▌XTuner 简介

xtuner help

xtuner version

xtuner list-cfg

xtuner list-cfg -p $NAME

xtuner copy-cfg $CONFIG $SAVE_PATH

xtuner train $CONFIG

xtuner convert pth_to_hf $CONFIG $PATH_TO_PTH_MODEL $SAVE_PATH_TO_HF_MODEL

▌资源选择

• 模块：【华为云】-【控制台】-【AI开发平台ModelArts】-【开发空间】-【Notebook】
• 镜像：pytorch_2.1.0-cann_8.0.rc1-py_3.9-euler_2.10.7-aarch64-snt9b

• 类型：【Ascend】
• 规格：【Ascend: 1*ascend-snt9b2|ARM: 24核 192GB】
• 磁盘规格：按需即可

▌安装环境

安装依赖

pip install einopspip install acceleratepip install dlinfer-ascendpip install deepspeedpip install loguru

安装 XTuner

git clone -b v0.1.23 https://github.com/InternLM/xtunergit clone -b v0.1.23 https://gitee.com/InternLM/xtuner #github 不行的话用这条cd xtuner

parser.add_argument(        '--device',        default='npu',        choices=('cuda', 'cpu', 'auto','npu'),        help='Indicate the device')  # choices里面添加一个'npu',也可以把default直接改成npu

pip install -e .

▌模型准备

· 创建目录

mkdir -p /home/ma-user/work/work_dir/cd /home/ma-user/work/work_dir/

· 下载模型

export HF_ENDPOINT=https://hf-mirror.com

huggingface-cli download --resume-download internlm/internlm2-chat-1_8b --local-dir /home/ma-user/work/model/internlm2-chat-1_8b

· 微调前的模型对话

import torchfrom transformers import AutoTokenizer, AutoModelForCausalLM

model_name_or_path = "/home/ma-user/work/model/internlm2-chat-1_8b"  # 模型所在的本地路径
tokenizer = AutoTokenizer.from_pretrained(model_name_or_path, trust_remote_code=True, device_map='npu')model = AutoModelForCausalLM.from_pretrained(model_name_or_path, trust_remote_code=True, torch_dtype=torch.bfloat16, device_map='npu')model = model.eval()
system_prompt = """You are an AI assistant whose name is InternLM (书生·浦语).- InternLM (书生·浦语) is a conversational language model that is developed by Shanghai AI Laboratory (上海人工智能实验室). It is designed to be helpful, honest, and harmless.- InternLM (书生·浦语) can understand and communicate fluently in the language chosen by the user such as English and 中文."""
messages = [(system_prompt, '')]
print("=============Welcome to InternLM chatbot, type 'exit' to exit.=============")
while True:    input_text = input("\nUser  >>> ")    input_text = input_text.replace(' ', '')    if input_text == "exit":        break
    length = 0    for response, _ in model.stream_chat(tokenizer, input_text, messages):        if response is not None:            print(response[length:], flush=True, end="")            length = len(response)

python cli_demo.py

· 微调

准备数据文件

cd /home/ma-user/work/work_dir/mkdir -p datastouch datas/assistant.json

创建一个脚本准备数据（xtuner_generate_assistant.py）

vim xtuner_generate_assistant.py

import json
# 设置用户的名字name = 'JeffDing同志'# 设置需要重复添加的数据次数n = 8000
# 初始化数据data = [    {"conversation": [{"input": "请介绍一下你自己", "output": "我是{}的小助手，内在是上海AI实验室书生·浦语的1.8B大模型哦".format(name)}]},    {"conversation": [{"input": "你在实战营做什么", "output": "我在这里帮助{}完成XTuner微调个人小助手的任务".format(name)}]}]
# 通过循环，将初始化的对话数据重复添加到data列表中for i in range(n):    data.append(data[0])    data.append(data[1])
# 将data列表中的数据写入到'datas/assistant.json'文件中with open('datas/assistant.json', 'w', encoding='utf-8') as f:    # 使用json.dump方法将数据以JSON格式写入文件    # ensure_ascii=False 确保中文字符正常显示    # indent=4 使得文件内容格式化，便于阅读    json.dump(data, f, ensure_ascii=False, indent=4)

- name = 'JeffDing同志'+ name = "你自己的名称"

python xtuner_generate_assistant.py

· 准备配置文件

xtuner list-cfg -p internlm2

· 复制配置文件

xtuner copy-cfg internlm2_chat_1_8b_qlora_alpaca_e3 .

· 修改配置文件

########################################################################                          PART 1  Settings                           ########################################################################- pretrained_model_name_or_path = 'internlm/internlm2-chat-1_8b'+ pretrained_model_name_or_path = '/home/ma-user/work/model/internlm2-chat-1_8b'
- alpaca_en_path = 'tatsu-lab/alpaca'+ alpaca_en_path = 'datas/assistant.json'
evaluation_inputs = [-    '请给我介绍五个上海的景点', 'Please tell me five scenic spots in Shanghai'+    '请介绍一下你自己', 'Please introduce yourself']########################################################################                      PART 3  Dataset & Dataloader                   ########################################################################alpaca_en = dict(type=process_hf_dataset,-   dataset=dict(type=load_dataset, path=alpaca_en_path),+   dataset=dict(type=load_dataset, path='json', data_files=dict(train=alpaca_en_path)),    tokenizer=tokenizer,    max_length=max_length,-   dataset_map_fn=alpaca_map_fn,+   dataset_map_fn=None,    template_map_fn=dict(type=template_map_fn_factory, template=prompt_template),    remove_unused_columns=True,    shuffle_before_pack=True,    pack_to_max_length=pack_to_max_length,    use_varlen_attn=use_varlen_attn)

########################################################################                      PART 2  Model & Tokenizer                      ########################################################################
- quantization_config=dict(-            type=BitsAndBytesConfig,-            load_in_4bit=True,-            load_in_8bit=False,-            llm_int8_threshold=6.0,-            llm_int8_has_fp16_weight=False,-            bnb_4bit_compute_dtype=torch.float16,-            bnb_4bit_use_double_quant=True,-            bnb_4bit_quant_type='nf4')

· 启动微调

xtuner train ./internlm2_chat_1_8b_qlora_alpaca_e3_copy.py

微调启动时 NPU 的使用情况，使用 npu-smi info 命令查看

· 模型格式转换

pth_file=`ls -t ./work_dirs/internlm2_chat_1_8b_qlora_alpaca_e3_copy/*.pth | head -n 1`xtuner convert pth_to_hf ./internlm2_chat_1_8b_qlora_alpaca_e3_copy.py ${pth_file} ./hf

· 模型合并

xtuner convert merge /home/ma-user/work/model/internlm2-chat-1_8b ./hf ./merged --max-shard-size 2GB --device npu

· 执行微调后的 demo

- model_name_or_path = "/home/ma-user/work/model/internlm2-chat-1_8b"  # 模型所在的本地路径+ model_name_or_path = "/home/ma-user/work/work_dir/merged"  # 模型所在的本地路径

python cli_demo.ppy