Qwen3-VL SFT n_try_fetch error

Checklist / 检查清单

• I have searched existing issues, and this is a new question or discussion topic. / 我已经搜索过现有的 issues，确认这是一个新的问题与讨论。

Question Description / 问题描述

boot cmd:
when FPS_MAX_FRAMES = 16 is ok, but FPS_MAX_FRAMES=360 or larger has error.

PYTORCH_CUDA_ALLOC_CONF='expandable_segments:True' \
IMAGE_MAX_TOKEN_NUM=1024 \
VIDEO_MAX_TOKEN_NUM=256 \
FPS_MAX_FRAMES=360 \
NPROC_PER_NODE=2 \
CUDA_VISIBLE_DEVICES=0,1 \
swift sft \
    --model local_path/Qwen3-VL-8B-Instruct \
    --dataset 'dataset/data_train.jsonl' \
    --val_dataset 'dataset/data_val.jsonl' \
    --load_from_cache_file true \
    --split_dataset_ratio 0.01 \
    --train_type lora \
    --torch_dtype bfloat16 \
    --num_train_epochs 1 \
    --per_device_train_batch_size 1 \
    --per_device_eval_batch_size 1 \
    --attn_impl flash_attn \
    --padding_free true \
    --learning_rate 1e-4 \
    --lora_rank 8 \
    --lora_alpha 32 \
    --target_modules all-linear \
    --freeze_vit true \
    --freeze_aligner true \
    --packing true \
    --gradient_checkpointing true \
    --vit_gradient_checkpointing false \
    --gradient_accumulation_steps 2 \
    --eval_steps 100 \
    --save_steps 100 \
    --save_total_limit 2 \
    --logging_steps 5 \
    --max_length 4096 \
    --output_dir output \
    --warmup_ratio 0.05 \
    --deepspeed zero2 \
    --dataset_num_proc 4 \
    --dataloader_num_workers 4

Error log:

[rank0]: Traceback (most recent call last):
[rank0]:   File "/usr/local/miniconda3/lib/python3.10/site-packages/swift/cli/sft.py", line 20, in <module>
[rank0]:     sft_main()
[rank0]:   File "/usr/local/miniconda3/lib/python3.10/site-packages/swift/llm/train/sft.py", line 365, in sft_main
[rank0]:     return SwiftSft(args).main()
[rank0]:   File "/usr/local/miniconda3/lib/python3.10/site-packages/swift/llm/base.py", line 49, in main
[rank0]:     result = self.run()
[rank0]:   File "/usr/local/miniconda3/lib/python3.10/site-packages/swift/ray/base.py", line 170, in wrapper
[rank0]:     return func(self, *args, **kwargs)
[rank0]:   File "/usr/local/miniconda3/lib/python3.10/site-packages/swift/llm/train/sft.py", line 184, in run
[rank0]:     train_dataset, val_dataset = self._prepare_dataset()
[rank0]:   File "/usr/local/miniconda3/lib/python3.10/site-packages/swift/ray/base.py", line 170, in wrapper
[rank0]:     return func(self, *args, **kwargs)
[rank0]:   File "/usr/local/miniconda3/lib/python3.10/site-packages/swift/llm/train/sft.py", line 144, in _prepare_dataset
[rank0]:     datasets = self._post_process_datasets(datasets)
[rank0]:   File "/usr/local/miniconda3/lib/python3.10/site-packages/swift/llm/train/sft.py", line 160, in _post_process_datasets
[rank0]:     dataset = LazyLLMDataset(dataset, template.encode, strict=args.strict, random_state=args.data_seed)
[rank0]:   File "/usr/local/miniconda3/lib/python3.10/site-packages/swift/llm/dataset/utils.py", line 77, in __init__
[rank0]:     assert n_try_fetch >= 1
[rank0]: AssertionError

Start with swift/llm/train/sft.py, following _prepare_dataset and _post_process_datasets into swift/llm/dataset/utils.py and LazyLLMDataset. Reproduce the command with FPS_MAX_FRAMES=16 and 360, then inspect why n_try_fetch reaches the assertion; done means the higher-frame configuration no longer fails during dataset preparation.