LLaMA-Factory是一个相当优秀的微调工具。这里提供一个dockerfile和一个train脚本，用于多卡微调，供大家参考。

Dockerfile

FROM nvidia/cuda:12.1.0-cudnn8-devel-ubuntu22.04
# python3
RUN apt-get update && apt-get install -y python3.10 python3-pip
# torch
COPY torch-2.2.0+cu121-cp310-cp310-linux_x86_64.whl torch-2.2.0+cu121-cp310-cp310-linux_x86_64.whl
RUN pip3 install torch-2.2.0+cu121-cp310-cp310-linux_x86_64.whl
# llama factory requirements
RUN pip3 install transformers==4.37.2 datasets==2.16.1 accelerate==0.25.0 peft==0.7.1 trl==0.7.10 gradio==3.50.2 \
    deepspeed modelscope ipython scipy einops sentencepiece protobuf jieba rouge-chinese nltk sse-starlette matplotlib \
    --no-cache-dir -i https://pypi.tuna.tsinghua.edu.cn/simple
# unsloth
RUN apt-get install -y git
RUN pip install --upgrade pip
RUN pip install triton --no-cache-dir -i https://pypi.tuna.tsinghua.edu.cn/simple
RUN pip install "unsloth[cu121_ampere_torch220] @ git+https://github.com/unslothai/unsloth.git"

train.sh

docker run \
  -it \
  --rm \
  --name llm \
  --network=host \
  --shm-size 32G \
  --gpus all \
  -v /home/[user_name]/.cache/modelscope/hub/:/root/.cache/modelscope/hub/ \
  -v /home/[user_name]/LLaMA-Factory/:/LLaMA-Factory/ \
  -v /home/[user_name]/.cache/huggingface/accelerate/default_config.yaml:/root/.cache/huggingface/accelerate/default_config.yaml \
  -w /LLaMA-Factory \
  -e USE_MODELSCOPE_HUB=1 \
  llm:v1.1 \
  accelerate launch src/train_bash.py \
    --stage sft \
    --do_train True \
    --model_name_or_path ZhipuAI/chatglm3-6b \
    --finetuning_type lora \
    --use_unsloth True \
    --template chatglm3 \
    --dataset_dir data \
    --dataset alpaca_gpt4_zh \
    --cutoff_len 512 \
    --learning_rate 5e-05 \
    --num_train_epochs 2.0 \
    --max_samples 8000 \
    --per_device_train_batch_size 1 \
    --gradient_accumulation_steps 2 \
    --lr_scheduler_type cosine \
    --max_grad_norm 1.0 \
    --logging_steps 5 \
    --save_steps 1000 \
    --warmup_steps 0 \
    --lora_rank 8 \
    --lora_dropout 0.1 \
    --lora_target query_key_value \
    --output_dir saves/ChatGLM3-6B-Chat/lora/train_20240212 \
    --fp16 True \
    --plot_loss True

注意事项：

1. –shm-size 32G --gpus all 这两个参数是必要的
2. –use_unsloth True 可以调用unsloth实现加速
3. 需要保证–gradient_accumulation_steps 2在deepspeed配置中的一致性

default_config.yaml

compute_environment: LOCAL_MACHINE
debug: false
# distributed_type: MULTI_GPU
deepspeed_config:
  deepspeed_multinode_launcher: standard
  gradient_accumulation_steps: 2
  offload_optimizer_device: none
  offload_param_device: none
  zero3_init_flag: false
  zero3_save_16bit_model: false
  zero_stage: 2
distributed_type: DEEPSPEED 
downcast_bf16: 'no'
gpu_ids: all
machine_rank: 0
main_training_function: main
mixed_precision: bf16
num_machines: 1
num_processes: 2
rdzv_backend: static
same_network: true
tpu_env: []
tpu_use_cluster: false
tpu_use_sudo: false
use_cpu: false

感谢以下两篇博客：

单卡 3 小时训练专属大模型 Agent：基于 LLaMA Factory 实战

Accelerate 0.24.0文档 二：DeepSpeed集成