XTuner 大模型训练

本篇介绍如何使用 XTuner 进行大模型训练与微调。

环境配置

# 在终端执行这些命令
conda create --name xtuner0.1.9 python=3.10 -y
conda activate xtuner0.1.0
pip install ipykernel
python -m ipykernel install --user --name xtuner --display-name xtuner

刷新 notebook，选择新内核 xtuner

1 2	%cd ~ %mkdir xtuner019 && cd xtuner019

1	!git clone -b v0.1.9 https://github.com/InternLM/xtuner

1	%cd xtuner

1	%pip install -e '.[all]'

1 2	%mkdir ~/ft-oasst1 %cd ~/ft-oasst1

1	import os, sys

1 2	PATH = os.environ['PATH'] PATH

1 2	basedir = os.path.dirname(os.path.dirname(sys.exec_prefix)) basedir

%env CONDA_EXE={os.path.join(basedir, 'bin/conda')}
%env CONDA_PREFIX={sys.exec_prefix}
%env CONDA_PYTHON_EXE={os.path.join(basedir, 'bin/python')}
%env PATH={os.path.join(sys.exec_prefix, 'bin')}:$PATH

微调

准备配置文件

1	!xtuner list-cfg

1	%cd ~/ft-oasst1

%pwd

1	!xtuner copy-cfg internlm_chat_7b_qlora_oasst1_e3 .

!ls

模型下载

1	%cp -r /root/share/temp/model_repos/internlm-chat-7b/ ~/ft-oasst1/

1	!ls internlm-chat-7b/

数据集下载

1	%cd ~/ft-oasst1

1	%cp -r /root/share/temp/datasets/openassistant-guanaco/ .

1	!ls openassistant-guanaco/

1	!tree ~/ft-oasst1

修改配置文件

1	!sed -i "s#pretrained_model_name_or_path = 'internlm/internlm-chat-7b'#pretrained_model_name_or_path = './internlm-chat-7b'#g" /root/ft-oasst1/internlm_chat_7b_qlora_oasst1_e3_copy.py

1	!sed -i "s#data_path = 'timdettmers/openassistant-guanaco'#data_path = './openassistant-guanaco'#g" /root/ft-oasst1/internlm_chat_7b_qlora_oasst1_e3_copy.py

1	!sed -i "s#max_epochs = 3#max_epochs = 1#g" /root/ft-oasst1/internlm_chat_7b_qlora_oasst1_e3_copy.py

开始微调

1 2	# # 常规训练，耗时长 # !xtuner train /root/ft-oasst1/internlm_chat_7b_qlora_oasst1_e3_copy.py

1 2	# 加速训练，耗时相对较短 !xtuner train /root/ft-oasst1/internlm_chat_7b_qlora_oasst1_e3_copy.py --deepspeed deepspeed_zero2

1	!tree /root/ft-oasst1

将 PTH 转为 HuggingFace 模型

即生成 Adapter 文件夹

1 2	%cd /root/ft-oasst1/ %mkdir hf

1	%env MKL_SERVICE_FORCE_INTEL=1

1	!xtuner convert pth_to_hf ./internlm_chat_7b_qlora_oasst1_e3_copy.py ./work_dirs/internlm_chat_7b_qlora_oasst1_e3_copy/epoch_1.pth ./hf

!tree .

此时，hf 文件夹即为我们平时所理解的 LoRA 模型文件，即 Adapter

部署

合并 HuggingFace Adapter 到大模型

1	%cd /root/ft-oasst1/

1	%mkdir merged

1	!xtuner convert merge ./internlm-chat-7b/ ./hf/ ./merged --max-shard-size 2GB

与合并后的大模型对话

1	!xtuner chat ./merged --prompt-template internlm_chat --bits 4

%ls

%%writefile cli_demo.py
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM


model_name_or_path = "./merged"

tokenizer = AutoTokenizer.from_pretrained(model_name_or_path, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(model_name_or_path, trust_remote_code=True, torch_dtype=torch.bfloat16, device_map='auto')
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("User  >>> ")
    input_text.replace(' ', '')
    if input_text == "exit":
        break
    response, history = model.chat(tokenizer, input_text, history=messages)
    messages.append((input_text, response))
    print(f"robot >>> {response}")

1	!python cli_demo.py

自定义微调

1 2	%mkdir ~/ft-medqa %cd ~/ft-medqa

1	%cp -r /root/ft-oasst1/internlm-chat-7b/ .

1	!git clone https://github.com/InternLM/tutorial

1	%cp /root/ft-medqa/tutorial/xtuner/MedQA2019-structured-train.jsonl .

%ls

准备配置文件

1	!xtuner copy-cfg internlm_chat_7b_qlora_oasst1_e3 .

1	%mv internlm_chat_7b_qlora_oasst1_e3_copy.py internlm_chat_7b_qlora_medqa2019_e3.py

1	!sed -i "s#from xtuner.dataset.map_fns import oasst1_map_fn, template_map_fn_factory#from xtuner.dataset.map_fns import template_map_fn_factory#g" internlm_chat_7b_qlora_medqa2019_e3.py

1	!sed -i "s#pretrained_model_name_or_path = 'internlm/internlm-chat-7b'#pretrained_model_name_or_path = './internlm-chat-7b'#g" internlm_chat_7b_qlora_medqa2019_e3.py

1	!sed -i "s#data_path = 'timdettmers/openassistant-guanaco'#data_path = 'MedQA2019-structured-train.jsonl'#g" internlm_chat_7b_qlora_medqa2019_e3.py

1	!sed -i "s#dataset=dict(type=load_dataset, path=data_path),#dataset=dict(type=load_dataset, path='json', data_files=dict(train=data_path)),#g" internlm_chat_7b_qlora_medqa2019_e3.py

1	!sed -i "s#dataset_map_fn=oasst1_map_fn,#dataset_map_fn=None,#g" internlm_chat_7b_qlora_medqa2019_e3.py

1	!xtuner train internlm_chat_7b_qlora_medqa2019_e3.py --deepspeed deepspeed_zero2

1	!tree work_dirs/

把 PTH 转为 HuggingFace

%ls

%mkdir hf

1	%env MKL_SERVICE_FORCE_INTEL=1

1	!xtuner convert pth_to_hf ./internlm_chat_7b_qlora_medqa2019_e3.py ./work_dirs/internlm_chat_7b_qlora_medqa2019_e3/epoch_3.pth/ ./hf

!tree hf

1	%mkdir merged

1	!xtuner convert merge ./internlm-chat-7b/ ./hf ./merged/ --max-shard-size 2GB

1	!xtuner chat ./merged/ --prompt-template internlm_chat --bits 4

使用 XTuner 微调 InternLM-Chat 模型，使得其能够知道它的主人是谁

1	!conda create --name personal_assistant python=3.10 -y

1	!which conda

1	!/opt/miniconda/bin/conda/envs/personal_assistant/bin/ pip install ipykernel && ipykernel install --user --name personal_assistant --display-name personal_assistant