source ~/.bashrc
df -h
uname -m && cat /etc/*release
gcc –version
sudo apt install git
conda create –name lf python==3.10
conda activate lf
conda install pytorch==2.3.1 torchvision==0.18.1 torchaudio==2.3.1 pytorch-cuda=12.1 -c pytorch -c nvidia
python -c ‘import torch;print(torch.__version__)’
python -c ‘import torch;print(torch.cuda.is_available())’
pip install -r requirements.txt -i https://pypi.tuna.tsinghua.edu.cn/simple
cd LLama-Factory/
pip install -e .[metrics]
如果要多卡训练：pip3 install deepspeed==0.14.0
单卡启动：CUDA_VISIBLE_DEVICES=0 python src/webui.py
可选：nohup CUDA_VISIBLE_DEVICES=0 python src/webui.py
多卡启动：CUDA_VISIBLE_DEVICES=0,1 python src/webui.py
| 参数 | 参数说明 |
| version | 显示版本信息 |
| train | 命令行版本训练 |
| chat | 命令行版本推理 chat |
| export | 模型合并和导出 |
| api | 启动 API server，供接口调用 |
| eval | 使用 mmmlu 等标准数据集做评测 |
| webchat | 前端版本纯推理的 chat 页面 |
| webui | 启动 LlamaBoard 前端页面，包含可视 |
llamafactory-cli train \
–stage sft \
–do_train True \
–model_name_or_path /root/model_path \
–preprocessing_num_workers 16 \
–finetuning_type lora \
–template qwen \
–flash_attn auto \
–dataset_dir data \
–dataset name \
–cutoff_len 4096 \
–learning_rate 5e-05 \
–num_train_epochs 100.0 \
–max_samples 100000 \
–per_device_train_batch_size 4 \
–gradient_accumulation_steps 8 \
–lr_scheduler_type cosine \
–max_grad_norm 1.0 \
–logging_steps 5 \
–save_steps 100 \
–warmup_steps 0 \
–packing False \
–report_to none \
–output_dir saves_path \
–fp16 True \
–plot_loss True \
–trust_remote_code True \
–ddp_timeout 180000000 \
–include_num_input_tokens_seen True \
–optim adamw_torch \
–lora_rank 8 \
–lora_alpha 16 \
–lora_dropout 0 \
–lora_target all \
llamafactory-cli train examples/train_lora/qwen_lora_sft.yaml
保存并导出完整模型
启动合成后模型
CUDA_VISIBLE_DEVICES=0 llamafactory-cli webchat \
–model_name_or_path /root/Qwen2.5-7B-Instruct \
–template qwen
json文本报错
数据集切分报错，json文件可能包含字节顺序标记
解决方法：使用notepad++转换