修改的地方：
**`Qwen2.5-VL/LLaMA-Factory/train&&inference.ipynb`: 训练指令及推理代码（推理代码是ipynb文件最后一个cell）**

**`Qwen2.5-VL/LLaMA-Factory/src/llamafactory/train/sft/workflow.py`: 训练代码**

**`Qwen2.5-VL/LLaMA-Factory/src/llamafactory/train/sft/SVIPModel.py`: 模型类定义代码**

**`Qwen2.5-VL/LLaMA-Factory/src/llamafactory/data/processor/supervised.py`: 训练数据转换代码**


scp -o 'ProxyJump hongwei@141.147.112.74' -P 1049 gmh@10.130.129.21:/home1/gmh/InternVL hongwei@10.16.7.157:/mnt/weka/cs_workplace/

CUDA_VISIBLE_DEVICES=1 python -m vllm.entrypoints.openai.api_server --model /home1/gmh/InternVL/InternVL_2_5 --tensor-parallel-size 1 --port 8000 --gpu-memory-utilization 0.9 --max-model-len 8096 --trust-remote-code --max-num-seqs 32 --max-num-batched-tokens 8192