Ascend LlamaFactory微调书生模型

1.环境安装

conda create -y -n llamafactory_lab python=3.10 conda activate llamafactory_lab git clone https://gh.llkk.cc/https://github.com/hiyouga/LLaMA-Factory.git cd LLaMA-Factory git checkout v0.9.3 pip install -e ".[torch-npu,metrics]" -i https://pypi.tuna.tsinghua.edu.cn/simple pip install torch==2.6.0 torch-npu==2.6.0 torchvision

2.安装校验

使用以下指令对 LLaMA-Factory × 昇腾的安装进行校验： llamafactory-cli env

3.微调前的原模型效果呈现

安装依赖

pip install transformers==4.47.1 # 环境变量设置单卡GPU export ASCEND_RT_VISIBLE_DEVICES=0 # 启动推理 llamafactory-cli chat --model_name_or_path /share/new_models/internlm3/internlm3-8b-instruct --trust_remote_code True

输出结果：

4.微调书生模型

安装依赖

pip install transformers==4.47.1 pip install huggingface_hub==0.34.0 # 创建工作目录 mkdir /root/llamafactory_workdir

安装数据集

export HF_ENDPOINT=https://hf-mirror.com huggingface-cli download --repo-type dataset --resume-download llamafactory/alpaca_en --local-dir /root/dataset/alpaca_en huggingface-cli download --repo-type dataset --resume-download llamafactory/alpaca_zh --local-dir /root/dataset/alpaca_zh

创建dataset_info.json文件

mkdir /root/llamafactory_workdir/data cd /root/llamafactory_workdir/data touch dataset_info.json

dataset_info.json文件

{ "alpaca_en":{ "file_name": "/root/dataset/alpaca_en/alpaca_data_en_52k.json", "columns": { "prompt": "instruction", "query": "input", "response": "output" } }, "alpaca_zh":{ "file_name": "/root/dataset/alpaca_zh/alpaca_data_zh_51k.json", "columns": { "prompt": "instruction", "query": "input", "response": "output" } } }

Yaml配置文件

cd /root/llamafactory_workdir touch internlm3_lora_sft_ds.yaml

internlm3_lora_sft_ds.yaml文件内容

### model model_name_or_path: /share/new_models/internlm3/internlm3-8b-instruct trust_remote_code: True ### method stage: sft do_train: true finetuning_type: lora lora_target: q_proj,v_proj ### dataset dataset: alpaca_en,alpaca_zh template: intern2 cutoff_len: 1024 max_samples: 1000 overwrite_cache: true preprocessing_num_workers: 16 ### output output_dir: saves/internlm3-8b-instruct/lora/sft logging_steps: 10 save_steps: 500 plot_loss: true overwrite_output_dir: true ### train per_device_train_batch_size: 1 gradient_accumulation_steps: 2 learning_rate: 0.0001 num_train_epochs: 3.0 lr_scheduler_type: cosine warmup_ratio: 0.1 fp16: true ### eval val_size: 0.1 per_device_eval_batch_size: 1 eval_strategy: steps eval_steps: 500

开启微调dong

torchrun --nproc_per_node 1 \ --nnodes 1 \ --node_rank 0 \ --master_addr 127.0.0.1 \ --master_port 7007 \ /root/LLaMA-Factory/src/train.py internlm3_lora_sft_ds.yaml

动态合并LoRA的微调

llamafactory-cli chat --model_name_or_path /share/new_models/internlm3/internlm3-8b-instruct \ --adapter_name_or_path saves/internlm3-8b-instruct/lora/sft \ --template intern \ --finetuning_type lora \ --trust_remote_code True

从推理结果看和源模型的输出有所区别了，那就说明微调完成了

微调完成之后可以将微调的权重和原始的模型合并导出

ASCEND_RT_VISIBLE_DEVICES=0 llamafactory-cli export \ --model_name_or_path /share/new_models/internlm3/internlm3-8b-instruct \ --adapter_name_or_path saves/internlm3-8b-instruct/lora/sft \ --template intern2 \ --finetuning_type lora \ --export_dir saves/internlm3-8b-instruct/lora/megred-model-path \ --export_device auto \ --export_legacy_format False \ --trust_remote_code True