIZE & DATA FLOW CONFIGURATION
+2026-06-29 03:09:15.413 | INFO     | __main__:print_training_configuration:1005 - --------------------------------------------------------------------------------
+2026-06-29 03:09:15.413 | INFO     | __main__:print_training_configuration:1006 - Per GPU batch size                               : 2
+2026-06-29 03:09:15.413 | INFO     | __main__:print_training_configuration:1007 - DP degree batch size                             : 64
+2026-06-29 03:09:15.414 | INFO     | __main__:print_training_configuration:1008 - Gradient accumulation steps                      : 1
+2026-06-29 03:09:15.414 | INFO     | __main__:print_training_configuration:1009 - Num generations                                  : 4
+2026-06-29 03:09:15.415 | INFO     | __main__:print_training_configuration:1010 - Use same noise                                   : False
+2026-06-29 03:09:15.415 | INFO     | __main__:print_training_configuration:1011 - Num workers                                      : 4
+2026-06-29 03:09:15.415 | INFO     | __main__:print_training_configuration:1012 - Prefetch factor                                  : 2
+2026-06-29 03:09:15.416 | INFO     | __main__:print_training_configuration:1014 - --------------------------------------------------------------------------------
+2026-06-29 03:09:15.416 | INFO     | __main__:print_training_configuration:1015 - OPTIMIZATION CONFIGURATION
+2026-06-29 03:09:15.417 | INFO     | __main__:print_training_configuration:1016 - --------------------------------------------------------------------------------
+2026-06-29 03:09:15.417 | INFO     | __main__:print_training_configuration:1017 - Optimizer                                        : muon
+2026-06-29 03:09:15.418 | INFO     | __main__:print_training_configuration:1018 - Learning rate                                    : 1e-05
+2026-06-29 03:09:15.418 | INFO     | __main__:print_training_configuration:1019 - Weight decay                                     : 0.01
+2026-06-29 03:09:15.418 | INFO     | __main__:print_training_configuration:1020 - Max gradient norm                                : 1.0
+2026-06-29 03:09:15.419 | INFO     | __main__:print_training_configuration:1021 - LR scheduler                                     : constant
+2026-06-29 03:09:15.419 | INFO     | __main__:print_training_configuration:1022 - Warmup steps                                     : 500
+2026-06-29 03:09:15.419 | INFO     | __main__:print_training_configuration:1023 - Max training steps                               : 10000
+2026-06-29 03:09:15.420 | INFO     | __main__:print_training_configuration:1025 - --------------------------------------------------------------------------------
+2026-06-29 03:09:15.420 | INFO     | __main__:print_training_configuration:1026 - INFERENCE & VALIDATION CONFIGURATION
+2026-06-29 03:09:15.421 | INFO     | __main__:print_training_configuration:1027 - --------------------------------------------------------------------------------
+2026-06-29 03:09:15.421 | INFO     | __main__:print_training_configuration:1028 - Validation interval                              : 0
+2026-06-29 03:09:15.421 | INFO     | __main__:print_training_configuration:1029 - Validate at step 0                               : False
+2026-06-29 03:09:15.422 | INFO     | __main__:print_training_configuration:1030 - Validation video length                          : 121
+2026-06-29 03:09:15.422 | INFO     | __main__:print_training_configuration:1031 - Validation global seed                           : 930
+2026-06-29 03:09:15.423 | INFO     | __main__:print_training_configuration:1032 - Validation inference steps                        : 40
+2026-06-29 03:09:15.423 | INFO     | __main__:print_training_configuration:1033 - Validation aspect ratio                          : 16:9
+2026-06-29 03:09:15.423 | INFO     | __main__:print_training_configuration:1034 - Validation fixed size                            : 480x864
+2026-06-29 03:09:15.424 | INFO     | __main__:print_training_configuration:1035 - Validation prompt rewrite                         : False
+2026-06-29 03:09:15.424 | INFO     | __main__:print_training_configuration:1036 - Validation flow shift                             : 5.0
+2026-06-29 03:09:15.425 | INFO     | __main__:print_training_configuration:1037 - Sampling CFG scale                                : 6.0
+2026-06-29 03:09:15.425 | INFO     | __main__:print_training_configuration:1038 - Sampling eta                                      : 0.5
+2026-06-29 03:09:15.425 | INFO     | __main__:print_training_configuration:1039 - Sampling steps (GRPO rollout)                     : 20
+2026-06-29 03:09:15.426 | INFO     | __main__:print_training_configuration:1040 - Token credit loss mode                            : video_adv
+2026-06-29 03:09:15.426 | INFO     | __main__:print_training_configuration:1041 - Token credit ratio mode                           : weighted_logprob
+2026-06-29 03:09:15.427 | INFO     | __main__:print_training_configuration:1042 - Token credit shuffle weights                      : False
+2026-06-29 03:09:15.427 | INFO     | __main__:print_training_configuration:1043 - Token credit spatial granularity                  : window
+2026-06-29 03:09:15.427 | INFO     | __main__:print_training_configuration:1044 - Token credit spatial bins                         : 16
+2026-06-29 03:09:15.428 | INFO     | __main__:print_training_configuration:1045 - Token credit spatial window                       : 3
+2026-06-29 03:09:15.428 | INFO     | __main__:print_training_configuration:1046 - Token credit chunk size                           : 1
+2026-06-29 03:09:15.429 | INFO     | __main__:print_training_configuration:1047 - Token credit source model                         : None
+2026-06-29 03:09:15.429 | INFO     | __main__:print_training_configuration:1048 - Token credit advantage source                     : token
+2026-06-29 03:09:15.429 | INFO     | __main__:print_training_configuration:1049 - Token credit VLM model path                       : None
+2026-06-29 03:09:15.430 | INFO     | __main__:print_training_configuration:1050 - Token credit VLM model family                     : None
+2026-06-29 03:09:15.430 | INFO     | __main__:print_training_configuration:1051 - Random train timestep count                       : False
+2026-06-29 03:09:15.430 | INFO     | __main__:print_training_configuration:1052 - Random train timestep range                       : 1-20
+2026-06-29 03:09:15.431 | INFO     | __main__:print_training_configuration:1056 - SDE type                                          : sage_grpo
+2026-06-29 03:09:15.431 | INFO     | __main__:print_training_configuration:1058 - --------------------------------------------------------------------------------
+2026-06-29 03:09:15.432 | INFO     | __main__:print_training_configuration:1059 - GRPO & KL CONFIGURATION
+2026-06-29 03:09:15.432 | INFO     | __main__:print_training_configuration:1060 - --------------------------------------------------------------------------------
+2026-06-29 03:09:15.432 | INFO     | __main__:print_training_configuration:1061 - Reward model                                     : snapvideo_vlm
+2026-06-29 03:09:15.433 | INFO     | __main__:print_training_configuration:1062 - Reward gradient credit                          : False
+2026-06-29 03:09:15.433 | INFO     | __main__:print_training_configuration:1063 - Reward credit frames                            : 0
+2026-06-29 03:09:15.434 | INFO     | __main__:print_training_configuration:1064 - Reward credit max questions                     : 0
+2026-06-29 03:09:15.434 | INFO     | __main__:print_training_configuration:1065 - Reward credit metric                            : weighted
+2026-06-29 03:09:15.434 | INFO     | __main__:print_training_configuration:1067 - VLM reward model path                            : /nfs/ywang29/ckpts/Qwen3.5-9B
+2026-06-29 03:09:15.435 | INFO     | __main__:print_training_configuration:1068 - VLM reward model family                          : qwen3_5
+2026-06-29 03:09:15.435 | INFO     | __main__:print_training_configuration:1069 - VLM reward frames                                : 20
+2026-06-29 03:09:15.435 | INFO     | __main__:print_training_configuration:1070 - VLM reward max pixels                            : 50176
+2026-06-29 03:09:15.436 | INFO     | __main__:print_training_configuration:1071 - VLM reward score type                            : token_credit
+2026-06-29 03:09:15.436 | INFO     | __main__:print_training_configuration:1072 - VLM reward max new tokens                        : 128
+2026-06-29 03:09:15.437 | INFO     | __main__:print_training_configuration:1073 - VLM reward batch size                            : 1
+2026-06-29 03:09:15.437 | INFO     | __main__:print_training_configuration:1074 - VLM reward missing score                         : 0.0
+2026-06-29 03:09:15.438 | INFO     | __main__:print_training_configuration:1077 - KL weight                                        : 1e-05
+2026-06-29 03:09:15.438 | INFO     | __main__:print_training_configuration:1078 - KL coef                                          : 1e-07
+2026-06-29 03:09:15.438 | INFO     | __main__:print_training_configuration:1079 - Use moving KL                                    : True
+2026-06-29 03:09:15.439 | INFO     | __main__:print_training_configuration:1080 - Update ref model step                            : 10
+2026-06-29 03:09:15.439 | INFO     | __main__:print_training_configuration:1081 - Use dual KL                                      : True
+2026-06-29 03:09:15.440 | INFO     | __main__:print_training_configuration:1082 - Dual KL moving weight                            : 1.0
+2026-06-29 03:09:15.440 | INFO     | __main__:print_training_configuration:1083 - Dual KL step weight                              : 0.1
+2026-06-29 03:09:15.440 | INFO     | __main__:print_training_configuration:1084 - KL compute mode                                  : rollout_phase
+2026-06-29 03:09:15.441 | INFO     | __main__:print_training_configuration:1085 - Use grad balancing                               : True
+2026-06-29 03:09:15.441 | INFO     | __main__:print_training_configuration:1086 - Enable timestep permutation                      : True
+2026-06-29 03:09:15.441 | INFO     | __main__:print_training_configuration:1087 - Reference mode offload                           : True
+2026-06-29 03:09:15.442 | INFO     | __main__:print_training_configuration:1089 - --------------------------------------------------------------------------------
+2026-06-29 03:09:15.442 | INFO     | __main__:print_training_configuration:1090 - LOGGING & CHECKPOINTING CONFIGURATION
+2026-06-29 03:09:15.443 | INFO     | __main__:print_training_configuration:1091 - --------------------------------------------------------------------------------
+2026-06-29 03:09:15.443 | INFO     | __main__:print_training_configuration:1092 - Output directory                                 : ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3
+2026-06-29 03:09:15.443 | INFO     | __main__:print_training_configuration:1093 - Checkpoints directory                            : s3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3
+2026-06-29 03:09:15.444 | INFO     | __main__:print_training_configuration:1094 - Checkpoint interval                              : 25
+2026-06-29 03:09:15.444 | INFO     | __main__:print_training_configuration:1095 - Timer checkpoint interval                        : 0
+2026-06-29 03:09:15.445 | INFO     | __main__:print_training_configuration:1096 - Auto resume from timer/checkpoint                : True
+2026-06-29 03:09:15.445 | INFO     | __main__:print_training_configuration:1097 - Timer checkpoint directory                       : s3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/timer
+2026-06-29 03:09:15.445 | INFO     | __main__:print_training_configuration:1098 - Log interval                                     : 1
+2026-06-29 03:09:15.446 | INFO     | __main__:print_training_configuration:1099 - Resume checkpoint                                : None
+2026-06-29 03:09:15.446 | INFO     | __main__:print_training_configuration:1100 - Train CSV                                         : ['/nfs/ywang29/Reward_finetuning/SAGE-GRPO/assets/demo_train_decomposed_binary5.csv']
+2026-06-29 03:09:15.447 | INFO     | __main__:print_training_configuration:1101 - Valid CSV                                         : assets/demo_sample.csv
+2026-06-29 03:09:15.447 | INFO     | __main__:print_training_configuration:1102 - Seed                                             : 42
+2026-06-29 03:09:15.448 | INFO     | __main__:print_training_configuration:1103 - WandB enabled                                    : True
+2026-06-29 03:09:15.448 | INFO     | __main__:print_training_configuration:1105 - WandB project                                    : SAGE-GRPO
+2026-06-29 03:09:15.449 | INFO     | __main__:print_training_configuration:1106 - WandB run name                                   : yw006-4-sage-qwen35-9b-token-credit-video-adv-window3
+2026-06-29 03:09:15.449 | INFO     | __main__:__init__:958 - [Effective Config] world_size=64, rank=0, local_rank=0, sp_size=2, dp_size=32, dp_rank=0, sp_rank=0, batch_size=2, num_generations=4, grad_accum=1
+2026-06-29 03:09:15.449 | INFO     | __main__:_build_video_dataset:2088 - [Effective GRPO Grouping] video_batch_size=2, dp_replicas=32, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, product=64
+2026-06-29 03:09:16.099 | INFO     | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1
+2026-06-29 03:09:16.100 | INFO     | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done
+2026-06-29 03:09:16.100 | INFO     | __main__:_train_grpo:2171 - Starting GRPO training...
+2026-06-29 03:09:16.101 | INFO     | __main__:_train_grpo:2172 - Max steps: 10000
+2026-06-29 03:09:16.101 | INFO     | __main__:_train_grpo:2173 - Batch size: 2
+2026-06-29 03:09:16.102 | INFO     | __main__:_train_grpo:2174 - Learning rate: 1e-05
+2026-06-29 03:09:16.952 | INFO     | __main__:_resolve_resume_checkpoint:1860 - No timer checkpoint found at s3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/timer; auto-resuming from latest regular checkpoint: s3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/checkpoint-150
+2026-06-29 03:09:17.081 | INFO     | __main__:load_checkpoint:1992 - Loading checkpoint from s3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/checkpoint-150
+/usr/local/lib/python3.12/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)

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260629_030914-rgt7q8d2/files/requirements.txt

@@ -0,0 +1,287 @@
+awscli==1.45.30
+boto3==1.43.30
+botocore==1.43.30
+certifi==2026.5.20
+charset-normalizer==3.4.7
+colorama==0.4.6
+docutils==0.19
+idna==3.18
+jmespath==1.1.0
+pip==26.1.2
+pyasn1==0.6.3
+python-dateutil==2.9.0.post0
+PyYAML==6.0.3
+requests==2.34.2
+rsa==4.7.2
+s3transfer==0.18.0
+six==1.17.0
+urllib3==2.7.0
+py-cpuinfo==9.0.0
+hjson==3.1.0
+easydict==1.13
+braceexpand==0.1.7
+antlr4-python3-runtime==4.9.3
+websocket-client==1.9.0
+webdataset==1.0.2
+triton==3.6.0
+torchcodec==0.14.0+cu130
+tifffile==2026.6.1
+tenacity==9.1.4
+smmap==5.0.3
+sentry-sdk==2.63.0
+sentencepiece==0.2.1
+regex==2026.6.28
+pyarrow==24.0.0
+pyaml==26.2.1
+propcache==0.5.2
+opencv-python-headless==4.13.0.92
+omegaconf==2.3.1
+nvidia-nvtx==13.0.85
+ImageIO==2.37.3
+nvidia-nvshmem-cu13==3.4.5
+nvidia-nvjitlink==13.0.88
+av==17.1.0
+nvidia-nccl-cu13==2.28.9
+nvidia-cufile==1.15.1.6
+imageio-ffmpeg==0.6.0
+nvidia-cuda-runtime==13.0.96
+nvidia-cuda-nvrtc==13.0.88
+attrs==26.1.0
+nvidia-cuda-cupti==13.0.85
+nvidia-cublas==13.1.0.3
+multidict==6.7.1
+msgpack==1.2.1
+Mako==1.3.12
+loguru==0.7.3
+lmdb==2.2.1
+lightning-utilities==0.15.3
+lazy-loader==0.5
+greenlet==3.5.3
+future==1.0.0
+ftfy==6.3.1
+frozenlist==1.8.0
+frozendict==2.4.7
+docstring_parser==0.18.0
+decord==0.6.0
+cuda-pathfinder==1.5.5
+aiohappyeyeballs==2.6.2
+yarl==1.24.2
+typed-argument-parser==1.12.0
+SQLAlchemy==2.0.51
+scikit-image==0.26.0
+qwen-vl-utils==0.0.14
+typeguard==4.5.2
+nvidia-cusparse==12.6.3.3
+nvidia-cufft==12.0.0.61
+nvidia-cudnn-cu13==9.15.1.9
+gitdb==4.0.12
+ffmpeg-python==0.2.0
+einx==0.4.3
+cupy-cuda12x==14.1.1
+cuda-bindings==13.0.3
+cattrs==26.1.0
+aiosignal==1.4.0
+tokenizers==0.22.2
+termcolor==3.3.0
+nvidia-cusolver==12.0.4.66
+mediapy==1.2.6
+ipdb==0.13.13
+GitPython==3.1.50
+diffusers==0.38.0
+coremltools==9.0
+alembic==1.18.5
+aiohttp==3.14.1
+wandb==0.28.0
+torch==2.10.0+cu130
+dataset==2.0.0
+dashscope==1.26.0
+torchvision==0.25.0+cu130
+torchmetrics==1.9.0
+sniffio==1.3.1
+torchdata==0.8.0
+tqdm==4.67.3
+torchaudio==2.10.0+cu130
+deepspeed==0.19.2
+timm==1.0.27
+pytorch-lightning==2.6.5
+peft==0.19.1
+lpips==0.1.4
+lightning==2.6.5
+flash_attn==2.8.3+cu130torch2.10
+xxhash==3.8.0
+shortuuid==1.0.13
+setuptools==82.0.0
+safetensors==0.7.0
+pillow==12.1.1
+numpy==2.4.2
+jiter==0.15.0
+fsspec==2026.2.0
+distro==1.9.0
+dill==0.4.1
+tyro==1.0.15
+tiktoken==0.13.0
+multiprocess==0.70.19
+modelscope==1.36.0
+fire==0.7.0
+openai==2.8.0
+datasets==4.8.4
+trl==0.8.6
+angelslim==0.2.2
+huggingface_hub==1.21.0
+transformers==5.12.1
+black==26.5.1
+pathspec==1.1.1
+pytokens==0.4.1
+importlib_metadata==9.0.0
+ml_dtypes==0.5.4
+onnx==1.22.0
+onnx-ir==0.2.1
+onnxscript==0.7.0
+transformer_engine==2.9.0+70f53666
+zipp==4.1.0
+nvidia-mathdx==25.6.0
+einops==0.8.2
+absl-py==2.4.0
+accelerate==1.14.0
+annotated-doc==0.0.4
+annotated-types==0.7.0
+anyio==4.14.0
+apsw==3.53.2.0
+apswutils==0.1.2
+beartype==0.22.9
+beautifulsoup4==4.15.0
+blis==1.2.1
+catalogue==2.0.10
+cloudpathlib==0.24.0
+cloudpickle==3.1.2
+confection==0.1.5
+contourpy==1.3.3
+cycler==0.12.1
+cymem==2.0.13
+fastai==2.8.7
+fastcore==1.13.6
+fastdownload==0.0.7
+fastlite==0.2.4
+fastprogress==1.1.6
+fasttransform==0.0.2
+fonttools==4.63.0
+grpcio==1.81.1
+h11==0.16.0
+hf-xet==1.5.1
+httpcore==1.0.9
+httpcore2==2.4.0
+httptools==0.8.0
+httpx==0.28.1
+httpx2==2.4.0
+itsdangerous==2.2.0
+joblib==1.5.3
+kiwisolver==1.5.0
+langcodes==3.5.1
+Markdown==3.10.2
+markdown-it-py==4.2.0
+matplotlib==3.11.0
+mdurl==0.1.2
+mpmath==1.3.0
+murmurhash==1.0.15
+mypy_extensions==1.1.0
+narwhals==2.22.1
+networkx==3.6.1
+nvidia-curand==10.4.0.35
+nvidia-cusparselt-cu13==0.8.0
+oauthlib==3.3.1
+pandas==3.0.3
+plum-dispatch==2.9.0
+preshed==3.0.13
+protobuf==7.35.1
+pydantic==2.13.4
+pydantic_core==2.46.4
+pyparsing==3.3.2
+pyre-extensions==0.0.32
+python-dotenv==1.2.2
+python-fasthtml==0.14.3
+python-multipart==0.0.32
+rich==15.0.0
+s3torchconnector==1.5.0
+s3torchconnectorclient==1.5.0
+scikit-learn==1.9.0
+shellingham==1.5.4
+smart_open==7.6.1
+soupsieve==2.8.4
+spacy==3.8.7
+spacy-legacy==3.0.12
+spacy-loggers==1.0.5
+srsly==2.5.3
+starlette==1.3.1
+sympy==1.14.0
+tabulate==0.10.0
+tensorboard==2.20.0
+tensorboard-data-server==0.7.2
+thinc==8.3.4
+threadpoolctl==3.6.0
+torchtnt==0.2.4
+truststore==0.10.4
+typer==0.25.1
+typer-slim==0.24.0
+typing-inspect==0.9.0
+typing-inspection==0.4.2
+uvicorn==0.49.0
+uvloop==0.22.1
+wasabi==1.1.3
+watchfiles==1.2.0
+weasel==0.4.3
+websockets==16.0
+Werkzeug==3.1.8
+wrapt==2.2.2
+asttokens==3.0.1
+cffi==2.0.0
+click==8.4.1
+comm==0.2.3
+cryptography==49.0.0
+Cython==3.2.5
+debugpy==1.8.21
+decorator==5.3.1
+executing==2.2.1
+filelock==3.29.4
+h5py==3.16.0
+intel-cmplr-lib-ur==2026.0.0
+intel-openmp==2026.0.0
+ipykernel==7.3.0
+ipython==9.14.1
+ipython_pygments_lexers==1.1.1
+jedi==0.20.0
+Jinja2==3.1.6
+jupyter_client==8.9.1
+jupyter_core==5.9.1
+MarkupSafe==3.0.3
+matplotlib-inline==0.2.2
+mkl==2026.0.0
+mkl-include==2026.0.0
+mpi4py==4.1.2
+nest-asyncio2==1.7.2
+ninja==1.13.0
+onemkl-license==2026.0.0
+opencv-python==4.11.0.86
+packaging==26.2
+parso==0.8.7
+pexpect==4.9.0
+platformdirs==4.10.0
+prompt_toolkit==3.0.52
+psutil==7.2.2
+ptyprocess==0.7.0
+pure_eval==0.2.3
+pybind11==3.0.4
+pycparser==3.0
+Pygments==2.20.0
+pyOpenSSL==26.3.0
+pyzmq==27.1.0
+scipy==1.18.0
+stack-data==0.6.3
+tbb==2023.0.0
+tcmlib==1.5.0
+tornado==6.5.7
+traitlets==5.15.1
+typing_extensions==4.15.0
+tzdata==2026.2
+umf==1.1.0
+wheel==0.47.0
+wcwidth==0.8.1

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260629_030914-rgt7q8d2/files/wandb-metadata.json

@@ -0,0 +1,185 @@
+{
+  "os": "Linux-6.1.163-186.299.amzn2023.x86_64-x86_64-with-glibc2.35",
+  "python": "CPython 3.12.10",
+  "startedAt": "2026-06-29T03:09:14.100599Z",
+  "args": [
+    "--pretrained_model_root",
+    "./ckpts",
+    "--train_video_csv",
+    "/nfs/ywang29/Reward_finetuning/SAGE-GRPO/assets/demo_train_decomposed_binary5.csv",
+    "--reward_model",
+    "snapvideo_vlm",
+    "--vlm_reward_model_path",
+    "/nfs/ywang29/ckpts/Qwen3.5-9B",
+    "--vlm_reward_model_family",
+    "qwen3_5",
+    "--vlm_reward_num_frames",
+    "20",
+    "--vlm_reward_max_pixels",
+    "50176",
+    "--vlm_reward_max_new_tokens",
+    "128",
+    "--vlm_reward_batch_size",
+    "1",
+    "--vlm_reward_score_type",
+    "token_credit",
+    "--vlm_reward_token_credit_max_questions",
+    "5",
+    "--token_credit_loss_mode",
+    "video_adv",
+    "--token_credit_ratio_mode",
+    "weighted_logprob",
+    "--token_credit_spatial_granularity",
+    "window",
+    "--token_credit_spatial_bins",
+    "16",
+    "--token_credit_spatial_window",
+    "3",
+    "--token_credit_chunk_size",
+    "1",
+    "--learning_rate",
+    "1e-5",
+    "--batch_size",
+    "2",
+    "--num_generations",
+    "4",
+    "--max_steps",
+    "10000",
+    "--output_dir",
+    "./outputs",
+    "--enable_wandb",
+    "True",
+    "--project_name",
+    "SAGE-GRPO",
+    "--run_name",
+    "yw006-4-sage-qwen35-9b-token-credit-video-adv-window3",
+    "--enable_fsdp",
+    "--enable_gradient_checkpointing",
+    "--sp_size",
+    "2",
+    "--reward_checkpoint_mode",
+    "v3",
+    "--validation_interval",
+    "0",
+    "--validate_at_step0",
+    "False",
+    "--validate_video_length",
+    "121",
+    "--validation_timestep_shift",
+    "5.0",
+    "--use_grad_balancing",
+    "True",
+    "--enable_timestep_permutation",
+    "True",
+    "--sde_type",
+    "sage_grpo",
+    "--kl_weight",
+    "1e-5",
+    "--kl_coef",
+    "1e-7",
+    "--use_moving_KL",
+    "True",
+    "--update_ref_model_step",
+    "10",
+    "--use_dual_kl",
+    "True",
+    "--dual_kl_moving_weight",
+    "1.0",
+    "--dual_kl_step_weight",
+    "0.1",
+    "--reference_mode_offload",
+    "True",
+    "--save_interval",
+    "25",
+    "--timer_interval",
+    "0",
+    "--auto_resume_from_timer",
+    "True",
+    "--checkpoints_directory",
+    "s3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3"
+  ],
+  "program": "/nfs/ywang29/Reward_finetuning/SAGE-GRPO/post_train.py",
+  "codePath": "post_train.py",
+  "codePathLocal": "post_train.py",
+  "git": {
+    "remote": "https://github.com/Tencent-Hunyuan/SAGE-GRPO.git",
+    "commit": "779776e42732e0262b3f2844dd206d9d8f043b2f"
+  },
+  "root": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb",
+  "host": "ywang29-hunyuan-sage-qwen35-tcva-w3-worker-0",
+  "username": "yi-fan-wang1216",
+  "executable": "/usr/local/bin/python",
+  "cpu_count": 48,
+  "cpu_count_logical": 96,
+  "gpu": "NVIDIA A100-SXM4-80GB",
+  "gpu_count": 8,
+  "disk": {
+    "/": {
+      "total": "7516777873408",
+      "used": "619664207872"
+    }
+  },
+  "memory": {
+    "total": "1204539252736"
+  },
+  "gpu_nvidia": [
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-f6eafbbf-b72f-1725-61a0-5da7fda21fe2"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-8faea6eb-ec0a-0561-c1a7-608821eb24f5"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-6d3845d3-5be7-3130-93f6-aa8bc1fc8f9b"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-a80bc0c5-98ed-a833-ec83-51e290f6ac9c"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-a8dcbb50-ac55-1c6b-f4a3-66a2e17a2e9f"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-a6632210-2f96-54eb-7c24-6689d8c3ab90"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-e2324dd0-03ea-7053-4212-42facf1acdf7"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-20e27620-2f39-65f1-ff0c-215b32505ff5"
+    }
+  ],
+  "cudaVersion": "13.0",
+  "writerId": "16wzahfmyr6npkufv9suqg37clvgtzox"
+}

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260629_030914-rgt7q8d2/logs/debug-internal.log

@@ -0,0 +1,27 @@
+{"time":"2026-06-29T03:09:14.353140666Z","level":"INFO","msg":"wandb-core"}
+{"time":"2026-06-29T03:09:14.354287437Z","level":"INFO","msg":"stream: starting","core version":"0.28.0"}
+{"time":"2026-06-29T03:09:14.632461218Z","level":"INFO","msg":"stream: created new stream","id":"rgt7q8d2"}
+{"time":"2026-06-29T03:09:14.632696173Z","level":"INFO","msg":"handler: started"}
+{"time":"2026-06-29T03:09:14.634847523Z","level":"INFO","msg":"stream: started"}
+{"time":"2026-06-29T03:09:14.634881842Z","level":"INFO","msg":"sender: started"}
+{"time":"2026-06-29T03:09:14.634870066Z","level":"INFO","msg":"writer: started","stream_id":"rgt7q8d2"}
+{"time":"2026-06-29T03:09:15.39109758Z","level":"INFO","msg":"filestream: sending request","total_files":1,"console_offset":532,"console_lines":2}
+{"time":"2026-06-29T03:09:15.823711279Z","level":"INFO","msg":"filestream: request sent","status":"200 OK"}
+{"time":"2026-06-29T03:09:30.391473379Z","level":"INFO","msg":"filestream: sending request","total_files":2,"events_offset":26,"events_lines":2,"console_offset":533,"console_lines":132,"uploaded_len":2}
+{"time":"2026-06-29T03:09:30.619287843Z","level":"INFO","msg":"filestream: request sent","status":"200 OK"}
+{"time":"2026-06-29T03:09:45.391834622Z","level":"INFO","msg":"filestream: sending request","total_files":1,"events_offset":28,"events_lines":2}
+{"time":"2026-06-29T03:09:45.647058741Z","level":"INFO","msg":"filestream: request sent","status":"200 OK"}
+{"time":"2026-06-29T03:10:00.391587418Z","level":"INFO","msg":"filestream: sending request","total_files":1,"events_offset":30,"events_lines":2}
+{"time":"2026-06-29T03:10:00.656101186Z","level":"INFO","msg":"filestream: request sent","status":"200 OK"}
+{"time":"2026-06-29T03:10:15.391545441Z","level":"INFO","msg":"filestream: sending request","total_files":1,"events_offset":32,"events_lines":2}
+{"time":"2026-06-29T03:10:15.748897158Z","level":"INFO","msg":"filestream: request sent","status":"200 OK"}
+{"time":"2026-06-29T03:10:30.392169205Z","level":"INFO","msg":"filestream: sending request","total_files":1,"events_offset":34,"events_lines":2}
+{"time":"2026-06-29T03:10:30.655413925Z","level":"INFO","msg":"filestream: request sent","status":"200 OK"}
+{"time":"2026-06-29T03:10:45.391855697Z","level":"INFO","msg":"filestream: sending request","total_files":1,"events_offset":36,"events_lines":2}
+{"time":"2026-06-29T03:10:45.59849883Z","level":"INFO","msg":"filestream: request sent","status":"200 OK"}
+{"time":"2026-06-29T03:11:00.391116493Z","level":"INFO","msg":"filestream: sending request","total_files":1,"events_offset":38,"events_lines":2}
+{"time":"2026-06-29T03:11:00.646388969Z","level":"INFO","msg":"filestream: request sent","status":"200 OK"}
+{"time":"2026-06-29T03:11:15.391517771Z","level":"INFO","msg":"filestream: sending request","total_files":1,"events_offset":40,"events_lines":2}
+{"time":"2026-06-29T03:11:15.61549784Z","level":"INFO","msg":"filestream: request sent","status":"200 OK"}
+{"time":"2026-06-29T03:11:30.391740479Z","level":"INFO","msg":"filestream: sending request","total_files":1,"events_offset":42,"events_lines":2}
+{"time":"2026-06-29T03:11:30.587070916Z","level":"INFO","msg":"filestream: request sent","status":"200 OK"}

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260629_030914-rgt7q8d2/logs/debug.log

@@ -0,0 +1,20 @@
+2026-06-29 03:09:14,120 INFO    MainThread:9369 [wandb_setup.py:_flush():81] Current SDK version is 0.28.0
+2026-06-29 03:09:14,120 INFO    MainThread:9369 [wandb_setup.py:_flush():81] Configure stats pid to 9369
+2026-06-29 03:09:14,121 INFO    MainThread:9369 [wandb_setup.py:_flush():81] Loading settings from environment variables
+2026-06-29 03:09:14,121 INFO    MainThread:9369 [wandb_init.py:setup_run_log_directory():725] Logging user logs to outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260629_030914-rgt7q8d2/logs/debug.log
+2026-06-29 03:09:14,121 INFO    MainThread:9369 [wandb_init.py:setup_run_log_directory():726] Logging internal logs to outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260629_030914-rgt7q8d2/logs/debug-internal.log
+2026-06-29 03:09:14,121 INFO    MainThread:9369 [wandb_init.py:init():768] calling init triggers
+2026-06-29 03:09:14,121 INFO    MainThread:9369 [wandb_init.py:init():773] wandb.init called with sweep_config: {}
+config: {'pretrained_model_root': './ckpts', 'pretrained_transformer_version': '480p_t2v', 'post_train_type': 'grpo', 'learning_rate': 1e-05, 'weight_decay': 0.01, 'max_steps': 10000, 'warmup_steps': 500, 'gradient_accumulation_steps': 1, 'max_grad_norm': 1.0, 'use_muon': True, 'num_train_timesteps': 1000, 'train_timestep_shift': 3.0, 'validation_timestep_shift': 5.0, 'snr_type': 'lognorm', 'task_type': 't2v', 'i2v_prob': 0.3, 'enable_fsdp': True, 'enable_gradient_checkpointing': True, 'sp_size': 2, 'dp_replicate': 1, 'batch_size': 2, 'num_workers': 4, 'prefetch_factor': 2, 'num_generations': 4, 'use_same_noise': False, 'output_dir': './outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3', 'checkpoints_directory': 's3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3', 'save_interval': 25, 'timer_interval': 0, 'auto_resume_from_timer': True, 'log_interval': 1, 'enable_wandb': True, 'project_name': 'SAGE-GRPO', 'run_name': 'yw006-4-sage-qwen35-9b-token-credit-video-adv-window3', 'wandb_key': None, 'wandb_username': None, 'wandb_base_url': 'https://api.wandb.ai', 'wandb_entity': None, 'wandb_silent': True, 'dtype': 'bf16', 'master_weight_type': 'fp32', 'seed': 42, 'validation_global_seed': 930, 'validation_interval': 0, 'validation_prompts': None, 'validate_video_length': 121, 'validation_aspect_ratio': '16:9', 'validation_fixed_size': '480x864', 'validation_num_inference_steps': 40, 'validation_negative_prompt': '', 'validation_enable_sr': False, 'validation_prompt_rewrite': False, 'validate_at_step0': False, 'resume_from_checkpoint': None, 'use_lora': False, 'lora_r': 8, 'lora_alpha': 16, 'lora_dropout': 0.0, 'lora_target_modules': None, 'pretrained_lora_path': None, 'train_video_csv': ['/nfs/ywang29/Reward_finetuning/SAGE-GRPO/assets/demo_train_decomposed_binary5.csv'], 'valid_video_csv': 'assets/demo_sample.csv', 'reward_model': 'snapvideo_vlm', 'reward_config': None, 'reward_checkpoint_mode': 'v3', 'reward_weights': '{"VQ":0.5,"MQ":0.5,"TA":1.0}', 'reward_gradient_credit': False, 'reward_credit_num_frames': 0, 'reward_credit_max_questions': 0, 'reward_credit_metric': 'weighted', 'token_credit_source_model': None, 'token_credit_advantage_source': 'token', 'token_credit_remote_url': None, 'token_credit_vlm_model_path': None, 'token_credit_vlm_model_family': None, 'token_credit_vlm_prompt_template': None, 'remote_reward_url': None, 'vlm_reward_model_path': '/nfs/ywang29/ckpts/Qwen3.5-9B', 'vlm_reward_model_family': 'qwen3_5', 'vlm_reward_prompt_template': None, 'vlm_reward_num_frames': 20, 'vlm_reward_max_pixels': 50176, 'vlm_reward_max_new_tokens': 128, 'vlm_reward_batch_size': 1, 'vlm_reward_score_type': 'token_credit', 'vlm_reward_missing_score': 0.0, 'vlm_reward_video_sample_type': 'uniform', 'vlm_reward_token_credit_max_questions': 5, 'token_credit_loss_mode': 'video_adv', 'token_credit_ratio_mode': 'weighted_logprob', 'token_credit_shuffle_weights': False, 'token_credit_spatial_granularity': 'window', 'token_credit_spatial_bins': 16, 'token_credit_spatial_window': 3, 'token_credit_chunk_size': 1, 'enable_grpo_policy_update': True, 'grpo_sampling_steps': 20, 'sample_n_frames': 81, 'eta': 0.5, 'cfg_scale': 6.0, 'neg_prompt': None, 'data_type': 'video', 'model_type': '16164', 'video_fps': 24, 'vae_name': '16164-32c-hy20250605', 'vae_tiling': False, 'infer_flow_shift_video': 5.0, 'video_bucket_hw_base_size': 480, 'video_bucket_hw_bucket_stride': 32, 'multitask_mask_training_type': 'concat', 'vision_num_semantic_tokens': 729, 'vision_states_dim': 1152, 'sde_type': 'sage_grpo', 'mini_batch_size_per_update': 1, 'clip_range': 0.0001, 'adv_clip_max': 5.0, 'timestep_fraction': 1.0, 'random_train_timestep_count': False, 'random_train_timestep_min': 1, 'random_train_timestep_max': 20, 'use_grad_balancing': True, 'enable_timestep_permutation': True, 'reference_mode_offload': True, 'kl_weight': 1e-05, 'kl_coef': 1e-07, 'kl_min_coef': 1e-07, 'use_moving_KL': True, 'update_ref_model_step': 10, 'use_dual_kl': True, 'dual_kl_moving_weight': 1.0, 'dual_kl_step_weight': 0.1, 'kl_compute_mode': 'rollout_phase', 'glyph_byT5_v2': True, 'byt5_max_length': 256, 'video_uncond_p_byt5': 0.0, 'precision': 'bf16', 'vae_precision': 'fp16', 'val_disable_autocast': False, 'debug_grad_flow': False, 'debug_train_diagnostics_interval': 1, '_wandb': {}}
+2026-06-29 03:09:14,121 INFO    MainThread:9369 [wandb_init.py:init():816] starting backend
+2026-06-29 03:09:14,343 INFO    MainThread:9369 [wandb_init.py:init():831] sending inform_init request
+2026-06-29 03:09:14,635 INFO    MainThread:9369 [wandb_init.py:init():836] backend started and connected
+2026-06-29 03:09:14,638 INFO    MainThread:9369 [wandb_init.py:init():906] updated telemetry
+2026-06-29 03:09:14,668 INFO    MainThread:9369 [wandb_init.py:init():929] communicating run to backend with 90.0 second timeout
+2026-06-29 03:09:15,188 INFO    MainThread:9369 [wandb_init.py:init():969] run resumed
+2026-06-29 03:09:15,189 INFO    MainThread:9369 [wandb_init.py:init():974] starting run threads in backend
+2026-06-29 03:09:15,387 INFO    MainThread:9369 [wandb_run.py:_console_start():2523] atexit reg
+2026-06-29 03:09:15,387 INFO    MainThread:9369 [wandb_run.py:_redirect():2373] redirect: wrap_raw
+2026-06-29 03:09:15,387 INFO    MainThread:9369 [wandb_run.py:_redirect():2442] Wrapping output streams.
+2026-06-29 03:09:15,387 INFO    MainThread:9369 [wandb_run.py:_redirect():2465] Redirects installed.
+2026-06-29 03:09:15,388 INFO    MainThread:9369 [wandb_init.py:init():1012] run started, returning control to user process

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260629_033809-rgt7q8d2/files/output.log

@@ -0,0 +1,133 @@
+2026-06-29 03:38:11.198 | INFO     | __main__:print_training_configuration:970 - SYSTEM & HARDWARE CONFIGURATION
+2026-06-29 03:38:11.199 | INFO     | __main__:print_training_configuration:971 - --------------------------------------------------------------------------------
+2026-06-29 03:38:11.199 | INFO     | __main__:print_training_configuration:972 - Number of GPUs                                   : 64
+2026-06-29 03:38:11.200 | INFO     | __main__:print_training_configuration:973 - Local rank                                       : 0
+2026-06-29 03:38:11.200 | INFO     | __main__:print_training_configuration:974 - Global rank                                      : 0
+2026-06-29 03:38:11.201 | INFO     | __main__:print_training_configuration:975 - Device                                           : cuda:0
+2026-06-29 03:38:11.201 | INFO     | __main__:print_training_configuration:976 - Master weight dtype                              : torch.float32
+2026-06-29 03:38:11.201 | INFO     | __main__:print_training_configuration:978 - --------------------------------------------------------------------------------
+2026-06-29 03:38:11.202 | INFO     | __main__:print_training_configuration:979 - MODEL PARAMETERS & ARCHITECTURE
+2026-06-29 03:38:11.202 | INFO     | __main__:print_training_configuration:980 - --------------------------------------------------------------------------------
+2026-06-29 03:38:11.214 | INFO     | __main__:print_training_configuration:983 - Number of total                                  : 8,326,608,160
+2026-06-29 03:38:11.214 | INFO     | __main__:print_training_configuration:984 - Total trainable parameters                        : 8,326,608,160
+2026-06-29 03:38:11.215 | INFO     | __main__:print_training_configuration:985 - DIT model root                                   : ./ckpts
+2026-06-29 03:38:11.215 | INFO     | __main__:print_training_configuration:986 - Transformer version                              : 480p_t2v
+2026-06-29 03:38:11.215 | INFO     | __main__:print_training_configuration:987 - Attention mode                                   : flash
+2026-06-29 03:38:11.216 | INFO     | __main__:print_training_configuration:988 - Gradient checkpointing                           : True
+2026-06-29 03:38:11.216 | INFO     | __main__:print_training_configuration:989 - FSDP enabled                                     : True
+2026-06-29 03:38:11.217 | INFO     | __main__:print_training_configuration:990 - LoRA enabled                                     : False
+2026-06-29 03:38:11.217 | INFO     | __main__:print_training_configuration:991 - Glyph byT5 v2                                    : True
+2026-06-29 03:38:11.218 | INFO     | __main__:print_training_configuration:992 - ByT5 max length                                  : 256
+2026-06-29 03:38:11.218 | INFO     | __main__:print_training_configuration:994 - --------------------------------------------------------------------------------
+2026-06-29 03:38:11.219 | INFO     | __main__:print_training_configuration:995 - DISTRIBUTED TRAINING CONFIGURATION
+2026-06-29 03:38:11.219 | INFO     | __main__:print_training_configuration:996 - --------------------------------------------------------------------------------
+2026-06-29 03:38:11.220 | INFO     | __main__:print_training_configuration:997 - World size                                       : 64
+2026-06-29 03:38:11.220 | INFO     | __main__:print_training_configuration:998 - DP degree                                        : 32
+2026-06-29 03:38:11.221 | INFO     | __main__:print_training_configuration:999 - DP rank                                          : 0
+2026-06-29 03:38:11.221 | INFO     | __main__:print_training_configuration:1000 - SP size                                          : 2
+2026-06-29 03:38:11.221 | INFO     | __main__:print_training_configuration:1001 - SP rank                                          : 0
+2026-06-29 03:38:11.222 | INFO     | __main__:print_training_configuration:1003 - --------------------------------------------------------------------------------
+2026-06-29 03:38:11.222 | INFO     | __main__:print_training_configuration:1004 - BATCH SIZE & DATA FLOW CONFIGURATION
+2026-06-29 03:38:11.223 | INFO     | __main__:print_training_configuration:1005 - --------------------------------------------------------------------------------
+2026-06-29 03:38:11.223 | INFO     | __main__:print_training_configuration:1006 - Per GPU batch size                               : 2
+2026-06-29 03:38:11.223 | INFO     | __main__:print_training_configuration:1007 - DP degree batch size                             : 64
+2026-06-29 03:38:11.224 | INFO     | __main__:print_training_configuration:1008 - Gradient accumulation steps                      : 1
+2026-06-29 03:38:11.224 | INFO     | __main__:print_training_configuration:1009 - Num generations                                  : 4
+2026-06-29 03:38:11.225 | INFO     | __main__:print_training_configuration:1010 - Use same noise                                   : False
+2026-06-29 03:38:11.225 | INFO     | __main__:print_training_configuration:1011 - Num workers                                      : 4
+2026-06-29 03:38:11.225 | INFO     | __main__:print_training_configuration:1012 - Prefetch factor                                  : 2
+2026-06-29 03:38:11.226 | INFO     | __main__:print_training_configuration:1014 - --------------------------------------------------------------------------------
+2026-06-29 03:38:11.226 | INFO     | __main__:print_training_configuration:1015 - OPTIMIZATION CONFIGURATION
+2026-06-29 03:38:11.227 | INFO     | __main__:print_training_configuration:1016 - --------------------------------------------------------------------------------
+2026-06-29 03:38:11.227 | INFO     | __main__:print_training_configuration:1017 - Optimizer                                        : muon
+2026-06-29 03:38:11.227 | INFO     | __main__:print_training_configuration:1018 - Learning rate                                    : 1e-05
+2026-06-29 03:38:11.228 | INFO     | __main__:print_training_configuration:1019 - Weight decay                                     : 0.01
+2026-06-29 03:38:11.228 | INFO     | __main__:print_training_configuration:1020 - Max gradient norm                                : 1.0
+2026-06-29 03:38:11.229 | INFO     | __main__:print_training_configuration:1021 - LR scheduler                                     : constant
+2026-06-29 03:38:11.229 | INFO     | __main__:print_training_configuration:1022 - Warmup steps                                     : 500
+2026-06-29 03:38:11.229 | INFO     | __main__:print_training_configuration:1023 - Max training steps                               : 10000
+2026-06-29 03:38:11.230 | INFO     | __main__:print_training_configuration:1025 - --------------------------------------------------------------------------------
+2026-06-29 03:38:11.230 | INFO     | __main__:print_training_configuration:1026 - INFERENCE & VALIDATION CONFIGURATION
+2026-06-29 03:38:11.230 | INFO     | __main__:print_training_configuration:1027 - --------------------------------------------------------------------------------
+2026-06-29 03:38:11.231 | INFO     | __main__:print_training_configuration:1028 - Validation interval                              : 0
+2026-06-29 03:38:11.231 | INFO     | __main__:print_training_configuration:1029 - Validate at step 0                               : False
+2026-06-29 03:38:11.232 | INFO     | __main__:print_training_configuration:1030 - Validation video length                          : 121
+2026-06-29 03:38:11.232 | INFO     | __main__:print_training_configuration:1031 - Validation global seed                           : 930
+2026-06-29 03:38:11.232 | INFO     | __main__:print_training_configuration:1032 - Validation inference steps                        : 40
+2026-06-29 03:38:11.233 | INFO     | __main__:print_training_configuration:1033 - Validation aspect ratio                          : 16:9
+2026-06-29 03:38:11.233 | INFO     | __main__:print_training_configuration:1034 - Validation fixed size                            : 480x864
+2026-06-29 03:38:11.234 | INFO     | __main__:print_training_configuration:1035 - Validation prompt rewrite                         : False
+2026-06-29 03:38:11.234 | INFO     | __main__:print_training_configuration:1036 - Validation flow shift                             : 5.0
+2026-06-29 03:38:11.234 | INFO     | __main__:print_training_configuration:1037 - Sampling CFG scale                                : 6.0
+2026-06-29 03:38:11.235 | INFO     | __main__:print_training_configuration:1038 - Sampling eta                                      : 0.5
+2026-06-29 03:38:11.235 | INFO     | __main__:print_training_configuration:1039 - Sampling steps (GRPO rollout)                     : 20
+2026-06-29 03:38:11.236 | INFO     | __main__:print_training_configuration:1040 - Token credit loss mode                            : video_adv
+2026-06-29 03:38:11.236 | INFO     | __main__:print_training_configuration:1041 - Token credit ratio mode                           : weighted_logprob
+2026-06-29 03:38:11.236 | INFO     | __main__:print_training_configuration:1042 - Token credit shuffle weights                      : False
+2026-06-29 03:38:11.237 | INFO     | __main__:print_training_configuration:1043 - Token credit spatial granularity                  : window
+2026-06-29 03:38:11.237 | INFO     | __main__:print_training_configuration:1044 - Token credit spatial bins                         : 16
+2026-06-29 03:38:11.238 | INFO     | __main__:print_training_configuration:1045 - Token credit spatial window                       : 3
+2026-06-29 03:38:11.238 | INFO     | __main__:print_training_configuration:1046 - Token credit chunk size                           : 1
+2026-06-29 03:38:11.238 | INFO     | __main__:print_training_configuration:1047 - Token credit source model                         : None
+2026-06-29 03:38:11.239 | INFO     | __main__:print_training_configuration:1048 - Token credit advantage source                     : token
+2026-06-29 03:38:11.239 | INFO     | __main__:print_training_configuration:1049 - Token credit VLM model path                       : None
+2026-06-29 03:38:11.239 | INFO     | __main__:print_training_configuration:1050 - Token credit VLM model family                     : None
+2026-06-29 03:38:11.240 | INFO     | __main__:print_training_configuration:1051 - Random train timestep count                       : False
+2026-06-29 03:38:11.240 | INFO     | __main__:print_training_configuration:1052 - Random train timestep range                       : 1-20
+2026-06-29 03:38:11.241 | INFO     | __main__:print_training_configuration:1056 - SDE type                                          : sage_grpo
+2026-06-29 03:38:11.241 | INFO     | __main__:print_training_configuration:1058 - --------------------------------------------------------------------------------
+2026-06-29 03:38:11.241 | INFO     | __main__:print_training_configuration:1059 - GRPO & KL CONFIGURATION
+2026-06-29 03:38:11.242 | INFO     | __main__:print_training_configuration:1060 - --------------------------------------------------------------------------------
+2026-06-29 03:38:11.242 | INFO     | __main__:print_training_configuration:1061 - Reward model                                     : snapvideo_vlm
+2026-06-29 03:38:11.242 | INFO     | __main__:print_training_configuration:1062 - Reward gradient credit                          : False
+2026-06-29 03:38:11.243 | INFO     | __main__:print_training_configuration:1063 - Reward credit frames                            : 0
+2026-06-29 03:38:11.243 | INFO     | __main__:print_training_configuration:1064 - Reward credit max questions                     : 0
+2026-06-29 03:38:11.244 | INFO     | __main__:print_training_configuration:1065 - Reward credit metric                            : weighted
+2026-06-29 03:38:11.244 | INFO     | __main__:print_training_configuration:1067 - VLM reward model path                            : /nfs/ywang29/ckpts/Qwen3.5-9B
+2026-06-29 03:38:11.244 | INFO     | __main__:print_training_configuration:1068 - VLM reward model family                          : qwen3_5
+2026-06-29 03:38:11.245 | INFO     | __main__:print_training_configuration:1069 - VLM reward frames                                : 20
+2026-06-29 03:38:11.245 | INFO     | __main__:print_training_configuration:1070 - VLM reward max pixels                            : 50176
+2026-06-29 03:38:11.245 | INFO     | __main__:print_training_configuration:1071 - VLM reward score type                            : token_credit
+2026-06-29 03:38:11.246 | INFO     | __main__:print_training_configuration:1072 - VLM reward max new tokens                        : 128
+2026-06-29 03:38:11.246 | INFO     | __main__:print_training_configuration:1073 - VLM reward batch size                            : 1
+2026-06-29 03:38:11.246 | INFO     | __main__:print_training_configuration:1074 - VLM reward missing score                         : 0.0
+2026-06-29 03:38:11.247 | INFO     | __main__:print_training_configuration:1077 - KL weight                                        : 1e-05
+2026-06-29 03:38:11.247 | INFO     | __main__:print_training_configuration:1078 - KL coef                                          : 1e-07
+2026-06-29 03:38:11.247 | INFO     | __main__:print_training_configuration:1079 - Use moving KL                                    : True
+2026-06-29 03:38:11.247 | INFO     | __main__:print_training_configuration:1080 - Update ref model step                            : 10
+2026-06-29 03:38:11.248 | INFO     | __main__:print_training_configuration:1081 - Use dual KL                                      : True
+2026-06-29 03:38:11.248 | INFO     | __main__:print_training_configuration:1082 - Dual KL moving weight                            : 1.0
+2026-06-29 03:38:11.249 | INFO     | __main__:print_training_configuration:1083 - Dual KL step weight                              : 0.1
+2026-06-29 03:38:11.250 | INFO     | __main__:print_training_configuration:1084 - KL compute mode                                  : rollout_phase
+2026-06-29 03:38:11.250 | INFO     | __main__:print_training_configuration:1085 - Use grad balancing                               : True
+2026-06-29 03:38:11.250 | INFO     | __main__:print_training_configuration:1086 - Enable timestep permutation                      : True
+2026-06-29 03:38:11.251 | INFO     | __main__:print_training_configuration:1087 - Reference mode offload                           : True
+2026-06-29 03:38:11.251 | INFO     | __main__:print_training_configuration:1089 - --------------------------------------------------------------------------------
+2026-06-29 03:38:11.251 | INFO     | __main__:print_training_configuration:1090 - LOGGING & CHECKPOINTING CONFIGURATION
+2026-06-29 03:38:11.252 | INFO     | __main__:print_training_configuration:1091 - --------------------------------------------------------------------------------
+2026-06-29 03:38:11.252 | INFO     | __main__:print_training_configuration:1092 - Output directory                                 : ./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3
+2026-06-29 03:38:11.253 | INFO     | __main__:print_training_configuration:1093 - Checkpoints directory                            : s3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3
+2026-06-29 03:38:11.253 | INFO     | __main__:print_training_configuration:1094 - Checkpoint interval                              : 25
+2026-06-29 03:38:11.254 | INFO     | __main__:print_training_configuration:1095 - Timer checkpoint interval                        : 0
+2026-06-29 03:38:11.254 | INFO     | __main__:print_training_configuration:1096 - Auto resume from timer/checkpoint                : True
+2026-06-29 03:38:11.255 | INFO     | __main__:print_training_configuration:1097 - Timer checkpoint directory                       : s3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/timer
+2026-06-29 03:38:11.255 | INFO     | __main__:print_training_configuration:1098 - Log interval                                     : 1
+2026-06-29 03:38:11.255 | INFO     | __main__:print_training_configuration:1099 - Resume checkpoint                                : None
+2026-06-29 03:38:11.256 | INFO     | __main__:print_training_configuration:1100 - Train CSV                                         : ['/nfs/ywang29/Reward_finetuning/SAGE-GRPO/assets/demo_train_decomposed_binary5.csv']
+2026-06-29 03:38:11.256 | INFO     | __main__:print_training_configuration:1101 - Valid CSV                                         : assets/demo_sample.csv
+2026-06-29 03:38:11.257 | INFO     | __main__:print_training_configuration:1102 - Seed                                             : 42
+2026-06-29 03:38:11.257 | INFO     | __main__:print_training_configuration:1103 - WandB enabled                                    : True
+2026-06-29 03:38:11.257 | INFO     | __main__:print_training_configuration:1105 - WandB project                                    : SAGE-GRPO
+2026-06-29 03:38:11.258 | INFO     | __main__:print_training_configuration:1106 - WandB run name                                   : yw006-4-sage-qwen35-9b-token-credit-video-adv-window3
+2026-06-29 03:38:11.258 | INFO     | __main__:__init__:958 - [Effective Config] world_size=64, rank=0, local_rank=0, sp_size=2, dp_size=32, dp_rank=0, sp_rank=0, batch_size=2, num_generations=4, grad_accum=1
+2026-06-29 03:38:11.259 | INFO     | __main__:_build_video_dataset:2088 - [Effective GRPO Grouping] video_batch_size=2, dp_replicas=32, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, product=64
+2026-06-29 03:38:11.927 | INFO     | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1178 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1
+2026-06-29 03:38:11.928 | INFO     | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1205 - Dataloader init done
+2026-06-29 03:38:11.929 | INFO     | __main__:_train_grpo:2171 - Starting GRPO training...
+2026-06-29 03:38:11.929 | INFO     | __main__:_train_grpo:2172 - Max steps: 10000
+2026-06-29 03:38:11.930 | INFO     | __main__:_train_grpo:2173 - Batch size: 2
+2026-06-29 03:38:11.930 | INFO     | __main__:_train_grpo:2174 - Learning rate: 1e-05
+2026-06-29 03:38:12.829 | INFO     | __main__:_resolve_resume_checkpoint:1860 - No timer checkpoint found at s3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/timer; auto-resuming from latest regular checkpoint: s3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/checkpoint-150
+2026-06-29 03:38:12.968 | INFO     | __main__:load_checkpoint:1992 - Loading checkpoint from s3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/checkpoint-150
+/usr/local/lib/python3.12/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)

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260629_033809-rgt7q8d2/files/requirements.txt

@@ -0,0 +1,287 @@
+awscli==1.45.30
+boto3==1.43.30
+botocore==1.43.30
+certifi==2026.5.20
+charset-normalizer==3.4.7
+colorama==0.4.6
+docutils==0.19
+idna==3.18
+jmespath==1.1.0
+pip==26.1.2
+pyasn1==0.6.3
+python-dateutil==2.9.0.post0
+PyYAML==6.0.3
+requests==2.34.2
+rsa==4.7.2
+s3transfer==0.18.0
+six==1.17.0
+urllib3==2.7.0
+py-cpuinfo==9.0.0
+hjson==3.1.0
+easydict==1.13
+braceexpand==0.1.7
+antlr4-python3-runtime==4.9.3
+websocket-client==1.9.0
+webdataset==1.0.2
+triton==3.6.0
+torchcodec==0.14.0+cu130
+tifffile==2026.6.1
+tenacity==9.1.4
+smmap==5.0.3
+sentry-sdk==2.63.0
+sentencepiece==0.2.1
+regex==2026.6.28
+pyarrow==24.0.0
+pyaml==26.2.1
+propcache==0.5.2
+opencv-python-headless==4.13.0.92
+omegaconf==2.3.1
+nvidia-nvtx==13.0.85
+ImageIO==2.37.3
+nvidia-nvshmem-cu13==3.4.5
+nvidia-nvjitlink==13.0.88
+av==17.1.0
+nvidia-nccl-cu13==2.28.9
+nvidia-cufile==1.15.1.6
+imageio-ffmpeg==0.6.0
+nvidia-cuda-runtime==13.0.96
+nvidia-cuda-nvrtc==13.0.88
+attrs==26.1.0
+nvidia-cuda-cupti==13.0.85
+nvidia-cublas==13.1.0.3
+multidict==6.7.1
+msgpack==1.2.1
+Mako==1.3.12
+loguru==0.7.3
+lmdb==2.2.1
+lightning-utilities==0.15.3
+lazy-loader==0.5
+greenlet==3.5.3
+future==1.0.0
+ftfy==6.3.1
+frozenlist==1.8.0
+frozendict==2.4.7
+docstring_parser==0.18.0
+decord==0.6.0
+cuda-pathfinder==1.5.5
+aiohappyeyeballs==2.6.2
+yarl==1.24.2
+typed-argument-parser==1.12.0
+SQLAlchemy==2.0.51
+scikit-image==0.26.0
+qwen-vl-utils==0.0.14
+typeguard==4.5.2
+nvidia-cusparse==12.6.3.3
+nvidia-cufft==12.0.0.61
+nvidia-cudnn-cu13==9.15.1.9
+gitdb==4.0.12
+ffmpeg-python==0.2.0
+einx==0.4.3
+cupy-cuda12x==14.1.1
+cuda-bindings==13.0.3
+cattrs==26.1.0
+aiosignal==1.4.0
+tokenizers==0.22.2
+termcolor==3.3.0
+nvidia-cusolver==12.0.4.66
+mediapy==1.2.6
+ipdb==0.13.13
+GitPython==3.1.50
+diffusers==0.38.0
+coremltools==9.0
+alembic==1.18.5
+aiohttp==3.14.1
+wandb==0.28.0
+torch==2.10.0+cu130
+dataset==2.0.0
+dashscope==1.26.0
+torchvision==0.25.0+cu130
+torchmetrics==1.9.0
+sniffio==1.3.1
+torchdata==0.8.0
+tqdm==4.67.3
+torchaudio==2.10.0+cu130
+deepspeed==0.19.2
+timm==1.0.27
+pytorch-lightning==2.6.5
+peft==0.19.1
+lpips==0.1.4
+lightning==2.6.5
+flash_attn==2.8.3+cu130torch2.10
+xxhash==3.8.0
+shortuuid==1.0.13
+setuptools==82.0.0
+safetensors==0.7.0
+pillow==12.1.1
+numpy==2.4.2
+jiter==0.15.0
+fsspec==2026.2.0
+distro==1.9.0
+dill==0.4.1
+tyro==1.0.15
+tiktoken==0.13.0
+multiprocess==0.70.19
+modelscope==1.36.0
+fire==0.7.0
+openai==2.8.0
+datasets==4.8.4
+trl==0.8.6
+angelslim==0.2.2
+huggingface_hub==1.21.0
+transformers==5.12.1
+black==26.5.1
+pathspec==1.1.1
+pytokens==0.4.1
+importlib_metadata==9.0.0
+ml_dtypes==0.5.4
+onnx==1.22.0
+onnx-ir==0.2.1
+onnxscript==0.7.0
+transformer_engine==2.9.0+70f53666
+zipp==4.1.0
+nvidia-mathdx==25.6.0
+einops==0.8.2
+absl-py==2.4.0
+accelerate==1.14.0
+annotated-doc==0.0.4
+annotated-types==0.7.0
+anyio==4.14.0
+apsw==3.53.2.0
+apswutils==0.1.2
+beartype==0.22.9
+beautifulsoup4==4.15.0
+blis==1.2.1
+catalogue==2.0.10
+cloudpathlib==0.24.0
+cloudpickle==3.1.2
+confection==0.1.5
+contourpy==1.3.3
+cycler==0.12.1
+cymem==2.0.13
+fastai==2.8.7
+fastcore==1.13.6
+fastdownload==0.0.7
+fastlite==0.2.4
+fastprogress==1.1.6
+fasttransform==0.0.2
+fonttools==4.63.0
+grpcio==1.81.1
+h11==0.16.0
+hf-xet==1.5.1
+httpcore==1.0.9
+httpcore2==2.4.0
+httptools==0.8.0
+httpx==0.28.1
+httpx2==2.4.0
+itsdangerous==2.2.0
+joblib==1.5.3
+kiwisolver==1.5.0
+langcodes==3.5.1
+Markdown==3.10.2
+markdown-it-py==4.2.0
+matplotlib==3.11.0
+mdurl==0.1.2
+mpmath==1.3.0
+murmurhash==1.0.15
+mypy_extensions==1.1.0
+narwhals==2.22.1
+networkx==3.6.1
+nvidia-curand==10.4.0.35
+nvidia-cusparselt-cu13==0.8.0
+oauthlib==3.3.1
+pandas==3.0.3
+plum-dispatch==2.9.0
+preshed==3.0.13
+protobuf==7.35.1
+pydantic==2.13.4
+pydantic_core==2.46.4
+pyparsing==3.3.2
+pyre-extensions==0.0.32
+python-dotenv==1.2.2
+python-fasthtml==0.14.3
+python-multipart==0.0.32
+rich==15.0.0
+s3torchconnector==1.5.0
+s3torchconnectorclient==1.5.0
+scikit-learn==1.9.0
+shellingham==1.5.4
+smart_open==7.6.1
+soupsieve==2.8.4
+spacy==3.8.7
+spacy-legacy==3.0.12
+spacy-loggers==1.0.5
+srsly==2.5.3
+starlette==1.3.1
+sympy==1.14.0
+tabulate==0.10.0
+tensorboard==2.20.0
+tensorboard-data-server==0.7.2
+thinc==8.3.4
+threadpoolctl==3.6.0
+torchtnt==0.2.4
+truststore==0.10.4
+typer==0.25.1
+typer-slim==0.24.0
+typing-inspect==0.9.0
+typing-inspection==0.4.2
+uvicorn==0.49.0
+uvloop==0.22.1
+wasabi==1.1.3
+watchfiles==1.2.0
+weasel==0.4.3
+websockets==16.0
+Werkzeug==3.1.8
+wrapt==2.2.2
+asttokens==3.0.1
+cffi==2.0.0
+click==8.4.1
+comm==0.2.3
+cryptography==49.0.0
+Cython==3.2.5
+debugpy==1.8.21
+decorator==5.3.1
+executing==2.2.1
+filelock==3.29.4
+h5py==3.16.0
+intel-cmplr-lib-ur==2026.0.0
+intel-openmp==2026.0.0
+ipykernel==7.3.0
+ipython==9.14.1
+ipython_pygments_lexers==1.1.1
+jedi==0.20.0
+Jinja2==3.1.6
+jupyter_client==8.9.1
+jupyter_core==5.9.1
+MarkupSafe==3.0.3
+matplotlib-inline==0.2.2
+mkl==2026.0.0
+mkl-include==2026.0.0
+mpi4py==4.1.2
+nest-asyncio2==1.7.2
+ninja==1.13.0
+onemkl-license==2026.0.0
+opencv-python==4.11.0.86
+packaging==26.2
+parso==0.8.7
+pexpect==4.9.0
+platformdirs==4.10.0
+prompt_toolkit==3.0.52
+psutil==7.2.2
+ptyprocess==0.7.0
+pure_eval==0.2.3
+pybind11==3.0.4
+pycparser==3.0
+Pygments==2.20.0
+pyOpenSSL==26.3.0
+pyzmq==27.1.0
+scipy==1.18.0
+stack-data==0.6.3
+tbb==2023.0.0
+tcmlib==1.5.0
+tornado==6.5.7
+traitlets==5.15.1
+typing_extensions==4.15.0
+tzdata==2026.2
+umf==1.1.0
+wheel==0.47.0
+wcwidth==0.8.1

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260629_033809-rgt7q8d2/files/wandb-metadata.json

@@ -0,0 +1,185 @@
+{
+  "os": "Linux-6.1.163-186.299.amzn2023.x86_64-x86_64-with-glibc2.35",
+  "python": "CPython 3.12.10",
+  "startedAt": "2026-06-29T03:38:09.984689Z",
+  "args": [
+    "--pretrained_model_root",
+    "./ckpts",
+    "--train_video_csv",
+    "/nfs/ywang29/Reward_finetuning/SAGE-GRPO/assets/demo_train_decomposed_binary5.csv",
+    "--reward_model",
+    "snapvideo_vlm",
+    "--vlm_reward_model_path",
+    "/nfs/ywang29/ckpts/Qwen3.5-9B",
+    "--vlm_reward_model_family",
+    "qwen3_5",
+    "--vlm_reward_num_frames",
+    "20",
+    "--vlm_reward_max_pixels",
+    "50176",
+    "--vlm_reward_max_new_tokens",
+    "128",
+    "--vlm_reward_batch_size",
+    "1",
+    "--vlm_reward_score_type",
+    "token_credit",
+    "--vlm_reward_token_credit_max_questions",
+    "5",
+    "--token_credit_loss_mode",
+    "video_adv",
+    "--token_credit_ratio_mode",
+    "weighted_logprob",
+    "--token_credit_spatial_granularity",
+    "window",
+    "--token_credit_spatial_bins",
+    "16",
+    "--token_credit_spatial_window",
+    "3",
+    "--token_credit_chunk_size",
+    "1",
+    "--learning_rate",
+    "1e-5",
+    "--batch_size",
+    "2",
+    "--num_generations",
+    "4",
+    "--max_steps",
+    "10000",
+    "--output_dir",
+    "./outputs",
+    "--enable_wandb",
+    "True",
+    "--project_name",
+    "SAGE-GRPO",
+    "--run_name",
+    "yw006-4-sage-qwen35-9b-token-credit-video-adv-window3",
+    "--enable_fsdp",
+    "--enable_gradient_checkpointing",
+    "--sp_size",
+    "2",
+    "--reward_checkpoint_mode",
+    "v3",
+    "--validation_interval",
+    "0",
+    "--validate_at_step0",
+    "False",
+    "--validate_video_length",
+    "121",
+    "--validation_timestep_shift",
+    "5.0",
+    "--use_grad_balancing",
+    "True",
+    "--enable_timestep_permutation",
+    "True",
+    "--sde_type",
+    "sage_grpo",
+    "--kl_weight",
+    "1e-5",
+    "--kl_coef",
+    "1e-7",
+    "--use_moving_KL",
+    "True",
+    "--update_ref_model_step",
+    "10",
+    "--use_dual_kl",
+    "True",
+    "--dual_kl_moving_weight",
+    "1.0",
+    "--dual_kl_step_weight",
+    "0.1",
+    "--reference_mode_offload",
+    "True",
+    "--save_interval",
+    "25",
+    "--timer_interval",
+    "0",
+    "--auto_resume_from_timer",
+    "True",
+    "--checkpoints_directory",
+    "s3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3"
+  ],
+  "program": "/nfs/ywang29/Reward_finetuning/SAGE-GRPO/post_train.py",
+  "codePath": "post_train.py",
+  "codePathLocal": "post_train.py",
+  "git": {
+    "remote": "https://github.com/Tencent-Hunyuan/SAGE-GRPO.git",
+    "commit": "779776e42732e0262b3f2844dd206d9d8f043b2f"
+  },
+  "root": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb",
+  "host": "ywang29-hunyuan-sage-qwen35-tcva-w3-worker-0",
+  "username": "yi-fan-wang1216",
+  "executable": "/usr/local/bin/python",
+  "cpu_count": 48,
+  "cpu_count_logical": 96,
+  "gpu": "NVIDIA A100-SXM4-80GB",
+  "gpu_count": 8,
+  "disk": {
+    "/": {
+      "total": "7516777873408",
+      "used": "619665272832"
+    }
+  },
+  "memory": {
+    "total": "1204539252736"
+  },
+  "gpu_nvidia": [
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-f6eafbbf-b72f-1725-61a0-5da7fda21fe2"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-8faea6eb-ec0a-0561-c1a7-608821eb24f5"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-6d3845d3-5be7-3130-93f6-aa8bc1fc8f9b"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-a80bc0c5-98ed-a833-ec83-51e290f6ac9c"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-a8dcbb50-ac55-1c6b-f4a3-66a2e17a2e9f"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-a6632210-2f96-54eb-7c24-6689d8c3ab90"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-e2324dd0-03ea-7053-4212-42facf1acdf7"
+    },
+    {
+      "name": "NVIDIA A100-SXM4-80GB",
+      "memoryTotal": "85899345920",
+      "cudaCores": 6912,
+      "architecture": "Ampere",
+      "uuid": "GPU-20e27620-2f39-65f1-ff0c-215b32505ff5"
+    }
+  ],
+  "cudaVersion": "13.0",
+  "writerId": "dlt04glepqwfbkknd3tz6wwrik6mxp1x"
+}

outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/wandb/wandb/run-20260629_033809-rgt7q8d2/logs/debug-internal.log

@@ -0,0 +1,11 @@
+{"time":"2026-06-29T03:38:10.234216477Z","level":"INFO","msg":"wandb-core"}
+{"time":"2026-06-29T03:38:10.235553516Z","level":"INFO","msg":"stream: starting","core version":"0.28.0"}
+{"time":"2026-06-29T03:38:10.621265925Z","level":"INFO","msg":"stream: created new stream","id":"rgt7q8d2"}
+{"time":"2026-06-29T03:38:10.621362184Z","level":"INFO","msg":"handler: started"}
+{"time":"2026-06-29T03:38:10.625022843Z","level":"INFO","msg":"stream: started"}
+{"time":"2026-06-29T03:38:10.625051177Z","level":"INFO","msg":"writer: started","stream_id":"rgt7q8d2"}
+{"time":"2026-06-29T03:38:10.625078885Z","level":"INFO","msg":"sender: started"}
+{"time":"2026-06-29T03:38:11.199728402Z","level":"INFO","msg":"filestream: sending request","total_files":1,"console_offset":665,"console_lines":1}
+{"time":"2026-06-29T03:38:11.42280913Z","level":"INFO","msg":"filestream: request sent","status":"200 OK"}
+{"time":"2026-06-29T03:38:26.201049848Z","level":"INFO","msg":"filestream: sending request","total_files":2,"events_offset":44,"events_lines":2,"console_offset":666,"console_lines":132,"uploaded_len":2}
+{"time":"2026-06-29T03:38:26.425015639Z","level":"INFO","msg":"filestream: request sent","status":"200 OK"}