华为昇腾910B1基于 LoRA 的 Qwen2.5-7B-Instruct 模型微调

系统环境

Ascend-hdk-910b-npu-driver_24.1.rc3_linux-aarch64.run
Ascend-hdk-910b-npu-firmware_7.5.0.1.129.run
Ascend-cann-toolkit_8.0.RC3.alpha003_linux-aarch64.run
Ascend-cann-kernels-910b_8.0.RC3.alpha003_linux-aarch64.run

虚拟环境

默认已安装conda，git

git clone https://github.com/hiyouga/LLaMA-Factory.git
cd LLaMA-Factory/
conda create -y -n llamafactory python=3.10
conda activate llamafactory
pip install -e ".[torch-npu,metrics]" -i https://pypi.tuna.tsinghua.edu.cn/simple
pip install -e ".[deepspeed,modelscope]" -i https://pypi.tuna.tsinghua.edu.cn/simple
llamafactory-cli env

微调模型

export USE_MODELSCOPE_HUB=1
ASCEND_RT_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 llamafactory-cli train examples/train_lora/qwen2__5_lora_sft.yaml

yaml文件

### model
model_name_or_path: qwen/Qwen2.5-7B-Instruct
trust_remote_code: true

### method
stage: sft
do_train: true
finetuning_type: lora
lora_target: q_proj,v_proj

### ddp
ddp_timeout: 180000000
deepspeed: examples/deepspeed/ds_z0_config.json

### dataset
dataset: alpaca_zh_demo
template: qwen
cutoff_len: 2048
max_samples: 1000
overwrite_cache: true
preprocessing_num_workers: 16

### output
output_dir: saves/Qwen2.5-7B-Instruct/lora/sft
logging_steps: 10
save_steps: 1000
plot_loss: true
overwrite_output_dir: true

### train
per_device_train_batch_size: 1
gradient_accumulation_steps: 8
learning_rate: 0.0001
num_train_epochs: 120
lr_scheduler_type: cosine
bf16: true

### eval
val_size: 0.1
per_device_eval_batch_size: 1
evaluation_strategy: steps
eval_steps: 500
    

training_loss

training_eval_loss