1.构建conda环境

conda create -n qwen-coder-sft python=3.9 conda activate qwen-coder-sft

2.安装依赖

pip install -r dev-requirements.txt

3.数据tokenize

样例数据在finetuning/sft/data下面

cd finetuning/sft
bash ./scripts/binarize_data.sh ${INPUT_PATH} ${OUTPUT_PATH} ${TOKENIZER_PATH}

4.模型微调(_mac支持在mac上进行微调)

全参数微调--7B可以跑起来，但是32B跑不起来
bash ./scripts/sft_qwencoder_mac.sh

lora微调--7B和32B都可以跑起来
bash ./scripts/sft_qwencoder_with_lora_mac.sh

5. 模型合并

BASE_MODEL_PATH=${1}
TRAIN_ADAPTERS_PATH=${2}
OUTPUT_PATH=${3}
bash ./scripts/merge_adapter.sh ${BASE_MODEL_PATH} ${TRAIN_ADAPTERS_PATH} ${OUTPUT_PATH}

6. 模型推理

cd inference/
修改合并后的模型地址
python infer.py