镜像：image.sourcefind.cn:5000/dcu/admin/base/jupyterlab-llamafactory:pytorch2.4.1-py3.10-dtk25.04.1-model 报错详情在这个帖子上：[光合开发者社区-论坛](https://forum.sourcefind.cn/thread/98d075ef-c10b-11f0-b077-0242ac150003)（👈点击查看 jupyterlab-llamafactory 微调时候报内核错误KERNEL VMFault） 复现方式: 在bw1000环境 ，用k8s 起一个pytorchjob， 如果想加速复现，可以起多个 数据集见附件 配置参数： ``` bf16: true cutoff_len: 1024 dataset: hh dataset_dir: /storage/main/users/wh1107/test-data ddp_timeout: 180000000 do_train: true enable_thinking: false eval_steps: 100 eval_strategy: 'no' finetuning_type: full gradient_accumulation_steps: 1 learning_rate: 0.0001 logging_steps: 10 lora_alpha: 16.0 lora_drop_out: 0.0 lora_target: all lr_scheduler_type: cosine max_samples: 1000 model_name_or_path: /model/Qwen2.5-0.5B-Instruct num_train_epochs: 2.0 output_dir: /storage/main/users/wh1107/train/output-sft/003/output overwrite_cache: true overwrite_output_dir: true packing: false per_device_eval_batch_size: 1 per_device_train_batch_size: 1 plot_loss: true preprocessing_num_workers: 16 report_to: tensorboard #save_strategy: 'no' #save_steps: 400 stage: sft template: qwen val_size: 0.1 warmup_ratio: 0.1 ```