Qwen3-Omni-30B-A3B-Instruct微调混合audio数据集会卡住

### model
model: /path/Qwen3-Omni-30B-A3B-Instruct
attn_impl: flash_attn

### method
train_type: full
deepspeed: zero3
freeze_vit: true
freeze_aligner: true

### dataset
dataset:
  - text_dataset
  - audio_dataset
dataloader_num_workers: 64
dataset_num_proc: 128
max_length: 8192
packing: true
# model_name: swift-bot
# model_author: swift

### output
output_dir: /path/checkpoint
logging_steps: 1
eval_steps: 10000000
save_steps: 10000000
save_total_limit: 1

### train
num_train_epochs: 3
per_device_train_batch_size: 1
learning_rate: 1e-4
padding_free: false
gradient_checkpointing: true
gradient_accumulation_steps: 4
warmup_ratio: 0.1
torch_dtype: bfloat16


### eval
per_device_eval_batch_size: 1
split_dataset_ratio: 0.0

yaml配置如上所示，32卡h200，使用了sft-packing技术，当max_length=4096训练的时候是正常的；当max_length=8192，训练像卡住了一样，

Train:   0%|          | 0/33 [00:00<?, ?it/s][INFO:swift] use_logits_to_keep: False
[INFO:swift] router_aux_loss_coef: 0.0
/usr/local/lib/python3.11/site-packages/transformers/models/qwen3_omni_moe/modeling_qwen3_omni_moe.py:1070: UserWarning: Specified kernel cache directory could not be created! This disables kernel caching. Specified directory is /root/.cache/torch/kernels. This warning will appear only once per process. (Triggered internally at /pytorch/aten/src/ATen/native/cuda/jit_utils.cpp:1487.)
  total_tokens = int(torch.prod(grid_thw, dim=1).sum().item())
/usr/local/lib/python3.11/site-packages/transformers/models/qwen3_omni_moe/modeling_qwen3_omni_moe.py:1070: UserWarning: Specified kernel cache directory could not be created! This disables kernel caching. Specified directory is /root/.cache/torch/kernels. This warning will appear only once per process. (Triggered internally at /pytorch/aten/src/ATen/native/cuda/jit_utils.cpp:1487.)
  total_tokens = int(torch.prod(grid_thw, dim=1).sum().item())
/usr/local/lib/python3.11/site-packages/transformers/models/qwen3_omni_moe/modeling_qwen3_omni_moe.py:1070: UserWarning: Specified kernel cache directory could not be created! This disables kernel caching. Specified directory is /root/.cache/torch/kernels. This warning will appear only once per process. (Triggered internally at /pytorch/aten/src/ATen/native/cuda/jit_utils.cpp:1487.)
  total_tokens = int(torch.prod(grid_thw, dim=1).sum().item())
/usr/local/lib/python3.11/site-packages/transformers/models/qwen3_omni_moe/modeling_qwen3_omni_moe.py:1070: UserWarning: Specified kernel cache directory could not be created! This disables kernel caching. Specified directory is /root/.cache/torch/kernels. This warning will appear only once per process. (Triggered internally at /pytorch/aten/src/ATen/native/cuda/jit_utils.cpp:1487.)
  total_tokens = int(torch.prod(grid_thw, dim=1).sum().item())
......

奇怪的是，我把audio_dataset去掉，只用文本的数据集做sft_packing，max_length=16384也是能正常训练的。不知道是什么问题造成的？

Start by reproducing the provided 32-card Qwen3-Omni configuration with both text_dataset and audio_dataset, comparing max_length 4096 and 8192. Then compare it with the text-only run at max_length 16384 and inspect where training stops after the kernel-cache warnings. Done means the mixed audio/text SFT-packing run progresses at max_length 8192 or the blocking condition is identified.