Qwen3.5系列模型评测时不断卡住

Checklist / 检查清单

• I have searched existing issues, and this is a new bug report. / 我已经搜索过现有的 issues，确认这是一个新的 bug report。

Bug Description / Bug 描述

Qwen3.5评测时卡住。

How to Reproduce / 如何复现

脚本如下：
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 swift eval --model megatron_output/Qwen3.5-35B-A3B/exp1-1/checkpoint-8600-merged --eval_backend Native --infer_backend transformers --eval_dataset mmmu real_world_qa omni_doc_bench --enable_thinking false
ms-swift 的版本4.0.2
transformers==5.2.0

Additional Information / 补充信息

Qwen3.5系列的评测经常正常一会，马上重复
[INFO:swift] {'num_prompt_tokens': 0, 'num_generated_tokens': 0, 'num_samples': 0, 'runtime': 20.01921532, 'samples/s': 0.0, 'tokens/s': 0.0}
一直retry

Start with the reported swift eval command using Native evaluation, the transformers inference backend, and the listed datasets. Reproduce with ms-swift 4.0.2 and transformers 5.2.0, then follow the retry path after the zero-sample runtime log. Done means the Qwen3.5 evaluation completes without repeatedly retrying.