| 2026-04-21 14:48:29.503 | INFO | __main__:__init__:806 - [RewardDebug] reward_weights(raw): {'VQ': 0.5, 'MQ': 0.5, 'TA': 1.0} |
| 2026-04-21 14:48:29.504 | INFO | __main__:__init__:807 - [RewardDebug] reward_config(effective): {'models': {'videoalign_local': {'weight': 1.0, 'sub_reward': {'VQ': 0.5, 'MQ': 0.5, 'TA': 1.0}}}, 'reward_checkpoint_mode': 'v3'} |
| 2026-04-21 14:48:31.474 | INFO | __main__:print_training_configuration:830 - SYSTEM & HARDWARE CONFIGURATION |
| 2026-04-21 14:48:31.475 | INFO | __main__:print_training_configuration:831 - -------------------------------------------------------------------------------- |
| 2026-04-21 14:48:31.475 | INFO | __main__:print_training_configuration:832 - Number of GPUs : 8 |
| 2026-04-21 14:48:31.475 | INFO | __main__:print_training_configuration:833 - Local rank : 0 |
| 2026-04-21 14:48:31.476 | INFO | __main__:print_training_configuration:834 - Global rank : 0 |
| 2026-04-21 14:48:31.476 | INFO | __main__:print_training_configuration:835 - Device : cuda:0 |
| 2026-04-21 14:48:31.477 | INFO | __main__:print_training_configuration:836 - Master weight dtype : torch.float32 |
| 2026-04-21 14:48:31.477 | INFO | __main__:print_training_configuration:838 - -------------------------------------------------------------------------------- |
| 2026-04-21 14:48:31.478 | INFO | __main__:print_training_configuration:839 - MODEL PARAMETERS & ARCHITECTURE |
| 2026-04-21 14:48:31.478 | INFO | __main__:print_training_configuration:840 - -------------------------------------------------------------------------------- |
| 2026-04-21 14:48:31.491 | INFO | __main__:print_training_configuration:843 - Number of total : 8,326,608,160 |
| 2026-04-21 14:48:31.491 | INFO | __main__:print_training_configuration:844 - Total trainable parameters : 8,326,608,160 |
| 2026-04-21 14:48:31.492 | INFO | __main__:print_training_configuration:845 - DIT model root : ./ckpts |
| 2026-04-21 14:48:31.493 | INFO | __main__:print_training_configuration:846 - Transformer version : 480p_t2v |
| 2026-04-21 14:48:31.493 | INFO | __main__:print_training_configuration:847 - Attention mode : flash |
| 2026-04-21 14:48:31.493 | INFO | __main__:print_training_configuration:848 - Gradient checkpointing : True |
| 2026-04-21 14:48:31.494 | INFO | __main__:print_training_configuration:849 - FSDP enabled : True |
| 2026-04-21 14:48:31.494 | INFO | __main__:print_training_configuration:850 - LoRA enabled : False |
| 2026-04-21 14:48:31.495 | INFO | __main__:print_training_configuration:851 - Glyph byT5 v2 : True |
| 2026-04-21 14:48:31.495 | INFO | __main__:print_training_configuration:852 - ByT5 max length : 256 |
| 2026-04-21 14:48:31.496 | INFO | __main__:print_training_configuration:854 - -------------------------------------------------------------------------------- |
| 2026-04-21 14:48:31.496 | INFO | __main__:print_training_configuration:855 - DISTRIBUTED TRAINING CONFIGURATION |
| 2026-04-21 14:48:31.497 | INFO | __main__:print_training_configuration:856 - -------------------------------------------------------------------------------- |
| 2026-04-21 14:48:31.497 | INFO | __main__:print_training_configuration:857 - World size : 8 |
| 2026-04-21 14:48:31.497 | INFO | __main__:print_training_configuration:858 - DP degree : 4 |
| 2026-04-21 14:48:31.498 | INFO | __main__:print_training_configuration:859 - DP rank : 0 |
| 2026-04-21 14:48:31.498 | INFO | __main__:print_training_configuration:860 - SP size : 2 |
| 2026-04-21 14:48:31.499 | INFO | __main__:print_training_configuration:861 - SP rank : 0 |
| 2026-04-21 14:48:31.499 | INFO | __main__:print_training_configuration:863 - -------------------------------------------------------------------------------- |
| 2026-04-21 14:48:31.500 | INFO | __main__:print_training_configuration:864 - BATCH SIZE & DATA FLOW CONFIGURATION |
| 2026-04-21 14:48:31.500 | INFO | __main__:print_training_configuration:865 - -------------------------------------------------------------------------------- |
| 2026-04-21 14:48:31.501 | INFO | __main__:print_training_configuration:866 - Per GPU batch size : 2 |
| 2026-04-21 14:48:31.501 | INFO | __main__:print_training_configuration:867 - DP degree batch size : 8 |
| 2026-04-21 14:48:31.502 | INFO | __main__:print_training_configuration:868 - Gradient accumulation steps : 1 |
| 2026-04-21 14:48:31.502 | INFO | __main__:print_training_configuration:869 - Num generations : 4 |
| 2026-04-21 14:48:31.502 | INFO | __main__:print_training_configuration:870 - Use same noise : False |
| 2026-04-21 14:48:31.503 | INFO | __main__:print_training_configuration:871 - Num workers : 4 |
| 2026-04-21 14:48:31.503 | INFO | __main__:print_training_configuration:872 - Prefetch factor : 2 |
| 2026-04-21 14:48:31.504 | INFO | __main__:print_training_configuration:874 - -------------------------------------------------------------------------------- |
| 2026-04-21 14:48:31.504 | INFO | __main__:print_training_configuration:875 - OPTIMIZATION CONFIGURATION |
| 2026-04-21 14:48:31.504 | INFO | __main__:print_training_configuration:876 - -------------------------------------------------------------------------------- |
| 2026-04-21 14:48:31.505 | INFO | __main__:print_training_configuration:877 - Optimizer : muon |
| 2026-04-21 14:48:31.505 | INFO | __main__:print_training_configuration:878 - Learning rate : 1e-05 |
| 2026-04-21 14:48:31.506 | INFO | __main__:print_training_configuration:879 - Weight decay : 0.01 |
| 2026-04-21 14:48:31.506 | INFO | __main__:print_training_configuration:880 - Max gradient norm : 1.0 |
| 2026-04-21 14:48:31.507 | INFO | __main__:print_training_configuration:881 - LR scheduler : constant |
| 2026-04-21 14:48:31.507 | INFO | __main__:print_training_configuration:882 - Warmup steps : 500 |
| 2026-04-21 14:48:31.508 | INFO | __main__:print_training_configuration:883 - Max training steps : 10000 |
| 2026-04-21 14:48:31.508 | INFO | __main__:print_training_configuration:885 - -------------------------------------------------------------------------------- |
| 2026-04-21 14:48:31.509 | INFO | __main__:print_training_configuration:886 - INFERENCE & VALIDATION CONFIGURATION |
| 2026-04-21 14:48:31.509 | INFO | __main__:print_training_configuration:887 - -------------------------------------------------------------------------------- |
| 2026-04-21 14:48:31.509 | INFO | __main__:print_training_configuration:888 - Validation interval : 100 |
| 2026-04-21 14:48:31.510 | INFO | __main__:print_training_configuration:889 - Validate at step 0 : False |
| 2026-04-21 14:48:31.510 | INFO | __main__:print_training_configuration:890 - Validation video length : 121 |
| 2026-04-21 14:48:31.511 | INFO | __main__:print_training_configuration:891 - Validation global seed : 930 |
| 2026-04-21 14:48:31.511 | INFO | __main__:print_training_configuration:892 - Validation inference steps : 40 |
| 2026-04-21 14:48:31.512 | INFO | __main__:print_training_configuration:893 - Validation aspect ratio : 16:9 |
| 2026-04-21 14:48:31.512 | INFO | __main__:print_training_configuration:894 - Validation fixed size : 480x864 |
| 2026-04-21 14:48:31.513 | INFO | __main__:print_training_configuration:895 - Validation prompt rewrite : False |
| 2026-04-21 14:48:31.513 | INFO | __main__:print_training_configuration:896 - Validation flow shift : 5.0 |
| 2026-04-21 14:48:31.513 | INFO | __main__:print_training_configuration:897 - Sampling CFG scale : 6.0 |
| 2026-04-21 14:48:31.514 | INFO | __main__:print_training_configuration:898 - Sampling eta : 0.5 |
| 2026-04-21 14:48:31.514 | INFO | __main__:print_training_configuration:899 - Sampling steps (GRPO rollout) : 20 |
| 2026-04-21 14:48:31.514 | INFO | __main__:print_training_configuration:900 - SDE type : sage_grpo |
| 2026-04-21 14:48:31.515 | INFO | __main__:print_training_configuration:902 - -------------------------------------------------------------------------------- |
| 2026-04-21 14:48:31.515 | INFO | __main__:print_training_configuration:903 - GRPO & KL CONFIGURATION |
| 2026-04-21 14:48:31.515 | INFO | __main__:print_training_configuration:904 - -------------------------------------------------------------------------------- |
| 2026-04-21 14:48:31.516 | INFO | __main__:print_training_configuration:905 - Reward model : videoalign_local |
| 2026-04-21 14:48:31.516 | INFO | __main__:print_training_configuration:906 - Reward checkpoint mode : v3 |
| 2026-04-21 14:48:31.517 | INFO | __main__:print_training_configuration:907 - KL weight : 1e-05 |
| 2026-04-21 14:48:31.517 | INFO | __main__:print_training_configuration:908 - KL coef : 1e-07 |
| 2026-04-21 14:48:31.518 | INFO | __main__:print_training_configuration:909 - Use moving KL : True |
| 2026-04-21 14:48:31.518 | INFO | __main__:print_training_configuration:910 - Update ref model step : 10 |
| 2026-04-21 14:48:31.518 | INFO | __main__:print_training_configuration:911 - Use dual KL : True |
| 2026-04-21 14:48:31.519 | INFO | __main__:print_training_configuration:912 - Dual KL moving weight : 1.0 |
| 2026-04-21 14:48:31.519 | INFO | __main__:print_training_configuration:913 - Dual KL step weight : 0.1 |
| 2026-04-21 14:48:31.520 | INFO | __main__:print_training_configuration:914 - KL compute mode : rollout_phase |
| 2026-04-21 14:48:31.520 | INFO | __main__:print_training_configuration:915 - Use grad balancing : True |
| 2026-04-21 14:48:31.520 | INFO | __main__:print_training_configuration:916 - Enable timestep permutation : True |
| 2026-04-21 14:48:31.521 | INFO | __main__:print_training_configuration:917 - Reference mode offload : True |
| 2026-04-21 14:48:31.521 | INFO | __main__:print_training_configuration:919 - -------------------------------------------------------------------------------- |
| 2026-04-21 14:48:31.522 | INFO | __main__:print_training_configuration:920 - LOGGING & CHECKPOINTING CONFIGURATION |
| 2026-04-21 14:48:31.522 | INFO | __main__:print_training_configuration:921 - -------------------------------------------------------------------------------- |
| 2026-04-21 14:48:31.522 | INFO | __main__:print_training_configuration:922 - Output directory : ./outputs/debug |
| 2026-04-21 14:48:31.523 | INFO | __main__:print_training_configuration:923 - Checkpoints directory : s3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/debug/debug |
| 2026-04-21 14:48:31.523 | INFO | __main__:print_training_configuration:924 - Checkpoint interval : 100 |
| 2026-04-21 14:48:31.524 | INFO | __main__:print_training_configuration:925 - Log interval : 1 |
| 2026-04-21 14:48:31.524 | INFO | __main__:print_training_configuration:926 - Resume checkpoint : None |
| 2026-04-21 14:48:31.524 | INFO | __main__:print_training_configuration:927 - Train CSV : assets/demo_train.csv |
| 2026-04-21 14:48:31.525 | INFO | __main__:print_training_configuration:928 - Valid CSV : assets/demo_sample.csv |
| 2026-04-21 14:48:31.525 | INFO | __main__:print_training_configuration:929 - Seed : 42 |
| 2026-04-21 14:48:31.526 | INFO | __main__:print_training_configuration:930 - WandB enabled : True |
| 2026-04-21 14:48:31.526 | INFO | __main__:print_training_configuration:932 - WandB project : SAGE-GRPO |
| 2026-04-21 14:48:31.526 | INFO | __main__:print_training_configuration:933 - WandB run name : debug |
| 2026-04-21 14:48:31.527 | INFO | __main__:__init__:818 - [Effective Config] world_size=8, rank=0, local_rank=0, sp_size=2, dp_size=4, dp_rank=0, sp_rank=0, batch_size=2, num_generations=4, grad_accum=1 |
| 2026-04-21 14:48:31.527 | INFO | __main__:_build_video_dataset:1675 - [Effective GRPO Grouping] video_batch_size=2, dp_replicas=4, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, product=8 |
| 2026-04-21 14:48:31.832 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1171 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 |
| 2026-04-21 14:48:31.833 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1198 - Dataloader init done |
| 2026-04-21 14:48:31.833 | INFO | __main__:_train_grpo:1751 - Starting GRPO training... |
| 2026-04-21 14:48:31.834 | INFO | __main__:_train_grpo:1752 - Max steps: 10000 |
| 2026-04-21 14:48:31.834 | INFO | __main__:_train_grpo:1753 - Batch size: 2 |
| 2026-04-21 14:48:31.835 | INFO | __main__:_train_grpo:1754 - Learning rate: 1e-05 |
| 2026-04-21 14:48:33.054 | INFO | __main__:_train_grpo:1780 - [FLAG] post_train train path -> train_one_step |
| 2026-04-21 14:48:33.065 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:493 - [ParallelGroups] Rank 0, dp_rank 0: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=0, rank_in_group=0 |
| 2026-04-21 14:48:33.191 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:621 - Rank 0 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 |
| 2026-04-21 14:48:33.191 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:646 - Rank 0 using 'all' strategy (no mixgrpo): determistic=False for all timesteps |
| 2026-04-21 14:49:39.663 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:748 - Rank 0: Saved video 1/1 to ./outputs/debug/rl_samples/0000000/group_0/index_7542_seed_28576_dp_0_rank_0_sample_0_slot_0.mp4 |
| 2026-04-21 14:49:40.444 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:621 - Rank 0 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 |
| 2026-04-21 14:49:40.445 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:646 - Rank 0 using 'all' strategy (no mixgrpo): determistic=False for all timesteps |
| 2026-04-21 14:50:42.947 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:748 - Rank 0: Saved video 1/1 to ./outputs/debug/rl_samples/0000000/group_0/index_7542_seed_28577_dp_0_rank_0_sample_1_slot_1.mp4 |
| 2026-04-21 14:50:43.716 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:268 - [GatherRewards] Rank 0: Using local metric names without global sync: ['avg', 'videoalign_local_mq', 'videoalign_local_overall', 'videoalign_local_ta', 'videoalign_local_vq'] |
| 2026-04-21 14:50:44.423 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:336 - [GatherRewards] Rank 0 saved 4 entries for group 0 |
| 2026-04-21 14:50:46.033 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:847 - Rank 0: Pre-computing Reference Model statistics (Batch-wise)... |
| 2026-04-21 14:52:51.342 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:990 - Rank 0: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 125.30s (2.09min) |
| 2026-04-21 14:52:51.927 | INFO | hyvideo.utils.grpo_utils:train_one_step:1433 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] |
| 2026-04-21 14:52:54.892 | INFO | hyvideo.utils.grpo_utils:train_one_step:1545 - Rank 0, sp_rank 0: init adaptive KL controller (init=1.000e-07, min=1.000e-07, max=1.000e-05) |
| 2026-04-21 15:50:30.201 | INFO | __main__:__init__:806 - [RewardDebug] reward_weights(raw): {'VQ': 0.5, 'MQ': 0.5, 'TA': 1.0} |
| 2026-04-21 15:50:30.201 | INFO | __main__:__init__:807 - [RewardDebug] reward_config(effective): {'models': {'videoalign_local': {'weight': 1.0, 'sub_reward': {'VQ': 0.5, 'MQ': 0.5, 'TA': 1.0}}}, 'reward_checkpoint_mode': 'v3'} |
| 2026-04-21 15:50:32.742 | INFO | __main__:print_training_configuration:830 - SYSTEM & HARDWARE CONFIGURATION |
| 2026-04-21 15:50:32.743 | INFO | __main__:print_training_configuration:831 - -------------------------------------------------------------------------------- |
| 2026-04-21 15:50:32.743 | INFO | __main__:print_training_configuration:832 - Number of GPUs : 8 |
| 2026-04-21 15:50:32.744 | INFO | __main__:print_training_configuration:833 - Local rank : 0 |
| 2026-04-21 15:50:32.745 | INFO | __main__:print_training_configuration:834 - Global rank : 0 |
| 2026-04-21 15:50:32.745 | INFO | __main__:print_training_configuration:835 - Device : cuda:0 |
| 2026-04-21 15:50:32.745 | INFO | __main__:print_training_configuration:836 - Master weight dtype : torch.float32 |
| 2026-04-21 15:50:32.746 | INFO | __main__:print_training_configuration:838 - -------------------------------------------------------------------------------- |
| 2026-04-21 15:50:32.746 | INFO | __main__:print_training_configuration:839 - MODEL PARAMETERS & ARCHITECTURE |
| 2026-04-21 15:50:32.746 | INFO | __main__:print_training_configuration:840 - -------------------------------------------------------------------------------- |
| 2026-04-21 15:50:32.758 | INFO | __main__:print_training_configuration:843 - Number of total : 8,326,608,160 |
| 2026-04-21 15:50:32.759 | INFO | __main__:print_training_configuration:844 - Total trainable parameters : 8,326,608,160 |
| 2026-04-21 15:50:32.759 | INFO | __main__:print_training_configuration:845 - DIT model root : ./ckpts |
| 2026-04-21 15:50:32.761 | INFO | __main__:print_training_configuration:846 - Transformer version : 480p_t2v |
| 2026-04-21 15:50:32.761 | INFO | __main__:print_training_configuration:847 - Attention mode : flash |
| 2026-04-21 15:50:32.761 | INFO | __main__:print_training_configuration:848 - Gradient checkpointing : True |
| 2026-04-21 15:50:32.762 | INFO | __main__:print_training_configuration:849 - FSDP enabled : True |
| 2026-04-21 15:50:32.763 | INFO | __main__:print_training_configuration:850 - LoRA enabled : False |
| 2026-04-21 15:50:32.763 | INFO | __main__:print_training_configuration:851 - Glyph byT5 v2 : True |
| 2026-04-21 15:50:32.764 | INFO | __main__:print_training_configuration:852 - ByT5 max length : 256 |
| 2026-04-21 15:50:32.764 | INFO | __main__:print_training_configuration:854 - -------------------------------------------------------------------------------- |
| 2026-04-21 15:50:32.764 | INFO | __main__:print_training_configuration:855 - DISTRIBUTED TRAINING CONFIGURATION |
| 2026-04-21 15:50:32.765 | INFO | __main__:print_training_configuration:856 - -------------------------------------------------------------------------------- |
| 2026-04-21 15:50:32.765 | INFO | __main__:print_training_configuration:857 - World size : 8 |
| 2026-04-21 15:50:32.766 | INFO | __main__:print_training_configuration:858 - DP degree : 4 |
| 2026-04-21 15:50:32.766 | INFO | __main__:print_training_configuration:859 - DP rank : 0 |
| 2026-04-21 15:50:32.766 | INFO | __main__:print_training_configuration:860 - SP size : 2 |
| 2026-04-21 15:50:32.767 | INFO | __main__:print_training_configuration:861 - SP rank : 0 |
| 2026-04-21 15:50:32.767 | INFO | __main__:print_training_configuration:863 - -------------------------------------------------------------------------------- |
| 2026-04-21 15:50:32.767 | INFO | __main__:print_training_configuration:864 - BATCH SIZE & DATA FLOW CONFIGURATION |
| 2026-04-21 15:50:32.767 | INFO | __main__:print_training_configuration:865 - -------------------------------------------------------------------------------- |
| 2026-04-21 15:50:32.768 | INFO | __main__:print_training_configuration:866 - Per GPU batch size : 2 |
| 2026-04-21 15:50:32.768 | INFO | __main__:print_training_configuration:867 - DP degree batch size : 8 |
| 2026-04-21 15:50:32.769 | INFO | __main__:print_training_configuration:868 - Gradient accumulation steps : 1 |
| 2026-04-21 15:50:32.769 | INFO | __main__:print_training_configuration:869 - Num generations : 4 |
| 2026-04-21 15:50:32.770 | INFO | __main__:print_training_configuration:870 - Use same noise : False |
| 2026-04-21 15:50:32.770 | INFO | __main__:print_training_configuration:871 - Num workers : 4 |
| 2026-04-21 15:50:32.771 | INFO | __main__:print_training_configuration:872 - Prefetch factor : 2 |
| 2026-04-21 15:50:32.771 | INFO | __main__:print_training_configuration:874 - -------------------------------------------------------------------------------- |
| 2026-04-21 15:50:32.772 | INFO | __main__:print_training_configuration:875 - OPTIMIZATION CONFIGURATION |
| 2026-04-21 15:50:32.772 | INFO | __main__:print_training_configuration:876 - -------------------------------------------------------------------------------- |
| 2026-04-21 15:50:32.772 | INFO | __main__:print_training_configuration:877 - Optimizer : muon |
| 2026-04-21 15:50:32.772 | INFO | __main__:print_training_configuration:878 - Learning rate : 1e-05 |
| 2026-04-21 15:50:32.773 | INFO | __main__:print_training_configuration:879 - Weight decay : 0.01 |
| 2026-04-21 15:50:32.773 | INFO | __main__:print_training_configuration:880 - Max gradient norm : 1.0 |
| 2026-04-21 15:50:32.774 | INFO | __main__:print_training_configuration:881 - LR scheduler : constant |
| 2026-04-21 15:50:32.774 | INFO | __main__:print_training_configuration:882 - Warmup steps : 500 |
| 2026-04-21 15:50:32.774 | INFO | __main__:print_training_configuration:883 - Max training steps : 10000 |
| 2026-04-21 15:50:32.775 | INFO | __main__:print_training_configuration:885 - -------------------------------------------------------------------------------- |
| 2026-04-21 15:50:32.775 | INFO | __main__:print_training_configuration:886 - INFERENCE & VALIDATION CONFIGURATION |
| 2026-04-21 15:50:32.775 | INFO | __main__:print_training_configuration:887 - -------------------------------------------------------------------------------- |
| 2026-04-21 15:50:32.776 | INFO | __main__:print_training_configuration:888 - Validation interval : 100 |
| 2026-04-21 15:50:32.776 | INFO | __main__:print_training_configuration:889 - Validate at step 0 : False |
| 2026-04-21 15:50:32.777 | INFO | __main__:print_training_configuration:890 - Validation video length : 121 |
| 2026-04-21 15:50:32.777 | INFO | __main__:print_training_configuration:891 - Validation global seed : 930 |
| 2026-04-21 15:50:32.778 | INFO | __main__:print_training_configuration:892 - Validation inference steps : 40 |
| 2026-04-21 15:50:32.778 | INFO | __main__:print_training_configuration:893 - Validation aspect ratio : 16:9 |
| 2026-04-21 15:50:32.778 | INFO | __main__:print_training_configuration:894 - Validation fixed size : 480x864 |
| 2026-04-21 15:50:32.779 | INFO | __main__:print_training_configuration:895 - Validation prompt rewrite : False |
| 2026-04-21 15:50:32.779 | INFO | __main__:print_training_configuration:896 - Validation flow shift : 5.0 |
| 2026-04-21 15:50:32.779 | INFO | __main__:print_training_configuration:897 - Sampling CFG scale : 6.0 |
| 2026-04-21 15:50:32.780 | INFO | __main__:print_training_configuration:898 - Sampling eta : 0.5 |
| 2026-04-21 15:50:32.780 | INFO | __main__:print_training_configuration:899 - Sampling steps (GRPO rollout) : 20 |
| 2026-04-21 15:50:32.780 | INFO | __main__:print_training_configuration:900 - SDE type : sage_grpo |
| 2026-04-21 15:50:32.781 | INFO | __main__:print_training_configuration:902 - -------------------------------------------------------------------------------- |
| 2026-04-21 15:50:32.781 | INFO | __main__:print_training_configuration:903 - GRPO & KL CONFIGURATION |
| 2026-04-21 15:50:32.782 | INFO | __main__:print_training_configuration:904 - -------------------------------------------------------------------------------- |
| 2026-04-21 15:50:32.782 | INFO | __main__:print_training_configuration:905 - Reward model : videoalign_local |
| 2026-04-21 15:50:32.782 | INFO | __main__:print_training_configuration:906 - Reward checkpoint mode : v3 |
| 2026-04-21 15:50:32.783 | INFO | __main__:print_training_configuration:907 - KL weight : 1e-05 |
| 2026-04-21 15:50:32.783 | INFO | __main__:print_training_configuration:908 - KL coef : 1e-07 |
| 2026-04-21 15:50:32.784 | INFO | __main__:print_training_configuration:909 - Use moving KL : True |
| 2026-04-21 15:50:32.784 | INFO | __main__:print_training_configuration:910 - Update ref model step : 10 |
| 2026-04-21 15:50:32.785 | INFO | __main__:print_training_configuration:911 - Use dual KL : True |
| 2026-04-21 15:50:32.785 | INFO | __main__:print_training_configuration:912 - Dual KL moving weight : 1.0 |
| 2026-04-21 15:50:32.786 | INFO | __main__:print_training_configuration:913 - Dual KL step weight : 0.1 |
| 2026-04-21 15:50:32.786 | INFO | __main__:print_training_configuration:914 - KL compute mode : rollout_phase |
| 2026-04-21 15:50:32.786 | INFO | __main__:print_training_configuration:915 - Use grad balancing : True |
| 2026-04-21 15:50:32.787 | INFO | __main__:print_training_configuration:916 - Enable timestep permutation : True |
| 2026-04-21 15:50:32.787 | INFO | __main__:print_training_configuration:917 - Reference mode offload : True |
| 2026-04-21 15:50:32.787 | INFO | __main__:print_training_configuration:919 - -------------------------------------------------------------------------------- |
| 2026-04-21 15:50:32.788 | INFO | __main__:print_training_configuration:920 - LOGGING & CHECKPOINTING CONFIGURATION |
| 2026-04-21 15:50:32.788 | INFO | __main__:print_training_configuration:921 - -------------------------------------------------------------------------------- |
| 2026-04-21 15:50:32.789 | INFO | __main__:print_training_configuration:922 - Output directory : ./outputs/debug |
| 2026-04-21 15:50:32.789 | INFO | __main__:print_training_configuration:923 - Checkpoints directory : s3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/debug/debug |
| 2026-04-21 15:50:32.789 | INFO | __main__:print_training_configuration:924 - Checkpoint interval : 100 |
| 2026-04-21 15:50:32.790 | INFO | __main__:print_training_configuration:925 - Log interval : 1 |
| 2026-04-21 15:50:32.790 | INFO | __main__:print_training_configuration:926 - Resume checkpoint : None |
| 2026-04-21 15:50:32.791 | INFO | __main__:print_training_configuration:927 - Train CSV : assets/demo_train.csv |
| 2026-04-21 15:50:32.791 | INFO | __main__:print_training_configuration:928 - Valid CSV : assets/demo_sample.csv |
| 2026-04-21 15:50:32.792 | INFO | __main__:print_training_configuration:929 - Seed : 42 |
| 2026-04-21 15:50:32.792 | INFO | __main__:print_training_configuration:930 - WandB enabled : True |
| 2026-04-21 15:50:32.792 | INFO | __main__:print_training_configuration:932 - WandB project : SAGE-GRPO |
| 2026-04-21 15:50:32.793 | INFO | __main__:print_training_configuration:933 - WandB run name : debug |
| 2026-04-21 15:50:32.793 | INFO | __main__:__init__:818 - [Effective Config] world_size=8, rank=0, local_rank=0, sp_size=2, dp_size=4, dp_rank=0, sp_rank=0, batch_size=2, num_generations=4, grad_accum=1 |
| 2026-04-21 15:50:32.794 | INFO | __main__:_build_video_dataset:1675 - [Effective GRPO Grouping] video_batch_size=2, dp_replicas=4, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, product=8 |
| 2026-04-21 15:50:33.106 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1171 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 |
| 2026-04-21 15:50:33.107 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1198 - Dataloader init done |
| 2026-04-21 15:50:33.107 | INFO | __main__:_train_grpo:1751 - Starting GRPO training... |
| 2026-04-21 15:50:33.108 | INFO | __main__:_train_grpo:1752 - Max steps: 10000 |
| 2026-04-21 15:50:33.108 | INFO | __main__:_train_grpo:1753 - Batch size: 2 |
| 2026-04-21 15:50:33.109 | INFO | __main__:_train_grpo:1754 - Learning rate: 1e-05 |
| 2026-04-21 15:50:34.529 | INFO | __main__:_train_grpo:1780 - [FLAG] post_train train path -> train_one_step |
| 2026-04-21 15:50:34.539 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:493 - [ParallelGroups] Rank 0, dp_rank 0: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=0, rank_in_group=0 |
| 2026-04-21 15:50:34.662 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:621 - Rank 0 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 |
| 2026-04-21 15:50:34.663 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:646 - Rank 0 using 'all' strategy (no mixgrpo): determistic=False for all timesteps |
| 2026-04-21 15:51:41.462 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:748 - Rank 0: Saved video 1/1 to ./outputs/debug/rl_samples/0000000/group_0/index_7542_seed_28576_dp_0_rank_0_sample_0_slot_0.mp4 |
| 2026-04-21 15:51:42.269 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:621 - Rank 0 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 |
| 2026-04-21 15:51:42.269 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:646 - Rank 0 using 'all' strategy (no mixgrpo): determistic=False for all timesteps |
| 2026-04-21 15:52:44.313 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:748 - Rank 0: Saved video 1/1 to ./outputs/debug/rl_samples/0000000/group_0/index_7542_seed_28577_dp_0_rank_0_sample_1_slot_1.mp4 |
| 2026-04-21 15:52:45.050 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:268 - [GatherRewards] Rank 0: Using local metric names without global sync: ['avg', 'videoalign_local_mq', 'videoalign_local_overall', 'videoalign_local_ta', 'videoalign_local_vq'] |
| 2026-04-21 15:52:45.766 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:336 - [GatherRewards] Rank 0 saved 4 entries for group 0 |
| 2026-04-21 15:52:47.513 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:847 - Rank 0: Pre-computing Reference Model statistics (Batch-wise)... |
| 2026-04-21 15:54:51.688 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:990 - Rank 0: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 124.14s (2.07min) |
| 2026-04-21 15:54:52.258 | INFO | hyvideo.utils.grpo_utils:train_one_step:1433 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] |
| 2026-04-21 15:54:55.172 | INFO | hyvideo.utils.grpo_utils:train_one_step:1545 - Rank 0, sp_rank 0: init adaptive KL controller (init=1.000e-07, min=1.000e-07, max=1.000e-05) |
| 2026-04-21 16:01:50.683 | INFO | hyvideo.utils.grpo_utils:train_one_step:1774 - Rank 0, sp_rank 0: update step: 1, train_steps: 40 |
| 2026-04-21 16:01:51.130 | INFO | __main__:_log_precision_diagnostics_once:1372 - [PrecisionDiag] trainable param dtypes: ['torch.float32'] |
| 2026-04-21 16:01:51.130 | INFO | __main__:_log_precision_diagnostics_once:1375 - [PrecisionDiag] Using composable FSDP2 mixed precision (param_dtype=bf16, reduce_dtype=float32). This path does not maintain an explicit FP32 master-weight copy for parameters. |
| 2026-04-21 16:01:51.133 | INFO | __main__:_log_precision_diagnostics_once:1388 - [PrecisionDiag] optimizer tensor-state dtypes: ['torch.float32'] (count=2804) |
| 2026-04-21 16:01:51.134 | INFO | __main__:_log_progress_step:1734 - Training Progress: update_steps 1 train_steps 40 {'step': '1/10000', 'learning_rate': '0.00001000', 'loss': '-2.682e-07', 'grad_norm': '0.01066162', 'KL_weight': '0.00000396', 'param_delta_mean': '0.00000000', 'param_delta_max': '0.00000000', 'rollout_time': '257.61416626', 'grpo_loop_time': '418.37072754', 'ref_model_time': '124.14012909', 'approx_kl': '1.479e-15', 'clipfrac': '0.00000000', 'clipfrac_gt_one': '0.00000000', 'clipfrac_lt_one': '0.00000000', 'policy_loss': '-2.906e-07', 'kl_loss': '0.00863841', 'kl_moving': '0.00863842', 'kl_stepwise': '-1.902e-08', 'gathered_avg_reward_mean': '-1.07328284', 'gathered_avg_reward_std': '0.44008070', 'gathered_videoalign_local_mq_reward_mean': '-0.93518472', 'gathered_videoalign_local_mq_reward_std': '0.22331762', 'gathered_videoalign_local_overall_reward_mean': '-2.16913152', 'gathered_videoalign_local_overall_reward_std': '0.58840650', 'gathered_videoalign_local_ta_reward_mean': '0.02256578', 'gathered_videoalign_local_ta_reward_std': '0.30612063', 'gathered_videoalign_local_vq_reward_mean': '-1.25651264', 'gathered_videoalign_local_vq_reward_std': '0.17318933', 'gathered_ori_avg_reward_mean': '-1.07328284', 'gathered_ori_avg_reward_std': '0.44008070'} |
| 2026-04-21 16:01:51.172 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:493 - [ParallelGroups] Rank 0, dp_rank 0: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=0, rank_in_group=0 |
| 2026-04-21 16:01:51.294 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:621 - Rank 0 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 |
| 2026-04-21 16:01:51.295 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:646 - Rank 0 using 'all' strategy (no mixgrpo): determistic=False for all timesteps |
| 2026-04-21 16:02:54.542 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:748 - Rank 0: Saved video 1/1 to ./outputs/debug/rl_samples/0000001/group_0/index_3698_seed_1481_dp_0_rank_0_sample_0_slot_0.mp4 |
| 2026-04-21 16:02:55.155 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:621 - Rank 0 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 |
| 2026-04-21 16:02:55.156 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:646 - Rank 0 using 'all' strategy (no mixgrpo): determistic=False for all timesteps |
| 2026-04-21 16:03:57.266 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:748 - Rank 0: Saved video 1/1 to ./outputs/debug/rl_samples/0000001/group_0/index_3698_seed_1482_dp_0_rank_0_sample_1_slot_1.mp4 |
| 2026-04-21 16:03:57.931 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:268 - [GatherRewards] Rank 0: Using local metric names without global sync: ['avg', 'videoalign_local_mq', 'videoalign_local_overall', 'videoalign_local_ta', 'videoalign_local_vq'] |
| 2026-04-21 16:03:58.070 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:336 - [GatherRewards] Rank 0 saved 4 entries for group 0 |
| 2026-04-21 16:03:58.832 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:847 - Rank 0: Pre-computing Reference Model statistics (Batch-wise)... |
| 2026-04-21 16:06:01.982 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:990 - Rank 0: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 123.14s (2.05min) |
| 2026-04-21 16:06:02.536 | INFO | hyvideo.utils.grpo_utils:train_one_step:1433 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] |
| 2026-04-21 16:12:56.744 | INFO | hyvideo.utils.grpo_utils:train_one_step:1774 - Rank 0, sp_rank 0: update step: 2, train_steps: 80 |
| 2026-04-21 16:12:57.326 | INFO | __main__:_log_progress_step:1734 - Training Progress: update_steps 2 train_steps 80 {'step': '2/10000', 'learning_rate': '0.00001000', 'loss': '-1.788e-07', 'grad_norm': '0.01083679', 'KL_weight': '0.00000792', 'param_delta_mean': '0.00000000', 'param_delta_max': '0.00000000', 'rollout_time': '251.31460571', 'grpo_loop_time': '414.25421143', 'ref_model_time': '123.14079285', 'approx_kl': '1.560e-15', 'clipfrac': '0.00000000', 'clipfrac_gt_one': '0.00000000', 'clipfrac_lt_one': '0.00000000', 'policy_loss': '-2.813e-07', 'kl_loss': '0.00744735', 'kl_moving': '0.00744735', 'kl_stepwise': '-1.874e-08', 'gathered_avg_reward_mean': '-0.57466680', 'gathered_avg_reward_std': '1.04985189', 'gathered_videoalign_local_mq_reward_mean': '-0.57400572', 'gathered_videoalign_local_mq_reward_std': '0.33090693', 'gathered_videoalign_local_overall_reward_mean': '-1.30037904', 'gathered_videoalign_local_overall_reward_std': '1.27385473', 'gathered_videoalign_local_ta_reward_mean': '0.15104552', 'gathered_videoalign_local_ta_reward_std': '0.84300774', 'gathered_videoalign_local_vq_reward_mean': '-0.87741888', 'gathered_videoalign_local_vq_reward_std': '0.20758396', 'gathered_ori_avg_reward_mean': '-0.57466680', 'gathered_ori_avg_reward_std': '1.04985189'} |
| 2026-04-21 16:12:57.372 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:493 - [ParallelGroups] Rank 0, dp_rank 0: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=0, rank_in_group=0 |
| 2026-04-21 16:12:57.462 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:621 - Rank 0 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 |
| 2026-04-21 16:12:57.462 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:646 - Rank 0 using 'all' strategy (no mixgrpo): determistic=False for all timesteps |
| 2026-04-21 16:14:00.754 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:748 - Rank 0: Saved video 1/1 to ./outputs/debug/rl_samples/0000002/group_0/index_8086_seed_8771_dp_0_rank_0_sample_0_slot_0.mp4 |
| 2026-04-21 16:14:01.503 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:621 - Rank 0 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 |
| 2026-04-21 16:14:01.504 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:646 - Rank 0 using 'all' strategy (no mixgrpo): determistic=False for all timesteps |
| 2026-04-21 16:15:03.529 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:748 - Rank 0: Saved video 1/1 to ./outputs/debug/rl_samples/0000002/group_0/index_8086_seed_8772_dp_0_rank_0_sample_1_slot_1.mp4 |
| 2026-04-21 16:15:04.198 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:268 - [GatherRewards] Rank 0: Using local metric names without global sync: ['avg', 'videoalign_local_mq', 'videoalign_local_overall', 'videoalign_local_ta', 'videoalign_local_vq'] |
| 2026-04-21 16:15:04.350 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:336 - [GatherRewards] Rank 0 saved 4 entries for group 0 |
| 2026-04-21 16:15:05.072 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:847 - Rank 0: Pre-computing Reference Model statistics (Batch-wise)... |
| 2026-04-21 16:17:05.987 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:990 - Rank 0: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 120.81s (2.01min) |
| 2026-04-21 16:17:06.550 | INFO | hyvideo.utils.grpo_utils:train_one_step:1433 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] |
| 2026-04-21 16:24:00.783 | INFO | hyvideo.utils.grpo_utils:train_one_step:1774 - Rank 0, sp_rank 0: update step: 3, train_steps: 120 |
| 2026-04-21 16:24:01.483 | INFO | __main__:_log_progress_step:1734 - Training Progress: update_steps 3 train_steps 120 {'step': '3/10000', 'learning_rate': '0.00001000', 'loss': '-2.086e-07', 'grad_norm': '0.01063384', 'KL_weight': '0.00000120', 'param_delta_mean': '0.00000000', 'param_delta_max': '0.00000000', 'rollout_time': '249.13354492', 'grpo_loop_time': '414.43972778', 'ref_model_time': '120.80517578', 'approx_kl': '1.426e-15', 'clipfrac': '0.00000000', 'clipfrac_gt_one': '0.00000000', 'clipfrac_lt_one': '0.00000000', 'policy_loss': '-2.682e-07', 'kl_loss': '0.00618454', 'kl_moving': '0.00618454', 'kl_stepwise': '-2.075e-08', 'gathered_avg_reward_mean': '-1.57800758', 'gathered_avg_reward_std': '0.19400825', 'gathered_videoalign_local_mq_reward_mean': '-1.01983666', 'gathered_videoalign_local_mq_reward_std': '0.10223494', 'gathered_videoalign_local_overall_reward_mean': '-2.86369872', 'gathered_videoalign_local_overall_reward_std': '0.25308815', 'gathered_videoalign_local_ta_reward_mean': '-0.29231644', 'gathered_videoalign_local_ta_reward_std': '0.16989647', 'gathered_videoalign_local_vq_reward_mean': '-1.55154562', 'gathered_videoalign_local_vq_reward_std': '0.14786871', 'gathered_ori_avg_reward_mean': '-1.57800758', 'gathered_ori_avg_reward_std': '0.19400822'} |
| 2026-04-21 16:24:01.524 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:493 - [ParallelGroups] Rank 0, dp_rank 0: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=0, rank_in_group=0 |
| 2026-04-21 16:24:01.618 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:621 - Rank 0 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 |
| 2026-04-21 16:24:01.618 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:646 - Rank 0 using 'all' strategy (no mixgrpo): determistic=False for all timesteps |
| 2026-04-21 16:25:04.684 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:748 - Rank 0: Saved video 1/1 to ./outputs/debug/rl_samples/0000003/group_0/index_2282_seed_5648_dp_0_rank_0_sample_0_slot_0.mp4 |
| 2026-04-21 16:25:05.254 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:621 - Rank 0 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 |
| 2026-04-21 16:25:05.255 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:646 - Rank 0 using 'all' strategy (no mixgrpo): determistic=False for all timesteps |
| 2026-04-21 16:26:07.572 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:748 - Rank 0: Saved video 1/1 to ./outputs/debug/rl_samples/0000003/group_0/index_2282_seed_5649_dp_0_rank_0_sample_1_slot_1.mp4 |
| 2026-04-21 16:26:08.344 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:268 - [GatherRewards] Rank 0: Using local metric names without global sync: ['avg', 'videoalign_local_mq', 'videoalign_local_overall', 'videoalign_local_ta', 'videoalign_local_vq'] |
| 2026-04-21 16:26:08.500 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:336 - [GatherRewards] Rank 0 saved 4 entries for group 0 |
| 2026-04-21 16:26:09.275 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:847 - Rank 0: Pre-computing Reference Model statistics (Batch-wise)... |
| 2026-04-21 16:28:10.850 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:990 - Rank 0: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 121.50s (2.03min) |
| 2026-04-21 16:28:11.408 | INFO | hyvideo.utils.grpo_utils:train_one_step:1433 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] |
| 2026-04-21 16:35:05.697 | INFO | hyvideo.utils.grpo_utils:train_one_step:1774 - Rank 0, sp_rank 0: update step: 4, train_steps: 160 |
| 2026-04-21 16:35:06.210 | INFO | __main__:_log_progress_step:1734 - Training Progress: update_steps 4 train_steps 160 {'step': '4/10000', 'learning_rate': '0.00001000', 'loss': '-9.686e-08', 'grad_norm': '0.00794416', 'KL_weight': '1.000e-07', 'param_delta_mean': '0.00000000', 'param_delta_max': '0.00000000', 'rollout_time': '249.84045410', 'grpo_loop_time': '414.29733276', 'ref_model_time': '121.50069427', 'approx_kl': '1.450e-15', 'clipfrac': '0.00000000', 'clipfrac_gt_one': '0.00000000', 'clipfrac_lt_one': '0.00000000', 'policy_loss': '-9.686e-08', 'kl_loss': '0.00632224', 'kl_moving': '0.00632224', 'kl_stepwise': '-1.881e-08', 'gathered_avg_reward_mean': '-0.67030287', 'gathered_avg_reward_std': '0.92555404', 'gathered_videoalign_local_mq_reward_mean': '-0.70078433', 'gathered_videoalign_local_mq_reward_std': '0.37624747', 'gathered_videoalign_local_overall_reward_mean': '-1.50739539', 'gathered_videoalign_local_overall_reward_std': '1.35916543', 'gathered_videoalign_local_ta_reward_mean': '0.16678962', 'gathered_videoalign_local_ta_reward_std': '0.49920946', 'gathered_videoalign_local_vq_reward_mean': '-0.97340071', 'gathered_videoalign_local_vq_reward_std': '0.53128082', 'gathered_ori_avg_reward_mean': '-0.67030287', 'gathered_ori_avg_reward_std': '0.92555410'} |
| 2026-04-21 16:35:06.253 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:493 - [ParallelGroups] Rank 0, dp_rank 0: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=0, rank_in_group=0 |
| 2026-04-21 16:35:06.345 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:621 - Rank 0 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 |
| 2026-04-21 16:35:06.346 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:646 - Rank 0 using 'all' strategy (no mixgrpo): determistic=False for all timesteps |
| 2026-04-21 16:36:09.725 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:748 - Rank 0: Saved video 1/1 to ./outputs/debug/rl_samples/0000004/group_0/index_1462_seed_13994_dp_0_rank_0_sample_0_slot_0.mp4 |
| 2026-04-21 16:36:10.768 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:621 - Rank 0 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 |
| 2026-04-21 16:36:10.769 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:646 - Rank 0 using 'all' strategy (no mixgrpo): determistic=False for all timesteps |
| 2026-04-21 16:37:12.957 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:748 - Rank 0: Saved video 1/1 to ./outputs/debug/rl_samples/0000004/group_0/index_1462_seed_13995_dp_0_rank_0_sample_1_slot_1.mp4 |
| 2026-04-21 16:37:13.744 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:268 - [GatherRewards] Rank 0: Using local metric names without global sync: ['avg', 'videoalign_local_mq', 'videoalign_local_overall', 'videoalign_local_ta', 'videoalign_local_vq'] |
| 2026-04-21 16:37:13.896 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:336 - [GatherRewards] Rank 0 saved 4 entries for group 0 |
| 2026-04-21 16:37:14.715 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:847 - Rank 0: Pre-computing Reference Model statistics (Batch-wise)... |
| 2026-04-21 16:39:16.189 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:990 - Rank 0: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 121.46s (2.02min) |
| 2026-04-21 16:39:16.746 | INFO | hyvideo.utils.grpo_utils:train_one_step:1433 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] |
| 2026-04-22 04:28:29.335 | INFO | __main__:__init__:806 - [RewardDebug] reward_weights(raw): {'VQ': 0.5, 'MQ': 0.5, 'TA': 1.0} |
| 2026-04-22 04:28:29.335 | INFO | __main__:__init__:807 - [RewardDebug] reward_config(effective): {'models': {'videoalign_local': {'weight': 1.0, 'sub_reward': {'VQ': 0.5, 'MQ': 0.5, 'TA': 1.0}}}, 'reward_checkpoint_mode': 'v3'} |
| 2026-04-22 04:28:34.139 | INFO | __main__:print_training_configuration:830 - SYSTEM & HARDWARE CONFIGURATION |
| 2026-04-22 04:28:34.140 | INFO | __main__:print_training_configuration:831 - -------------------------------------------------------------------------------- |
| 2026-04-22 04:28:34.140 | INFO | __main__:print_training_configuration:832 - Number of GPUs : 8 |
| 2026-04-22 04:28:34.140 | INFO | __main__:print_training_configuration:833 - Local rank : 0 |
| 2026-04-22 04:28:34.141 | INFO | __main__:print_training_configuration:834 - Global rank : 0 |
| 2026-04-22 04:28:34.142 | INFO | __main__:print_training_configuration:835 - Device : cuda:0 |
| 2026-04-22 04:28:34.142 | INFO | __main__:print_training_configuration:836 - Master weight dtype : torch.float32 |
| 2026-04-22 04:28:34.142 | INFO | __main__:print_training_configuration:838 - -------------------------------------------------------------------------------- |
| 2026-04-22 04:28:34.143 | INFO | __main__:print_training_configuration:839 - MODEL PARAMETERS & ARCHITECTURE |
| 2026-04-22 04:28:34.143 | INFO | __main__:print_training_configuration:840 - -------------------------------------------------------------------------------- |
| 2026-04-22 04:28:34.155 | INFO | __main__:print_training_configuration:843 - Number of total : 8,326,608,160 |
| 2026-04-22 04:28:34.155 | INFO | __main__:print_training_configuration:844 - Total trainable parameters : 8,326,608,160 |
| 2026-04-22 04:28:34.156 | INFO | __main__:print_training_configuration:845 - DIT model root : ./ckpts |
| 2026-04-22 04:28:34.156 | INFO | __main__:print_training_configuration:846 - Transformer version : 480p_t2v |
| 2026-04-22 04:28:34.157 | INFO | __main__:print_training_configuration:847 - Attention mode : flash |
| 2026-04-22 04:28:34.157 | INFO | __main__:print_training_configuration:848 - Gradient checkpointing : True |
| 2026-04-22 04:28:34.158 | INFO | __main__:print_training_configuration:849 - FSDP enabled : True |
| 2026-04-22 04:28:34.158 | INFO | __main__:print_training_configuration:850 - LoRA enabled : False |
| 2026-04-22 04:28:34.158 | INFO | __main__:print_training_configuration:851 - Glyph byT5 v2 : True |
| 2026-04-22 04:28:34.159 | INFO | __main__:print_training_configuration:852 - ByT5 max length : 256 |
| 2026-04-22 04:28:34.159 | INFO | __main__:print_training_configuration:854 - -------------------------------------------------------------------------------- |
| 2026-04-22 04:28:34.159 | INFO | __main__:print_training_configuration:855 - DISTRIBUTED TRAINING CONFIGURATION |
| 2026-04-22 04:28:34.160 | INFO | __main__:print_training_configuration:856 - -------------------------------------------------------------------------------- |
| 2026-04-22 04:28:34.160 | INFO | __main__:print_training_configuration:857 - World size : 8 |
| 2026-04-22 04:28:34.160 | INFO | __main__:print_training_configuration:858 - DP degree : 4 |
| 2026-04-22 04:28:34.161 | INFO | __main__:print_training_configuration:859 - DP rank : 0 |
| 2026-04-22 04:28:34.161 | INFO | __main__:print_training_configuration:860 - SP size : 2 |
| 2026-04-22 04:28:34.161 | INFO | __main__:print_training_configuration:861 - SP rank : 0 |
| 2026-04-22 04:28:34.161 | INFO | __main__:print_training_configuration:863 - -------------------------------------------------------------------------------- |
| 2026-04-22 04:28:34.162 | INFO | __main__:print_training_configuration:864 - BATCH SIZE & DATA FLOW CONFIGURATION |
| 2026-04-22 04:28:34.162 | INFO | __main__:print_training_configuration:865 - -------------------------------------------------------------------------------- |
| 2026-04-22 04:28:34.163 | INFO | __main__:print_training_configuration:866 - Per GPU batch size : 2 |
| 2026-04-22 04:28:34.163 | INFO | __main__:print_training_configuration:867 - DP degree batch size : 8 |
| 2026-04-22 04:28:34.163 | INFO | __main__:print_training_configuration:868 - Gradient accumulation steps : 1 |
| 2026-04-22 04:28:34.164 | INFO | __main__:print_training_configuration:869 - Num generations : 4 |
| 2026-04-22 04:28:34.164 | INFO | __main__:print_training_configuration:870 - Use same noise : False |
| 2026-04-22 04:28:34.165 | INFO | __main__:print_training_configuration:871 - Num workers : 4 |
| 2026-04-22 04:28:34.165 | INFO | __main__:print_training_configuration:872 - Prefetch factor : 2 |
| 2026-04-22 04:28:34.165 | INFO | __main__:print_training_configuration:874 - -------------------------------------------------------------------------------- |
| 2026-04-22 04:28:34.166 | INFO | __main__:print_training_configuration:875 - OPTIMIZATION CONFIGURATION |
| 2026-04-22 04:28:34.166 | INFO | __main__:print_training_configuration:876 - -------------------------------------------------------------------------------- |
| 2026-04-22 04:28:34.166 | INFO | __main__:print_training_configuration:877 - Optimizer : muon |
| 2026-04-22 04:28:34.167 | INFO | __main__:print_training_configuration:878 - Learning rate : 1e-05 |
| 2026-04-22 04:28:34.167 | INFO | __main__:print_training_configuration:879 - Weight decay : 0.01 |
| 2026-04-22 04:28:34.168 | INFO | __main__:print_training_configuration:880 - Max gradient norm : 1.0 |
| 2026-04-22 04:28:34.168 | INFO | __main__:print_training_configuration:881 - LR scheduler : constant |
| 2026-04-22 04:28:34.168 | INFO | __main__:print_training_configuration:882 - Warmup steps : 500 |
| 2026-04-22 04:28:34.169 | INFO | __main__:print_training_configuration:883 - Max training steps : 10000 |
| 2026-04-22 04:28:34.169 | INFO | __main__:print_training_configuration:885 - -------------------------------------------------------------------------------- |
| 2026-04-22 04:28:34.169 | INFO | __main__:print_training_configuration:886 - INFERENCE & VALIDATION CONFIGURATION |
| 2026-04-22 04:28:34.170 | INFO | __main__:print_training_configuration:887 - -------------------------------------------------------------------------------- |
| 2026-04-22 04:28:34.170 | INFO | __main__:print_training_configuration:888 - Validation interval : 100 |
| 2026-04-22 04:28:34.171 | INFO | __main__:print_training_configuration:889 - Validate at step 0 : False |
| 2026-04-22 04:28:34.171 | INFO | __main__:print_training_configuration:890 - Validation video length : 121 |
| 2026-04-22 04:28:34.171 | INFO | __main__:print_training_configuration:891 - Validation global seed : 930 |
| 2026-04-22 04:28:34.171 | INFO | __main__:print_training_configuration:892 - Validation inference steps : 40 |
| 2026-04-22 04:28:34.172 | INFO | __main__:print_training_configuration:893 - Validation aspect ratio : 16:9 |
| 2026-04-22 04:28:34.172 | INFO | __main__:print_training_configuration:894 - Validation fixed size : 480x864 |
| 2026-04-22 04:28:34.173 | INFO | __main__:print_training_configuration:895 - Validation prompt rewrite : False |
| 2026-04-22 04:28:34.173 | INFO | __main__:print_training_configuration:896 - Validation flow shift : 5.0 |
| 2026-04-22 04:28:34.174 | INFO | __main__:print_training_configuration:897 - Sampling CFG scale : 6.0 |
| 2026-04-22 04:28:34.174 | INFO | __main__:print_training_configuration:898 - Sampling eta : 0.5 |
| 2026-04-22 04:28:34.174 | INFO | __main__:print_training_configuration:899 - Sampling steps (GRPO rollout) : 20 |
| 2026-04-22 04:28:34.175 | INFO | __main__:print_training_configuration:900 - SDE type : sage_grpo |
| 2026-04-22 04:28:34.175 | INFO | __main__:print_training_configuration:902 - -------------------------------------------------------------------------------- |
| 2026-04-22 04:28:34.176 | INFO | __main__:print_training_configuration:903 - GRPO & KL CONFIGURATION |
| 2026-04-22 04:28:34.176 | INFO | __main__:print_training_configuration:904 - -------------------------------------------------------------------------------- |
| 2026-04-22 04:28:34.176 | INFO | __main__:print_training_configuration:905 - Reward model : videoalign_local |
| 2026-04-22 04:28:34.177 | INFO | __main__:print_training_configuration:906 - Reward checkpoint mode : v3 |
| 2026-04-22 04:28:34.177 | INFO | __main__:print_training_configuration:907 - KL weight : 1e-05 |
| 2026-04-22 04:28:34.178 | INFO | __main__:print_training_configuration:908 - KL coef : 1e-07 |
| 2026-04-22 04:28:34.178 | INFO | __main__:print_training_configuration:909 - Use moving KL : True |
| 2026-04-22 04:28:34.178 | INFO | __main__:print_training_configuration:910 - Update ref model step : 10 |
| 2026-04-22 04:28:34.178 | INFO | __main__:print_training_configuration:911 - Use dual KL : True |
| 2026-04-22 04:28:34.178 | INFO | __main__:print_training_configuration:912 - Dual KL moving weight : 1.0 |
| 2026-04-22 04:28:34.179 | INFO | __main__:print_training_configuration:913 - Dual KL step weight : 0.1 |
| 2026-04-22 04:28:34.180 | INFO | __main__:print_training_configuration:914 - KL compute mode : rollout_phase |
| 2026-04-22 04:28:34.180 | INFO | __main__:print_training_configuration:915 - Use grad balancing : True |
| 2026-04-22 04:28:34.180 | INFO | __main__:print_training_configuration:916 - Enable timestep permutation : True |
| 2026-04-22 04:28:34.180 | INFO | __main__:print_training_configuration:917 - Reference mode offload : True |
| 2026-04-22 04:28:34.181 | INFO | __main__:print_training_configuration:919 - -------------------------------------------------------------------------------- |
| 2026-04-22 04:28:34.181 | INFO | __main__:print_training_configuration:920 - LOGGING & CHECKPOINTING CONFIGURATION |
| 2026-04-22 04:28:34.182 | INFO | __main__:print_training_configuration:921 - -------------------------------------------------------------------------------- |
| 2026-04-22 04:28:34.182 | INFO | __main__:print_training_configuration:922 - Output directory : ./outputs/debug |
| 2026-04-22 04:28:34.183 | INFO | __main__:print_training_configuration:923 - Checkpoints directory : s3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/debug/debug |
| 2026-04-22 04:28:34.183 | INFO | __main__:print_training_configuration:924 - Checkpoint interval : 100 |
| 2026-04-22 04:28:34.184 | INFO | __main__:print_training_configuration:925 - Log interval : 1 |
| 2026-04-22 04:28:34.184 | INFO | __main__:print_training_configuration:926 - Resume checkpoint : None |
| 2026-04-22 04:28:34.184 | INFO | __main__:print_training_configuration:927 - Train CSV : assets/demo_train.csv |
| 2026-04-22 04:28:34.185 | INFO | __main__:print_training_configuration:928 - Valid CSV : assets/demo_sample.csv |
| 2026-04-22 04:28:34.185 | INFO | __main__:print_training_configuration:929 - Seed : 42 |
| 2026-04-22 04:28:34.185 | INFO | __main__:print_training_configuration:930 - WandB enabled : True |
| 2026-04-22 04:28:34.186 | INFO | __main__:print_training_configuration:932 - WandB project : SAGE-GRPO |
| 2026-04-22 04:28:34.186 | INFO | __main__:print_training_configuration:933 - WandB run name : debug |
| 2026-04-22 04:28:34.187 | INFO | __main__:__init__:818 - [Effective Config] world_size=8, rank=0, local_rank=0, sp_size=2, dp_size=4, dp_rank=0, sp_rank=0, batch_size=2, num_generations=4, grad_accum=1 |
| 2026-04-22 04:28:34.187 | INFO | __main__:_build_video_dataset:1675 - [Effective GRPO Grouping] video_batch_size=2, dp_replicas=4, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, product=8 |
| 2026-04-22 04:28:34.497 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1171 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 |
| 2026-04-22 04:28:34.498 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1198 - Dataloader init done |
| 2026-04-22 04:28:34.499 | INFO | __main__:_train_grpo:1751 - Starting GRPO training... |
| 2026-04-22 04:28:34.499 | INFO | __main__:_train_grpo:1752 - Max steps: 10000 |
| 2026-04-22 04:28:34.500 | INFO | __main__:_train_grpo:1753 - Batch size: 2 |
| 2026-04-22 04:28:34.500 | INFO | __main__:_train_grpo:1754 - Learning rate: 1e-05 |
| 2026-04-22 04:28:35.784 | INFO | __main__:_train_grpo:1780 - [FLAG] post_train train path -> train_one_step |
| 2026-04-22 04:28:35.795 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:493 - [ParallelGroups] Rank 0, dp_rank 0: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=0, rank_in_group=0 |
| 2026-04-22 04:28:35.922 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:621 - Rank 0 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 |
| 2026-04-22 04:28:35.923 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:646 - Rank 0 using 'all' strategy (no mixgrpo): determistic=False for all timesteps |
| 2026-04-22 04:29:42.397 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:748 - Rank 0: Saved video 1/1 to ./outputs/debug/rl_samples/0000000/group_0/index_7542_seed_28576_dp_0_rank_0_sample_0_slot_0.mp4 |
| 2026-04-22 04:29:43.172 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:621 - Rank 0 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 |
| 2026-04-22 04:29:43.172 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:646 - Rank 0 using 'all' strategy (no mixgrpo): determistic=False for all timesteps |
| 2026-04-22 04:30:45.697 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:748 - Rank 0: Saved video 1/1 to ./outputs/debug/rl_samples/0000000/group_0/index_7542_seed_28577_dp_0_rank_0_sample_1_slot_1.mp4 |
| 2026-04-22 04:30:46.458 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:268 - [GatherRewards] Rank 0: Using local metric names without global sync: ['avg', 'videoalign_local_mq', 'videoalign_local_overall', 'videoalign_local_ta', 'videoalign_local_vq'] |
| 2026-04-22 04:30:47.202 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:336 - [GatherRewards] Rank 0 saved 4 entries for group 0 |
| 2026-04-22 04:30:48.875 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:847 - Rank 0: Pre-computing Reference Model statistics (Batch-wise)... |
| 2026-04-22 04:32:52.399 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:990 - Rank 0: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 123.52s (2.06min) |
| 2026-04-22 04:32:52.974 | INFO | hyvideo.utils.grpo_utils:train_one_step:1433 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] |
| 2026-04-22 04:32:55.902 | INFO | hyvideo.utils.grpo_utils:train_one_step:1545 - Rank 0, sp_rank 0: init adaptive KL controller (init=1.000e-07, min=1.000e-07, max=1.000e-05) |
| 2026-04-22 15:25:56.540 | INFO | __main__:__init__:806 - [RewardDebug] reward_weights(raw): {'VQ': 0.5, 'MQ': 0.5, 'TA': 1.0} |
| 2026-04-22 15:25:56.540 | INFO | __main__:__init__:807 - [RewardDebug] reward_config(effective): {'models': {'videoalign_local': {'weight': 1.0, 'sub_reward': {'VQ': 0.5, 'MQ': 0.5, 'TA': 1.0}}}, 'reward_checkpoint_mode': 'v3'} |
| 2026-04-22 15:25:59.319 | INFO | __main__:print_training_configuration:830 - SYSTEM & HARDWARE CONFIGURATION |
| 2026-04-22 15:25:59.320 | INFO | __main__:print_training_configuration:831 - -------------------------------------------------------------------------------- |
| 2026-04-22 15:25:59.320 | INFO | __main__:print_training_configuration:832 - Number of GPUs : 8 |
| 2026-04-22 15:25:59.321 | INFO | __main__:print_training_configuration:833 - Local rank : 0 |
| 2026-04-22 15:25:59.321 | INFO | __main__:print_training_configuration:834 - Global rank : 0 |
| 2026-04-22 15:25:59.322 | INFO | __main__:print_training_configuration:835 - Device : cuda:0 |
| 2026-04-22 15:25:59.322 | INFO | __main__:print_training_configuration:836 - Master weight dtype : torch.float32 |
| 2026-04-22 15:25:59.323 | INFO | __main__:print_training_configuration:838 - -------------------------------------------------------------------------------- |
| 2026-04-22 15:25:59.323 | INFO | __main__:print_training_configuration:839 - MODEL PARAMETERS & ARCHITECTURE |
| 2026-04-22 15:25:59.324 | INFO | __main__:print_training_configuration:840 - -------------------------------------------------------------------------------- |
| 2026-04-22 15:25:59.335 | INFO | __main__:print_training_configuration:843 - Number of total : 8,326,608,160 |
| 2026-04-22 15:25:59.336 | INFO | __main__:print_training_configuration:844 - Total trainable parameters : 8,326,608,160 |
| 2026-04-22 15:25:59.336 | INFO | __main__:print_training_configuration:845 - DIT model root : ./ckpts |
| 2026-04-22 15:25:59.337 | INFO | __main__:print_training_configuration:846 - Transformer version : 480p_t2v |
| 2026-04-22 15:25:59.337 | INFO | __main__:print_training_configuration:847 - Attention mode : flash |
| 2026-04-22 15:25:59.337 | INFO | __main__:print_training_configuration:848 - Gradient checkpointing : True |
| 2026-04-22 15:25:59.338 | INFO | __main__:print_training_configuration:849 - FSDP enabled : True |
| 2026-04-22 15:25:59.338 | INFO | __main__:print_training_configuration:850 - LoRA enabled : False |
| 2026-04-22 15:25:59.339 | INFO | __main__:print_training_configuration:851 - Glyph byT5 v2 : True |
| 2026-04-22 15:25:59.339 | INFO | __main__:print_training_configuration:852 - ByT5 max length : 256 |
| 2026-04-22 15:25:59.340 | INFO | __main__:print_training_configuration:854 - -------------------------------------------------------------------------------- |
| 2026-04-22 15:25:59.340 | INFO | __main__:print_training_configuration:855 - DISTRIBUTED TRAINING CONFIGURATION |
| 2026-04-22 15:25:59.341 | INFO | __main__:print_training_configuration:856 - -------------------------------------------------------------------------------- |
| 2026-04-22 15:25:59.341 | INFO | __main__:print_training_configuration:857 - World size : 8 |
| 2026-04-22 15:25:59.342 | INFO | __main__:print_training_configuration:858 - DP degree : 4 |
| 2026-04-22 15:25:59.342 | INFO | __main__:print_training_configuration:859 - DP rank : 0 |
| 2026-04-22 15:25:59.343 | INFO | __main__:print_training_configuration:860 - SP size : 2 |
| 2026-04-22 15:25:59.343 | INFO | __main__:print_training_configuration:861 - SP rank : 0 |
| 2026-04-22 15:25:59.343 | INFO | __main__:print_training_configuration:863 - -------------------------------------------------------------------------------- |
| 2026-04-22 15:25:59.344 | INFO | __main__:print_training_configuration:864 - BATCH SIZE & DATA FLOW CONFIGURATION |
| 2026-04-22 15:25:59.344 | INFO | __main__:print_training_configuration:865 - -------------------------------------------------------------------------------- |
| 2026-04-22 15:25:59.345 | INFO | __main__:print_training_configuration:866 - Per GPU batch size : 2 |
| 2026-04-22 15:25:59.345 | INFO | __main__:print_training_configuration:867 - DP degree batch size : 8 |
| 2026-04-22 15:25:59.345 | INFO | __main__:print_training_configuration:868 - Gradient accumulation steps : 1 |
| 2026-04-22 15:25:59.346 | INFO | __main__:print_training_configuration:869 - Num generations : 4 |
| 2026-04-22 15:25:59.347 | INFO | __main__:print_training_configuration:870 - Use same noise : False |
| 2026-04-22 15:25:59.347 | INFO | __main__:print_training_configuration:871 - Num workers : 4 |
| 2026-04-22 15:25:59.347 | INFO | __main__:print_training_configuration:872 - Prefetch factor : 2 |
| 2026-04-22 15:25:59.348 | INFO | __main__:print_training_configuration:874 - -------------------------------------------------------------------------------- |
| 2026-04-22 15:25:59.348 | INFO | __main__:print_training_configuration:875 - OPTIMIZATION CONFIGURATION |
| 2026-04-22 15:25:59.349 | INFO | __main__:print_training_configuration:876 - -------------------------------------------------------------------------------- |
| 2026-04-22 15:25:59.349 | INFO | __main__:print_training_configuration:877 - Optimizer : muon |
| 2026-04-22 15:25:59.350 | INFO | __main__:print_training_configuration:878 - Learning rate : 1e-05 |
| 2026-04-22 15:25:59.350 | INFO | __main__:print_training_configuration:879 - Weight decay : 0.01 |
| 2026-04-22 15:25:59.351 | INFO | __main__:print_training_configuration:880 - Max gradient norm : 1.0 |
| 2026-04-22 15:25:59.351 | INFO | __main__:print_training_configuration:881 - LR scheduler : constant |
| 2026-04-22 15:25:59.352 | INFO | __main__:print_training_configuration:882 - Warmup steps : 500 |
| 2026-04-22 15:25:59.352 | INFO | __main__:print_training_configuration:883 - Max training steps : 10000 |
| 2026-04-22 15:25:59.353 | INFO | __main__:print_training_configuration:885 - -------------------------------------------------------------------------------- |
| 2026-04-22 15:25:59.353 | INFO | __main__:print_training_configuration:886 - INFERENCE & VALIDATION CONFIGURATION |
| 2026-04-22 15:25:59.354 | INFO | __main__:print_training_configuration:887 - -------------------------------------------------------------------------------- |
| 2026-04-22 15:25:59.354 | INFO | __main__:print_training_configuration:888 - Validation interval : 100 |
| 2026-04-22 15:25:59.354 | INFO | __main__:print_training_configuration:889 - Validate at step 0 : False |
| 2026-04-22 15:25:59.354 | INFO | __main__:print_training_configuration:890 - Validation video length : 121 |
| 2026-04-22 15:25:59.355 | INFO | __main__:print_training_configuration:891 - Validation global seed : 930 |
| 2026-04-22 15:25:59.355 | INFO | __main__:print_training_configuration:892 - Validation inference steps : 40 |
| 2026-04-22 15:25:59.356 | INFO | __main__:print_training_configuration:893 - Validation aspect ratio : 16:9 |
| 2026-04-22 15:25:59.356 | INFO | __main__:print_training_configuration:894 - Validation fixed size : 480x864 |
| 2026-04-22 15:25:59.357 | INFO | __main__:print_training_configuration:895 - Validation prompt rewrite : False |
| 2026-04-22 15:25:59.357 | INFO | __main__:print_training_configuration:896 - Validation flow shift : 5.0 |
| 2026-04-22 15:25:59.357 | INFO | __main__:print_training_configuration:897 - Sampling CFG scale : 6.0 |
| 2026-04-22 15:25:59.358 | INFO | __main__:print_training_configuration:898 - Sampling eta : 0.5 |
| 2026-04-22 15:25:59.358 | INFO | __main__:print_training_configuration:899 - Sampling steps (GRPO rollout) : 20 |
| 2026-04-22 15:25:59.358 | INFO | __main__:print_training_configuration:900 - SDE type : sage_grpo |
| 2026-04-22 15:25:59.358 | INFO | __main__:print_training_configuration:902 - -------------------------------------------------------------------------------- |
| 2026-04-22 15:25:59.359 | INFO | __main__:print_training_configuration:903 - GRPO & KL CONFIGURATION |
| 2026-04-22 15:25:59.360 | INFO | __main__:print_training_configuration:904 - -------------------------------------------------------------------------------- |
| 2026-04-22 15:25:59.360 | INFO | __main__:print_training_configuration:905 - Reward model : videoalign_local |
| 2026-04-22 15:25:59.360 | INFO | __main__:print_training_configuration:906 - Reward checkpoint mode : v3 |
| 2026-04-22 15:25:59.361 | INFO | __main__:print_training_configuration:907 - KL weight : 1e-05 |
| 2026-04-22 15:25:59.361 | INFO | __main__:print_training_configuration:908 - KL coef : 1e-07 |
| 2026-04-22 15:25:59.362 | INFO | __main__:print_training_configuration:909 - Use moving KL : True |
| 2026-04-22 15:25:59.362 | INFO | __main__:print_training_configuration:910 - Update ref model step : 1 |
| 2026-04-22 15:25:59.363 | INFO | __main__:print_training_configuration:911 - Use dual KL : True |
| 2026-04-22 15:25:59.363 | INFO | __main__:print_training_configuration:912 - Dual KL moving weight : 1.0 |
| 2026-04-22 15:25:59.364 | INFO | __main__:print_training_configuration:913 - Dual KL step weight : 0.1 |
| 2026-04-22 15:25:59.364 | INFO | __main__:print_training_configuration:914 - KL compute mode : rollout_phase |
| 2026-04-22 15:25:59.364 | INFO | __main__:print_training_configuration:915 - Use grad balancing : True |
| 2026-04-22 15:25:59.365 | INFO | __main__:print_training_configuration:916 - Enable timestep permutation : True |
| 2026-04-22 15:25:59.365 | INFO | __main__:print_training_configuration:917 - Reference mode offload : True |
| 2026-04-22 15:25:59.365 | INFO | __main__:print_training_configuration:919 - -------------------------------------------------------------------------------- |
| 2026-04-22 15:25:59.366 | INFO | __main__:print_training_configuration:920 - LOGGING & CHECKPOINTING CONFIGURATION |
| 2026-04-22 15:25:59.366 | INFO | __main__:print_training_configuration:921 - -------------------------------------------------------------------------------- |
| 2026-04-22 15:25:59.367 | INFO | __main__:print_training_configuration:922 - Output directory : ./outputs/debug |
| 2026-04-22 15:25:59.367 | INFO | __main__:print_training_configuration:923 - Checkpoints directory : s3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/debug/debug |
| 2026-04-22 15:25:59.367 | INFO | __main__:print_training_configuration:924 - Checkpoint interval : 100 |
| 2026-04-22 15:25:59.368 | INFO | __main__:print_training_configuration:925 - Log interval : 1 |
| 2026-04-22 15:25:59.368 | INFO | __main__:print_training_configuration:926 - Resume checkpoint : None |
| 2026-04-22 15:25:59.369 | INFO | __main__:print_training_configuration:927 - Train CSV : assets/demo_train.csv |
| 2026-04-22 15:25:59.369 | INFO | __main__:print_training_configuration:928 - Valid CSV : assets/demo_sample.csv |
| 2026-04-22 15:25:59.370 | INFO | __main__:print_training_configuration:929 - Seed : 42 |
| 2026-04-22 15:25:59.370 | INFO | __main__:print_training_configuration:930 - WandB enabled : True |
| 2026-04-22 15:25:59.370 | INFO | __main__:print_training_configuration:932 - WandB project : SAGE-GRPO |
| 2026-04-22 15:25:59.371 | INFO | __main__:print_training_configuration:933 - WandB run name : debug |
| 2026-04-22 15:25:59.371 | INFO | __main__:__init__:818 - [Effective Config] world_size=8, rank=0, local_rank=0, sp_size=2, dp_size=4, dp_rank=0, sp_rank=0, batch_size=2, num_generations=4, grad_accum=1 |
| 2026-04-22 15:25:59.372 | INFO | __main__:_build_video_dataset:1675 - [Effective GRPO Grouping] video_batch_size=2, dp_replicas=4, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, product=8 |
| 2026-04-22 15:25:59.693 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1171 - [Config Validation] video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1 |
| 2026-04-22 15:25:59.694 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1198 - Dataloader init done |
| 2026-04-22 15:25:59.695 | INFO | __main__:_train_grpo:1751 - Starting GRPO training... |
| 2026-04-22 15:25:59.695 | INFO | __main__:_train_grpo:1752 - Max steps: 10000 |
| 2026-04-22 15:25:59.696 | INFO | __main__:_train_grpo:1753 - Batch size: 2 |
| 2026-04-22 15:25:59.696 | INFO | __main__:_train_grpo:1754 - Learning rate: 1e-05 |
| 2026-04-22 15:26:01.042 | INFO | __main__:_train_grpo:1780 - [FLAG] post_train train path -> train_one_step |
| 2026-04-22 15:26:01.092 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:493 - [ParallelGroups] Rank 0, dp_rank 0: video_batch_size=2, num_generations=4, ranks_per_group=2, num_groups_per_rank=1, group_idx=0, rank_in_group=0 |
| 2026-04-22 15:26:01.210 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:621 - Rank 0 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 |
| 2026-04-22 15:26:01.211 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:646 - Rank 0 using 'all' strategy (no mixgrpo): determistic=False for all timesteps |
| 2026-04-22 15:27:09.193 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:748 - Rank 0: Saved video 1/1 to ./outputs/debug/rl_samples/0000000/group_0/index_7542_seed_28576_dp_0_rank_0_sample_0_slot_0.mp4 |
| 2026-04-22 15:27:10.000 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:621 - Rank 0 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 |
| 2026-04-22 15:27:10.001 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:646 - Rank 0 using 'all' strategy (no mixgrpo): determistic=False for all timesteps |
| 2026-04-22 15:28:12.089 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:748 - Rank 0: Saved video 1/1 to ./outputs/debug/rl_samples/0000000/group_0/index_7542_seed_28577_dp_0_rank_0_sample_1_slot_1.mp4 |
| 2026-04-22 15:28:12.910 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:268 - [GatherRewards] Rank 0: Using local metric names without global sync: ['avg', 'videoalign_local_mq', 'videoalign_local_overall', 'videoalign_local_ta', 'videoalign_local_vq'] |
| 2026-04-22 15:28:13.598 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:336 - [GatherRewards] Rank 0 saved 4 entries for group 0 |
| 2026-04-22 15:28:15.225 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:847 - Rank 0: Pre-computing Reference Model statistics (Batch-wise)... |
| 2026-04-22 15:30:18.905 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:990 - Rank 0: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 123.68s (2.06min) |
| 2026-04-22 15:30:19.500 | INFO | hyvideo.utils.grpo_utils:train_one_step:1433 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] |
| 2026-04-22 15:30:22.390 | INFO | hyvideo.utils.grpo_utils:train_one_step:1545 - Rank 0, sp_rank 0: init adaptive KL controller (init=1.000e-07, min=1.000e-07, max=1.000e-05) |
| 2026-04-22 15:52:10.303 | INFO | __main__:__init__:806 - [RewardDebug] reward_weights(raw): {'VQ': 0.5, 'MQ': 0.5, 'TA': 1.0} |
| 2026-04-22 15:52:10.304 | INFO | __main__:__init__:807 - [RewardDebug] reward_config(effective): {'models': {'videoalign_local': {'weight': 1.0, 'sub_reward': {'VQ': 0.5, 'MQ': 0.5, 'TA': 1.0}}}, 'reward_checkpoint_mode': 'v3'} |
| 2026-04-22 15:52:10.305 | INFO | __main__:print_training_configuration:830 - SYSTEM & HARDWARE CONFIGURATION |
| 2026-04-22 15:52:10.306 | INFO | __main__:print_training_configuration:831 - -------------------------------------------------------------------------------- |
| 2026-04-22 15:52:10.306 | INFO | __main__:print_training_configuration:832 - Number of GPUs : 8 |
| 2026-04-22 15:52:10.306 | INFO | __main__:print_training_configuration:833 - Local rank : 0 |
| 2026-04-22 15:52:10.307 | INFO | __main__:print_training_configuration:834 - Global rank : 0 |
| 2026-04-22 15:52:10.307 | INFO | __main__:print_training_configuration:835 - Device : cuda:0 |
| 2026-04-22 15:52:10.307 | INFO | __main__:print_training_configuration:836 - Master weight dtype : torch.float32 |
| 2026-04-22 15:52:10.307 | INFO | __main__:print_training_configuration:838 - -------------------------------------------------------------------------------- |
| 2026-04-22 15:52:10.308 | INFO | __main__:print_training_configuration:839 - MODEL PARAMETERS & ARCHITECTURE |
| 2026-04-22 15:52:10.308 | INFO | __main__:print_training_configuration:840 - -------------------------------------------------------------------------------- |
| 2026-04-22 15:52:10.320 | INFO | __main__:print_training_configuration:843 - Number of total : 8,326,608,160 |
| 2026-04-22 15:52:10.320 | INFO | __main__:print_training_configuration:844 - Total trainable parameters : 8,326,608,160 |
| 2026-04-22 15:52:10.320 | INFO | __main__:print_training_configuration:845 - DIT model root : ./ckpts |
| 2026-04-22 15:52:10.320 | INFO | __main__:print_training_configuration:846 - Transformer version : 480p_t2v |
| 2026-04-22 15:52:10.321 | INFO | __main__:print_training_configuration:847 - Attention mode : flash |
| 2026-04-22 15:52:10.321 | INFO | __main__:print_training_configuration:848 - Gradient checkpointing : True |
| 2026-04-22 15:52:10.321 | INFO | __main__:print_training_configuration:849 - FSDP enabled : True |
| 2026-04-22 15:52:10.321 | INFO | __main__:print_training_configuration:850 - LoRA enabled : False |
| 2026-04-22 15:52:10.322 | INFO | __main__:print_training_configuration:851 - Glyph byT5 v2 : True |
| 2026-04-22 15:52:10.322 | INFO | __main__:print_training_configuration:852 - ByT5 max length : 256 |
| 2026-04-22 15:52:10.322 | INFO | __main__:print_training_configuration:854 - -------------------------------------------------------------------------------- |
| 2026-04-22 15:52:10.322 | INFO | __main__:print_training_configuration:855 - DISTRIBUTED TRAINING CONFIGURATION |
| 2026-04-22 15:52:10.323 | INFO | __main__:print_training_configuration:856 - -------------------------------------------------------------------------------- |
| 2026-04-22 15:52:10.323 | INFO | __main__:print_training_configuration:857 - World size : 8 |
| 2026-04-22 15:52:10.323 | INFO | __main__:print_training_configuration:858 - DP degree : 8 |
| 2026-04-22 15:52:10.323 | INFO | __main__:print_training_configuration:859 - DP rank : 0 |
| 2026-04-22 15:52:10.324 | INFO | __main__:print_training_configuration:860 - SP size : 1 |
| 2026-04-22 15:52:10.324 | INFO | __main__:print_training_configuration:861 - SP rank : 0 |
| 2026-04-22 15:52:10.324 | INFO | __main__:print_training_configuration:863 - -------------------------------------------------------------------------------- |
| 2026-04-22 15:52:10.324 | INFO | __main__:print_training_configuration:864 - BATCH SIZE & DATA FLOW CONFIGURATION |
| 2026-04-22 15:52:10.325 | INFO | __main__:print_training_configuration:865 - -------------------------------------------------------------------------------- |
| 2026-04-22 15:52:10.325 | INFO | __main__:print_training_configuration:866 - Per GPU batch size : 2 |
| 2026-04-22 15:52:10.325 | INFO | __main__:print_training_configuration:867 - DP degree batch size : 16 |
| 2026-04-22 15:52:10.325 | INFO | __main__:print_training_configuration:868 - Gradient accumulation steps : 1 |
| 2026-04-22 15:52:10.326 | INFO | __main__:print_training_configuration:869 - Num generations : 1 |
| 2026-04-22 15:52:10.326 | INFO | __main__:print_training_configuration:870 - Use same noise : False |
| 2026-04-22 15:52:10.326 | INFO | __main__:print_training_configuration:871 - Num workers : 4 |
| 2026-04-22 15:52:10.326 | INFO | __main__:print_training_configuration:872 - Prefetch factor : 2 |
| 2026-04-22 15:52:10.327 | INFO | __main__:print_training_configuration:874 - -------------------------------------------------------------------------------- |
| 2026-04-22 15:52:10.327 | INFO | __main__:print_training_configuration:875 - OPTIMIZATION CONFIGURATION |
| 2026-04-22 15:52:10.327 | INFO | __main__:print_training_configuration:876 - -------------------------------------------------------------------------------- |
| 2026-04-22 15:52:10.327 | INFO | __main__:print_training_configuration:877 - Optimizer : muon |
| 2026-04-22 15:52:10.328 | INFO | __main__:print_training_configuration:878 - Learning rate : 1e-05 |
| 2026-04-22 15:52:10.328 | INFO | __main__:print_training_configuration:879 - Weight decay : 0.01 |
| 2026-04-22 15:52:10.328 | INFO | __main__:print_training_configuration:880 - Max gradient norm : 1.0 |
| 2026-04-22 15:52:10.328 | INFO | __main__:print_training_configuration:881 - LR scheduler : constant |
| 2026-04-22 15:52:10.329 | INFO | __main__:print_training_configuration:882 - Warmup steps : 500 |
| 2026-04-22 15:52:10.329 | INFO | __main__:print_training_configuration:883 - Max training steps : 10000 |
| 2026-04-22 15:52:10.329 | INFO | __main__:print_training_configuration:885 - -------------------------------------------------------------------------------- |
| 2026-04-22 15:52:10.329 | INFO | __main__:print_training_configuration:886 - INFERENCE & VALIDATION CONFIGURATION |
| 2026-04-22 15:52:10.330 | INFO | __main__:print_training_configuration:887 - -------------------------------------------------------------------------------- |
| 2026-04-22 15:52:10.330 | INFO | __main__:print_training_configuration:888 - Validation interval : 100 |
| 2026-04-22 15:52:10.330 | INFO | __main__:print_training_configuration:889 - Validate at step 0 : False |
| 2026-04-22 15:52:10.330 | INFO | __main__:print_training_configuration:890 - Validation video length : 121 |
| 2026-04-22 15:52:10.331 | INFO | __main__:print_training_configuration:891 - Validation global seed : 930 |
| 2026-04-22 15:52:10.331 | INFO | __main__:print_training_configuration:892 - Validation inference steps : 40 |
| 2026-04-22 15:52:10.331 | INFO | __main__:print_training_configuration:893 - Validation aspect ratio : 16:9 |
| 2026-04-22 15:52:10.331 | INFO | __main__:print_training_configuration:894 - Validation fixed size : 480x864 |
| 2026-04-22 15:52:10.332 | INFO | __main__:print_training_configuration:895 - Validation prompt rewrite : False |
| 2026-04-22 15:52:10.332 | INFO | __main__:print_training_configuration:896 - Validation flow shift : 5.0 |
| 2026-04-22 15:52:10.332 | INFO | __main__:print_training_configuration:897 - Sampling CFG scale : 6.0 |
| 2026-04-22 15:52:10.332 | INFO | __main__:print_training_configuration:898 - Sampling eta : 0.5 |
| 2026-04-22 15:52:10.333 | INFO | __main__:print_training_configuration:899 - Sampling steps (GRPO rollout) : 20 |
| 2026-04-22 15:52:10.333 | INFO | __main__:print_training_configuration:900 - SDE type : sage_grpo |
| 2026-04-22 15:52:10.333 | INFO | __main__:print_training_configuration:902 - -------------------------------------------------------------------------------- |
| 2026-04-22 15:52:10.333 | INFO | __main__:print_training_configuration:903 - GRPO & KL CONFIGURATION |
| 2026-04-22 15:52:10.334 | INFO | __main__:print_training_configuration:904 - -------------------------------------------------------------------------------- |
| 2026-04-22 15:52:10.334 | INFO | __main__:print_training_configuration:905 - Reward model : videoalign_local |
| 2026-04-22 15:52:10.334 | INFO | __main__:print_training_configuration:906 - Reward checkpoint mode : v3 |
| 2026-04-22 15:52:10.334 | INFO | __main__:print_training_configuration:907 - KL weight : 1e-05 |
| 2026-04-22 15:52:10.335 | INFO | __main__:print_training_configuration:908 - KL coef : 1e-07 |
| 2026-04-22 15:52:10.335 | INFO | __main__:print_training_configuration:909 - Use moving KL : True |
| 2026-04-22 15:52:10.335 | INFO | __main__:print_training_configuration:910 - Update ref model step : 1 |
| 2026-04-22 15:52:10.336 | INFO | __main__:print_training_configuration:911 - Use dual KL : True |
| 2026-04-22 15:52:10.336 | INFO | __main__:print_training_configuration:912 - Dual KL moving weight : 1.0 |
| 2026-04-22 15:52:10.336 | INFO | __main__:print_training_configuration:913 - Dual KL step weight : 0.1 |
| 2026-04-22 15:52:10.336 | INFO | __main__:print_training_configuration:914 - KL compute mode : rollout_phase |
| 2026-04-22 15:52:10.337 | INFO | __main__:print_training_configuration:915 - Use grad balancing : True |
| 2026-04-22 15:52:10.337 | INFO | __main__:print_training_configuration:916 - Enable timestep permutation : True |
| 2026-04-22 15:52:10.337 | INFO | __main__:print_training_configuration:917 - Reference mode offload : True |
| 2026-04-22 15:52:10.337 | INFO | __main__:print_training_configuration:919 - -------------------------------------------------------------------------------- |
| 2026-04-22 15:52:10.338 | INFO | __main__:print_training_configuration:920 - LOGGING & CHECKPOINTING CONFIGURATION |
| 2026-04-22 15:52:10.338 | INFO | __main__:print_training_configuration:921 - -------------------------------------------------------------------------------- |
| 2026-04-22 15:52:10.338 | INFO | __main__:print_training_configuration:922 - Output directory : ./outputs/debug |
| 2026-04-22 15:52:10.338 | INFO | __main__:print_training_configuration:923 - Checkpoints directory : s3://snap-genvid-us-east-2/ywang29/checkpoints_large_model/debug/debug |
| 2026-04-22 15:52:10.339 | INFO | __main__:print_training_configuration:924 - Checkpoint interval : 100 |
| 2026-04-22 15:52:10.339 | INFO | __main__:print_training_configuration:925 - Log interval : 1 |
| 2026-04-22 15:52:10.339 | INFO | __main__:print_training_configuration:926 - Resume checkpoint : None |
| 2026-04-22 15:52:10.339 | INFO | __main__:print_training_configuration:927 - Train CSV : assets/demo_train.csv |
| 2026-04-22 15:52:10.340 | INFO | __main__:print_training_configuration:928 - Valid CSV : assets/demo_sample.csv |
| 2026-04-22 15:52:10.340 | INFO | __main__:print_training_configuration:929 - Seed : 42 |
| 2026-04-22 15:52:10.340 | INFO | __main__:print_training_configuration:930 - WandB enabled : False |
| 2026-04-22 15:52:10.340 | INFO | __main__:__init__:818 - [Effective Config] world_size=8, rank=0, local_rank=0, sp_size=1, dp_size=8, dp_rank=0, sp_rank=0, batch_size=2, num_generations=1, grad_accum=1 |
| 2026-04-22 15:52:10.341 | INFO | __main__:_build_video_dataset:1675 - [Effective GRPO Grouping] video_batch_size=2, dp_replicas=8, num_generations=1, ranks_per_group=1, num_groups_per_rank=2, product=16 |
| 2026-04-22 15:52:10.713 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1157 - [Config Validation] video_batch_size=2, num_generations=1, ranks_per_group=1, num_groups_per_rank=2 |
| 2026-04-22 15:52:10.714 | INFO | hyvideo.pipelines.hunyuan_video_grpo_pipeline:get_post_train_video_dataloader:1198 - Dataloader init done |
| 2026-04-22 15:52:10.714 | INFO | __main__:_train_grpo:1751 - Starting GRPO training... |
| 2026-04-22 15:52:10.715 | INFO | __main__:_train_grpo:1752 - Max steps: 10000 |
| 2026-04-22 15:52:10.715 | INFO | __main__:_train_grpo:1753 - Batch size: 2 |
| 2026-04-22 15:52:10.715 | INFO | __main__:_train_grpo:1754 - Learning rate: 1e-05 |
| 2026-04-22 15:52:12.506 | INFO | __main__:_train_grpo:1780 - [FLAG] post_train train path -> train_one_step |
| 2026-04-22 15:52:12.795 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:493 - [ParallelGroups] Rank 0, dp_rank 0: video_batch_size=2, num_generations=1, ranks_per_group=1, num_groups_per_rank=2, group_idx=0, rank_in_group=0 |
| 2026-04-22 15:52:12.907 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:621 - Rank 0 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 |
| 2026-04-22 15:52:12.908 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:646 - Rank 0 using 'all' strategy (no mixgrpo): determistic=False for all timesteps |
| 2026-04-22 15:53:42.061 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:748 - Rank 0: Saved video 1/1 to ./outputs/debug/rl_samples/0000000/group_0/index_7542_seed_28576_dp_0_rank_0_sample_0_slot_0.mp4 |
| 2026-04-22 15:53:42.897 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:621 - Rank 0 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 |
| 2026-04-22 15:53:42.897 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:646 - Rank 0 using 'all' strategy (no mixgrpo): determistic=False for all timesteps |
| 2026-04-22 15:55:09.939 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:748 - Rank 0: Saved video 1/1 to ./outputs/debug/rl_samples/0000000/group_1/index_8214_seed_1872_dp_0_rank_0_sample_0_slot_1.mp4 |
| 2026-04-22 15:55:10.706 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:268 - [GatherRewards] Rank 0: Using local metric names without global sync: ['avg', 'videoalign_local_mq', 'videoalign_local_overall', 'videoalign_local_ta', 'videoalign_local_vq'] |
| 2026-04-22 15:55:10.712 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:356 - [GatherRewards] Rank 0 saved 1 entries for group 0 |
| 2026-04-22 15:55:10.717 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:356 - [GatherRewards] Rank 0 saved 1 entries for group 1 |
| 2026-04-22 15:55:11.976 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:847 - Rank 0: Pre-computing Reference Model statistics (Batch-wise)... |
| 2026-04-22 15:58:06.609 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:990 - Rank 0: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 174.63s (2.91min) |
| 2026-04-22 15:58:07.162 | INFO | hyvideo.utils.grpo_utils:train_one_step:1433 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] |
| 2026-04-22 15:58:11.222 | INFO | hyvideo.utils.grpo_utils:train_one_step:1545 - Rank 0, sp_rank 0: init adaptive KL controller (init=1.000e-07, min=1.000e-07, max=1.000e-05) |
| 2026-04-22 16:08:53.593 | INFO | hyvideo.utils.grpo_utils:train_one_step:1774 - Rank 0, sp_rank 0: update step: 1, train_steps: 40 |
| 2026-04-22 16:08:53.984 | INFO | __main__:_log_precision_diagnostics_once:1372 - [PrecisionDiag] trainable param dtypes: ['torch.float32'] |
| 2026-04-22 16:08:53.985 | INFO | __main__:_log_precision_diagnostics_once:1375 - [PrecisionDiag] Using composable FSDP2 mixed precision (param_dtype=bf16, reduce_dtype=float32). This path does not maintain an explicit FP32 master-weight copy for parameters. |
| 2026-04-22 16:08:53.987 | INFO | __main__:_log_precision_diagnostics_once:1388 - [PrecisionDiag] optimizer tensor-state dtypes: ['torch.float32'] (count=2804) |
| 2026-04-22 16:08:53.988 | INFO | __main__:_log_progress_step:1734 - Training Progress: update_steps 1 train_steps 40 {'step': '1/10000', 'learning_rate': '0.00001000', 'loss': '1.425e-08', 'grad_norm': '3.688e-07', 'KL_weight': '0.00000396', 'param_delta_mean': '0.00000000', 'param_delta_max': '0.00000000', 'rollout_time': '354.27218628', 'grpo_loop_time': '646.50683594', 'ref_model_time': '174.63209534', 'approx_kl': '1.497e-15', 'clipfrac': '0.00000000', 'clipfrac_gt_one': '0.00000000', 'clipfrac_lt_one': '0.00000000', 'policy_loss': '0.00000000', 'kl_loss': '0.00712395', 'kl_moving': '0.00712396', 'kl_stepwise': '-2.105e-08', 'gathered_avg_reward_mean': '-0.95339620', 'gathered_avg_reward_std': '0.69877505', 'gathered_videoalign_local_mq_reward_mean': '-0.77515584', 'gathered_videoalign_local_mq_reward_std': '0.25760680', 'gathered_videoalign_local_overall_reward_mean': '-1.94635177', 'gathered_videoalign_local_overall_reward_std': '0.87107003', 'gathered_videoalign_local_ta_reward_mean': '0.03955926', 'gathered_videoalign_local_ta_reward_std': '0.56261855', 'gathered_videoalign_local_vq_reward_mean': '-1.21075511', 'gathered_videoalign_local_vq_reward_std': '0.23360021', 'gathered_ori_avg_reward_mean': '-0.95339626', 'gathered_ori_avg_reward_std': '0.69877505'} |
| 2026-04-22 16:09:26.196 | INFO | __main__:maybe_update_ref_model_for_kl:685 - [Moving KL] Updated ref_model at update_step=1 (missing=0, unexpected=0) |
| 2026-04-22 16:09:30.578 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:493 - [ParallelGroups] Rank 0, dp_rank 0: video_batch_size=2, num_generations=1, ranks_per_group=1, num_groups_per_rank=2, group_idx=0, rank_in_group=0 |
| 2026-04-22 16:09:30.662 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:621 - Rank 0 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 |
| 2026-04-22 16:09:30.663 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:646 - Rank 0 using 'all' strategy (no mixgrpo): determistic=False for all timesteps |
| 2026-04-22 16:10:59.861 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:748 - Rank 0: Saved video 1/1 to ./outputs/debug/rl_samples/0000001/group_0/index_6771_seed_33047_dp_0_rank_0_sample_0_slot_0.mp4 |
| 2026-04-22 16:11:00.527 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:621 - Rank 0 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 |
| 2026-04-22 16:11:00.528 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:646 - Rank 0 using 'all' strategy (no mixgrpo): determistic=False for all timesteps |
| 2026-04-22 16:12:27.762 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:748 - Rank 0: Saved video 1/1 to ./outputs/debug/rl_samples/0000001/group_1/index_3217_seed_28160_dp_0_rank_0_sample_0_slot_1.mp4 |
| 2026-04-22 16:12:28.621 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:268 - [GatherRewards] Rank 0: Using local metric names without global sync: ['avg', 'videoalign_local_mq', 'videoalign_local_overall', 'videoalign_local_ta', 'videoalign_local_vq'] |
| 2026-04-22 16:12:28.626 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:356 - [GatherRewards] Rank 0 saved 1 entries for group 0 |
| 2026-04-22 16:12:28.629 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:356 - [GatherRewards] Rank 0 saved 1 entries for group 1 |
| 2026-04-22 16:12:29.436 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:847 - Rank 0: Pre-computing Reference Model statistics (Batch-wise)... |
| 2026-04-22 16:15:21.091 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:990 - Rank 0: Reference stats computed (Batched). Shape: torch.Size([2, 20, 32, 21, 22, 39]), Time: 171.61s (2.86min) |
| 2026-04-22 16:15:21.627 | INFO | hyvideo.utils.grpo_utils:train_one_step:1433 - [All Timesteps] Training on timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19] |
| 2026-04-22 16:27:07.140 | INFO | hyvideo.utils.grpo_utils:train_one_step:1774 - Rank 0, sp_rank 0: update step: 2, train_steps: 80 |
| 2026-04-22 16:27:07.557 | INFO | __main__:_log_progress_step:1734 - Training Progress: update_steps 2 train_steps 80 {'step': '2/10000', 'learning_rate': '0.00001000', 'loss': '4.378e-08', 'grad_norm': '0.00000142', 'KL_weight': '0.00000792', 'param_delta_mean': '0.00000000', 'param_delta_max': '0.00000000', 'rollout_time': '351.02148438', 'grpo_loop_time': '705.48779297', 'ref_model_time': '171.60736084', 'approx_kl': '1.522e-15', 'clipfrac': '0.00000000', 'clipfrac_gt_one': '0.00000000', 'clipfrac_lt_one': '0.00000000', 'policy_loss': '0.00000000', 'kl_loss': '0.00720933', 'kl_moving': '0.00720933', 'kl_stepwise': '-2.007e-08', 'gathered_avg_reward_mean': '-1.10273480', 'gathered_avg_reward_std': '0.74919885', 'gathered_videoalign_local_mq_reward_mean': '-0.93683374', 'gathered_videoalign_local_mq_reward_std': '0.21332532', 'gathered_videoalign_local_overall_reward_mean': '-2.19450641', 'gathered_videoalign_local_overall_reward_std': '0.88613653', 'gathered_videoalign_local_ta_reward_mean': '-0.01096334', 'gathered_videoalign_local_ta_reward_std': '0.67627674', 'gathered_videoalign_local_vq_reward_mean': '-1.24670935', 'gathered_videoalign_local_vq_reward_std': '0.34652811', 'gathered_ori_avg_reward_mean': '-1.10273480', 'gathered_ori_avg_reward_std': '0.74919885'} |
| 2026-04-22 16:27:39.176 | INFO | __main__:maybe_update_ref_model_for_kl:685 - [Moving KL] Updated ref_model at update_step=2 (missing=0, unexpected=0) |
| 2026-04-22 16:27:40.846 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:493 - [ParallelGroups] Rank 0, dp_rank 0: video_batch_size=2, num_generations=1, ranks_per_group=1, num_groups_per_rank=2, group_idx=0, rank_in_group=0 |
| 2026-04-22 16:27:40.929 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:621 - Rank 0 generating batch 1/2 (samples 1-1/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 |
| 2026-04-22 16:27:40.929 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:646 - Rank 0 using 'all' strategy (no mixgrpo): determistic=False for all timesteps |
| 2026-04-22 16:29:09.858 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:748 - Rank 0: Saved video 1/1 to ./outputs/debug/rl_samples/0000002/group_0/index_5387_seed_9522_dp_0_rank_0_sample_0_slot_0.mp4 |
| 2026-04-22 16:29:10.494 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:621 - Rank 0 generating batch 2/2 (samples 2-2/2), eta: 0.5, batch_size: 1, 81x352x624, flow_shift: 5.0, infer_steps: 20 |
| 2026-04-22 16:29:10.494 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:646 - Rank 0 using 'all' strategy (no mixgrpo): determistic=False for all timesteps |
| 2026-04-22 16:30:38.036 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:748 - Rank 0: Saved video 1/1 to ./outputs/debug/rl_samples/0000002/group_1/index_8238_seed_31698_dp_0_rank_0_sample_0_slot_1.mp4 |
| 2026-04-22 16:30:38.623 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:268 - [GatherRewards] Rank 0: Using local metric names without global sync: ['avg', 'videoalign_local_mq', 'videoalign_local_overall', 'videoalign_local_ta', 'videoalign_local_vq'] |
| 2026-04-22 16:30:38.628 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:356 - [GatherRewards] Rank 0 saved 1 entries for group 0 |
| 2026-04-22 16:30:38.632 | INFO | hyvideo.utils.grpo_utils:gather_and_process_rewards:356 - [GatherRewards] Rank 0 saved 1 entries for group 1 |
| 2026-04-22 16:30:39.474 | INFO | hyvideo.utils.grpo_utils:prepare_samples_online:847 - Rank 0: Pre-computing Reference Model statistics (Batch-wise)... |
|
|
