Qwen3-VL emb微调，训练到一半CPU OOM

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

• 32卡分布式训练，使用ZERO2。

• 训练log的最后一行loss日志：{'loss': 1.89589844, 'grad_norm': 0.04135821, 'learning_rate': 4.134e-05, 'epoch': 0.31, 'global_step/max_steps': '60/193', 'percentage': '31.09%', 'elapsed_time': '3h 36m 48s', 'remaining_time': '8h 0m 35s', 'memory(GiB)': 20.61, 'train_speed(iter/s)': 0.004612}

Rank(Worker): 16 caught unknown unhandled exception:
Exception type: OOM
Detail:
return code is -9, maybe OUT OF MEMORY, please check kmon

How to Reproduce / 如何复现

swift_version：4.0.0.dev0
"model": "Qwen/Qwen3-VL-Embedding-8B",
"dataset": 视频训练数据.json,
"task_type": "embedding",
"tuner_type": "lora",
"lora_rank": 8,
"lora_alpha": 32,
"target_modules": "all-linear",
"learning_rate": "5e-5",
"attn_impl": "flash_attn",
"max_length": 4096,
"load_from_cache_file": true,
"num_train_epochs": 1,
"per_device_train_batch_size": 16,
"per_device_eval_batch_size": 2,
"gradient_accumulation_steps": 2,
"dataloader_num_workers": 0,
"dataset_num_proc": 1,
"warmup_ratio": 0.05,
"loss_type": "infonce",
"infonce_temperature": 0.02,
"deepspeed": "zero2",
"padding_free": true,
"torch_dtype": "bfloat16",
"save_steps": 20,
"eval_steps": 500,
"save_total_limit": 2,
"logging_steps": 20,
"model_kwargs": {"fps_max_frames":6, "image_max_token_num":60, "video_max_token_num":60,"video_min_token_num":8}

Additional Information / 补充信息

No response

Start by reproducing the reported 32-card distributed Qwen3-VL embedding run with the supplied ZERO2, batch, video, and LoRA settings. Inspect the final training log, Rank 16 worker failure, and kmon output to determine what causes the CPU OOM. Done means the cause is identified and the same configuration can progress without the reported worker termination.