Activity

1. mi804 commented on Sep 1, 2026

   

   mi804

   on Sep 1, 2026

   Collaborator

   More actions

   Thanks for the report.

   This error means audio_latents was never produced for that sample, which happens when a video in your dataset has no audio track — without input_audio, the audio encoding step is skipped and model_fn fails on the missing argument. Only rank 5 crashed, so likely just a subset of your data is affected.

   Try adding --silent_on_missing_audio to your training command — it substitutes a silent track for those samples.

   If the error persists, please share your full training command.
2. stihuangyuan commented on Sep 14, 2026

   

   stihuangyuan

   on Sep 14, 2026

   Last edited by stihuangyuan

   Author

   More actions

   --silent_on_missing_audio

   This error still happen when stage2 training. my training command follow github doc：
   modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --include "minimax_h3/MiniMax-H3-FL2VA/*" --local_dir ./data/diffsynth_example_dataset

   T2VA - stage 1 (data process)

   accelerate launch examples/minimax_h3/model_training/train.py
   --dataset_base_path data/diffsynth_example_dataset/minimax_h3/MiniMax-H3-FL2VA
   --dataset_metadata_path data/diffsynth_example_dataset/minimax_h3/MiniMax-H3-FL2VA/metadata.csv
   --data_file_keys "video,input_audio"
   --extra_inputs "input_audio"
   --height 480
   --width 832
   --num_frames 124
   --dataset_repeat 1
   --model_id_with_origin_paths "MiniMax/MiniMax-H3:FL2VA/text_encoder/model*.safetensors,MiniMax/MiniMax-H3:FL2VA/video_vae/source/model.safetensors,MiniMax/MiniMax-H3:FL2VA/audio_vae/model.safetensors"
   --learning_rate 1e-5
   --num_epochs 1
   --remove_prefix_in_ckpt "pipe.dit."
   --output_path "./models/train/MiniMax-H3-T2VA-full-split-cache"
   --trainable_models "dit"
   --use_gradient_checkpointing
   --task "sft:data_process"

   T2VA - stage 2 (train)

   accelerate launch --config_file examples/minimax_h3/model_training/full/accelerate_config_zero3.yaml
   examples/minimax_h3/model_training/train.py
   --dataset_base_path ./models/train/MiniMax-H3-T2VA-full-split-cache
   --data_file_keys "video,input_audio"
   --extra_inputs "input_audio"
   --height 480
   --width 832
   --num_frames 124
   --dataset_repeat 100
   --model_id_with_origin_paths "MiniMax/MiniMax-H3:FL2VA/transformer/model*.safetensors"
   --learning_rate 1e-5
   --num_epochs 2
   --remove_prefix_in_ckpt "pipe.dit."
   --output_path "./models/train/MiniMax-H3-T2VA-full"
   --trainable_models "dit"
   --use_gradient_checkpointing
   --find_unused_parameters
   --task "sft:train"
3. stihuangyuan commented on Sep 16, 2026

   

   stihuangyuan

   on Sep 16, 2026

   Author

   More actions

   Thanks for the report.

   This error means audio_latents was never produced for that sample, which happens when a video in your dataset has no audio track — without input_audio, the audio encoding step is skipped and model_fn fails on the missing argument. Only rank 5 crashed, so likely just a subset of your data is affected.

   Try adding --silent_on_missing_audio to your training command — it substitutes a silent track for those samples.

   If the error persists, please share your full training command.

   @mi804 This error still happen when stage2 training. my training command follow github doc：
   modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --include "minimax_h3/MiniMax-H3-FL2VA/*" --local_dir ./data/diffsynth_example_dataset

   T2VA - stage 1 (data process)
   accelerate launch examples/minimax_h3/model_training/train.py
   --dataset_base_path data/diffsynth_example_dataset/minimax_h3/MiniMax-H3-FL2VA
   --dataset_metadata_path data/diffsynth_example_dataset/minimax_h3/MiniMax-H3-FL2VA/metadata.csv
   --data_file_keys "video,input_audio"
   --extra_inputs "input_audio"
   --height 480
   --width 832
   --num_frames 124
   --dataset_repeat 1
   --model_id_with_origin_paths "MiniMax/MiniMax-H3:FL2VA/text_encoder/model*.safetensors,MiniMax/MiniMax-H3:FL2VA/video_vae/source/model.safetensors,MiniMax/MiniMax-H3:FL2VA/audio_vae/model.safetensors"
   --learning_rate 1e-5
   --num_epochs 1
   --remove_prefix_in_ckpt "pipe.dit."
   --output_path "./models/train/MiniMax-H3-T2VA-full-split-cache"
   --trainable_models "dit"
   --use_gradient_checkpointing
   --task "sft:data_process"

   T2VA - stage 2 (train)
   accelerate launch --config_file examples/minimax_h3/model_training/full/accelerate_config_zero3.yaml
   examples/minimax_h3/model_training/train.py
   --dataset_base_path ./models/train/MiniMax-H3-T2VA-full-split-cache
   --data_file_keys "video,input_audio"
   --extra_inputs "input_audio"
   --height 480
   --width 832
   --num_frames 124
   --dataset_repeat 100
   --model_id_with_origin_paths "MiniMax/MiniMax-H3:FL2VA/transformer/model*.safetensors"
   --learning_rate 1e-5
   --num_epochs 2
   --remove_prefix_in_ckpt "pipe.dit."
   --output_path "./models/train/MiniMax-H3-T2VA-full"
   --trainable_models "dit"
   --use_gradient_checkpointing
   --find_unused_parameters
   --task "sft:train"
4. mi804 commented on Sep 20, 2026

   

   mi804

   on Sep 20, 2026

   Collaborator

   More actions

   Thanks for reporting this! The example videos include audio, and we’ve verified that training runs successfully with the example script. One possible cause is a missing audio dependency, such as torchcodec, in your Python environment, though we’d need the full error logs to confirm. Could you please share the complete traceback, along with any earlier warnings or errors? This would help us investigate further.
5. mvanhorn commented on Oct 13, 2026

   

   mvanhorn

   on Oct 13, 2026

   More actions

   FlowMatchSFTMiniMaxH3AudioVideoLoss sampled audio noise, stored audio_latents, and added the audio loss only when the key audio_input_latents was present.

   require_minimax_h3_audio_input_latents returns audio_input_latents only when it is a tensor and otherwise raises ValueError describing the skipped encode and a data-process rerun.

   #1748