XTuner概述
一个大语言模型微调工具箱。由 MMRazor 和 MMDeploy 联合开发。
支持的开源LLM (2023.11.01)
InternLM
Llama,Llama2
ChatGLM2,ChatGLM3
Qwen
Baichuan,Baichuan2
Zephyr
特色
傻瓜化: 以 配置文件 的形式封装了大部分微调场景,0基础的非专业人员也能一键开始微调。
轻量级: 对于 7B 参数量的LLM,微调所需的最小显存仅为 8GB : 消费级显卡,colab
微调原理
- 因此,你找到了一种叫 LoRA 的方法:只对玩具中的某些零件进行改动,而不是对整个玩具进行全面改动。
- 而 QLoRA 是 LoRA 的一种改进:如果你手里只有一把生锈的螺丝刀,也能改造你的玩具。
实践过程
安装
# 如果你是在 InternStudio 平台,则从本地 clone 一个已有 pytorch 2.0.1 的环境:
/root/share/install_conda_env_internlm_base.sh xtuner0.1.9
# 激活环境
conda activate xtuner0.1.9
# 进入家目录 (~的意思是 “当前用户的home路径”)
cd ~
# 创建版本文件夹并进入,以跟随本教程
mkdir xtuner019 && cd xtuner019
# 拉取 0.1.9 的版本源码
git clone -b v0.1.9 https://github.com/InternLM/xtuner
# 无法访问github的用户请从 gitee 拉取:
# git clone -b v0.1.9 https://gitee.com/Internlm/xtuner
# 进入源码目录
cd xtuner
# 从源码安装 XTuner
pip install -e '.[all]'
准备在 oasst1 数据集上微调 internlm-7b-chat
# 创建一个微调 oasst1 数据集的工作路径,进入
mkdir ~/ft-oasst1 && cd ~/ft-oasst1
微调
XTuner 提供多个开箱即用的配置文件,用户可以通过下列命令查看:
# 列出所有内置配置
xtuner list-cfg
拷贝一个配置文件到当前目录:
# xtuner copy-cfg ${CONFIG_NAME} ${SAVE_PATH}
本例中
cd ~/ft-oasst1
xtuner copy-cfg internlm_chat_7b_qlora_oasst1_e3 .
模型下载
直接复制模型
ln -s /share/temp/model_repos/internlm-chat-7b ~/ft-oasst1/
数据集下载
cd ~/ft-oasst1
# ...-guanaco 后面有个空格和英文句号啊
cp -r /root/share/temp/datasets/openassistant-guanaco .
修改配置文件
cd ~/ft-oasst1
vim internlm_chat_7b_qlora_oasst1_e3_copy.py
# 修改模型为本地路径
- pretrained_model_name_or_path = 'internlm/internlm-chat-7b'
+ pretrained_model_name_or_path = './internlm-chat-7b'
# 修改训练数据集为本地路径
- data_path = 'timdettmers/openassistant-guanaco'
+ data_path = './openassistant-guanaco'
开始微调
# 单卡
## 用刚才改好的config文件训练
xtuner train ./internlm_chat_7b_qlora_oasst1_e3_copy.py
# 多卡
NPROC_PER_NODE=${GPU_NUM} xtuner train ./internlm_chat_7b_qlora_oasst1_e3_copy.py
# 若要开启 deepspeed 加速,增加 --deepspeed deepspeed_zero2 即可
将得到的 PTH 模型转换为 HuggingFace 模型,即:生成 Adapter 文件夹
mkdir hf
export MKL_SERVICE_FORCE_INTEL=1
export MKL_THREADING_LAYER=GNU
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
部署与测试
将 HuggingFace adapter 合并到大语言模型
xtuner convert merge ./internlm-chat-7b ./hf ./merged --max-shard-size 2GB
与合并后的模型对话
xtuner chat ./merged --prompt-template internlm_chat
修改 cli_demo.py 中的模型路径
- model_name_or_path = "/root/model/Shanghai_AI_Laboratory/internlm-chat-7b"
+ model_name_or_path = "merged"
运行
python ./cli_demo.py
自定义微调
基于 InternLM-chat-7B 模型,用 MedQA 数据集进行微调,将其往医学问答领域对齐。
数据集
https://github.com/abachaa/Medication_QA_MedInfo2019
准备工作
将数据转为 XTuner 的数据格式(jsonL
目标格式:(.jsonL)
[{
"conversation":[
{
"system": "xxx",
"input": "xxx",
"output": "xxx"
}
]
},
{
"conversation":[
{
"system": "xxx",
"input": "xxx",
"output": "xxx"
}
]
}]
通过 python 脚本:将 .xlsx 中的 问题 和 回答 两列 提取出来,再放入 .jsonL 文件的每个 conversation 的 input 和 output 中。
这一步的 python 脚本可以请 ChatGPT 来完成。
Write a python file for me. using openpyxl. input file name is MedQA2019.xlsx
Step1: The input file is .xlsx. Exact the column A and column D in the sheet named "DrugQA" .
Step2: Put each value in column A into each "input" of each "conversation". Put each value in column D into each "output" of each "conversation".
Step3: The output file is .jsonL. It looks like:
[{
"conversation":[
{
"system": "xxx",
"input": "xxx",
"output": "xxx"
}
]
},
{
"conversation":[
{
"system": "xxx",
"input": "xxx",
"output": "xxx"
}
]
}]
Step4: All "system" value changes to "You are a professional, highly experienced doctor professor. You always provide accurate, comprehensive, and detailed answers based on the patients' questions."
生成的代码如下:
import openpyxl
import json
def process_excel_to_json(input_file, output_file):
# Load the workbook
wb = openpyxl.load_workbook(input_file)
# Select the "DrugQA" sheet
sheet = wb["DrugQA"]
# Initialize the output data structure
output_data = []
# Iterate through each row in column A and D
for row in sheet.iter_rows(min_row=2, max_col=4, values_only=True):
system_value = "You are a professional, highly experienced doctor professor. You always provide accurate, comprehensive, and detailed answers based on the patients' questions."
# Create the conversation dictionary
conversation = {
"system": system_value,
"input": row[0],
"output": row[3]
}
# Append the conversation to the output data
output_data.append({"conversation": [conversation]})
# Write the output data to a JSON file
with open(output_file, 'w', encoding='utf-8') as json_file:
json.dump(output_data, json_file, indent=4)
print(f"Conversion complete. Output written to {output_file}")
# Replace 'MedQA2019.xlsx' and 'output.jsonl' with your actual input and output file names
process_excel_to_json('MedQA2019.xlsx', 'output.jsonl')
执行脚本:
python xlsx2jsonl.py
**Tips:**如果数据集中有中文的话需要加上ensure_ascii=False
,否则中文会转换成一堆字符
划分训练集和测试集
my .jsonL file looks like:
[{
"conversation":[
{
"system": "xxx",
"input": "xxx",
"output": "xxx"
}
]
},
{
"conversation":[
{
"system": "xxx",
"input": "xxx",
"output": "xxx"
}
]
}]
Step1, read the .jsonL file.
Step2, count the amount of the "conversation" elements.
Step3, randomly split all "conversation" elements by 7:3. Targeted structure is same as the input.
Step4, save the 7/10 part as train.jsonl. save the 3/10 part as test.jsonl
代码如下:
import json
import random
def split_conversations(input_file, train_output_file, test_output_file):
# Read the input JSONL file
with open(input_file, 'r', encoding='utf-8') as jsonl_file:
data = json.load(jsonl_file)
# Count the number of conversation elements
num_conversations = len(data)
# Shuffle the data randomly
random.shuffle(data)
random.shuffle(data)
random.shuffle(data)
# Calculate the split points for train and test
split_point = int(num_conversations * 0.7)
# Split the data into train and test
train_data = data[:split_point]
test_data = data[split_point:]
# Write the train data to a new JSONL file
with open(train_output_file, 'w', encoding='utf-8') as train_jsonl_file:
json.dump(train_data, train_jsonl_file, indent=4)
# Write the test data to a new JSONL file
with open(test_output_file, 'w', encoding='utf-8') as test_jsonl_file:
json.dump(test_data, test_jsonl_file, indent=4)
print(f"Split complete. Train data written to {train_output_file}, Test data written to {test_output_file}")
# Replace 'input.jsonl', 'train.jsonl', and 'test.jsonl' with your actual file names
split_conversations('MedQA2019-structured.jsonl', 'MedQA2019-structured-train.jsonl', 'MedQA2019-structured-test.jsonl')
开始自定义微调
此时,我们重新建一个文件夹来玩“微调自定义数据集”
mkdir ~/ft-medqa && cd ~/ft-medqa
把前面下载好的internlm-chat-7b模型文件夹拷贝过来。
cp -r ~/ft-oasst1/internlm-chat-7b .
别忘了把自定义数据集,即几个 .jsonL,也传到服务器上。
git clone https://github.com/InternLM/tutorial
cp ~/tutorial/xtuner/MedQA2019-structured-train.jsonl .
准备配置文件
# 复制配置文件到当前目录
xtuner copy-cfg internlm_chat_7b_qlora_oasst1_e3 .
# 改个文件名
mv internlm_chat_7b_qlora_oasst1_e3_copy.py internlm_chat_7b_qlora_medqa2019_e3.py
# 修改配置文件内容
vim internlm_chat_7b_qlora_medqa2019_e3.py
减号代表要删除的行,加号代表要增加的行。
# 修改import部分
- from xtuner.dataset.map_fns import oasst1_map_fn, template_map_fn_factory
+ from xtuner.dataset.map_fns import template_map_fn_factory
# 修改模型为本地路径
- pretrained_model_name_or_path = 'internlm/internlm-chat-7b'
+ pretrained_model_name_or_path = './internlm-chat-7b'
# 修改训练数据为 MedQA2019-structured-train.jsonl 路径
- data_path = 'timdettmers/openassistant-guanaco'
+ data_path = 'MedQA2019-structured-train.jsonl'
# 修改 train_dataset 对象
train_dataset = dict(
type=process_hf_dataset,
- dataset=dict(type=load_dataset, path=data_path),
+ dataset=dict(type=load_dataset, path='json', data_files=dict(train=data_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)
启动
xtuner train internlm_chat_7b_qlora_medqa2019_e3.py --deepspeed deepspeed_zero2
将得到的 PTH 模型转换为 HuggingFace 模型,即:生成 Adapter 文件夹
mkdir hf
export MKL_SERVICE_FORCE_INTEL=1
export MKL_THREADING_LAYER=GNU
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
合并
将 HuggingFace adapter 合并到大语言模型
xtuner convert merge ./internlm-chat-7b ./hf ./merged --max-shard-size 2GB
用 MS-Agent 数据集 赋予 LLM 以 Agent 能力
MSAgent 数据集每条样本包含一个对话列表(conversations),其里面包含了 system、user、assistant 三种字段。其中:
-
system: 表示给模型前置的人设输入,其中有告诉模型如何调用插件以及生成请求
-
user: 表示用户的输入 prompt,分为两种,通用生成的prompt和调用插件需求的 prompt
-
assistant: 为模型的回复。其中会包括插件调用代码和执行代码,调用代码是要 LLM 生成的,而执行代码是调用服务来生成结果的
微调
xtuner 是从国内的 ModelScope 平台下载 MS-Agent 数据集,因此不用提前手动下载数据集文件。
# 准备工作
mkdir ~/ft-msagent && cd ~/ft-msagent
cp -r ~/ft-oasst1/internlm-chat-7b .
# 查看配置文件
xtuner list-cfg | grep msagent
# 复制配置文件到当前目录
xtuner copy-cfg internlm_7b_qlora_msagent_react_e3_gpu8 .
# 修改配置文件中的模型为本地路径
vim ./internlm_7b_qlora_msagent_react_e3_gpu8_copy.py
- pretrained_model_name_or_path = 'internlm/internlm-chat-7b'
+ pretrained_model_name_or_path = './internlm-chat-7b'
开始微调
xtuner train ./internlm_7b_qlora_msagent_react_e3_gpu8_copy.py --deepspeed deepspeed_zero2
使用
由于 msagent 的训练非常费时,大家如果想尽快把这个教程跟完,可以直接从 modelScope 拉取咱们已经微调好了的 Adapter。如下演示。
下载 Adapter
cd ~/ft-msagent
apt install git git-lfs
git lfs install
git lfs clone https://www.modelscope.cn/xtuner/internlm-7b-qlora-msagent-react.git
OK,现在目录应该长这样:
internlm_7b_qlora_msagent_react_e3_gpu8_copy.py
internlm-7b-qlora-msagent-react
internlm-chat-7b
work_dir(可有可无)`
有了这个在 msagent 上训练得到的Adapter,模型现在已经有 agent 能力了!就可以加 --lagent 以调用来自 lagent 的代理功能了!
添加 serper 环境变量
开始 chat 之前,还要加个 serper 的环境变量:
去 serper.dev 免费注册一个账号,生成自己的 api key。这个东西是用来给 lagent 去获取 google 搜索的结果的。等于是 serper.dev 帮你去访问 google,而不是从你自己本地去访问 google 了。
添加 serper api key 到环境变量:
export SERPER_API_KEY=你的API_KEY
启动
xtuner chat ./internlm-chat-7b --adapter internlm-7b-qlora-msagent-react --lagent
报错处理
xtuner chat 增加 --lagent 参数后,报错
TypeError: transfomers.modelsauto.auto factory. BaseAutoModelClass.from pretrained() got multiple values for keyword argument "trust remote code"
注释掉已安装包中的代码:
vim /root/xtuner019/xtuner/xtuner/tools/chat.py
作业
conda create --name personal_assistant --clone=/root/share/conda_envs/internlm-base
# 如果在其他平台:
# conda create --name personal_assistant python=3.10 -y
# 激活环境
conda activate personal_assistant
# 进入家目录 (~的意思是 “当前用户的home路径”)
cd ~
# 创建版本文件夹并进入,以跟随本教程
# personal_assistant用于存放本教程所使用的东西
mkdir /root/personal_assistant && cd /root/personal_assistant
mkdir /root/personal_assistant/xtuner019 && cd /root/personal_assistant/xtuner019
# 拉取 0.1.9 的版本源码
git clone -b v0.1.9 https://github.com/InternLM/xtuner
# 无法访问github的用户请从 gitee 拉取:
# git clone -b v0.1.9 https://gitee.com/Internlm/xtuner
# 进入源码目录
cd xtuner
# 从源码安装 XTuner
pip install -e '.[all]'
数据准备
mkdir -p /root/personal_assistant/data && cd /root/personal_assistant/data
在data目录下创建一个json文件personal_assistant.json作为本次微调所使用的数据集。json中内容可参考下方(复制粘贴n次做数据增广,数据量小无法有效微调,下面仅用于展示格式,下面也有生成脚本)
其中conversation表示一次对话的内容,input为输入,即用户会问的问题,output为输出,即想要模型回答的答案。
以下是一个python脚本,用于生成数据集。在data目录下新建一个generate_data.py文件,将以下代码复制进去,然后运行该脚本即可生成数据集。
import json
# 输入你的名字
name = 'Shengshenlan'
# 重复次数
n = 10000
data = [
{
"conversation": [
{
"input": "请做一下自我介绍",
"output": "我是{}的小助手,内在是上海AI实验室书生·浦语的7B大模型哦".format(name)
}
]
}
]
for i in range(n):
data.append(data[0])
with open('personal_assistant.json', 'w', encoding='utf-8') as f:
json.dump(data, f, ensure_ascii=False, indent=4)
配置准备
mkdir -p /root/personal_assistant/model/Shanghai_AI_Laboratory
cp -r /root/share/temp/model_repos/internlm-chat-7b /root/personal_assistant/model/Shanghai_AI_Laboratory
#创建用于存放配置的文件夹config并进入
mkdir /root/personal_assistant/config && cd /root/personal_assistant/config
xtuner copy-cfg internlm_chat_7b_qlora_oasst1_e3 .
修改拷贝后的文件internlm_chat_7b_qlora_oasst1_e3_copy.py
# Copyright (c) OpenMMLab. All rights reserved.
import torch
from bitsandbytes.optim import PagedAdamW32bit
from datasets import load_dataset
from mmengine.dataset import DefaultSampler
from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
LoggerHook, ParamSchedulerHook)
from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR
from peft import LoraConfig
from transformers import (AutoModelForCausalLM, AutoTokenizer,
BitsAndBytesConfig)
from xtuner.dataset import process_hf_dataset
from xtuner.dataset.collate_fns import default_collate_fn
from xtuner.dataset.map_fns import oasst1_map_fn, template_map_fn_factory
from xtuner.engine import DatasetInfoHook, EvaluateChatHook
from xtuner.model import SupervisedFinetune
from xtuner.utils import PROMPT_TEMPLATE
#######################################################################
# PART 1 Settings #
#######################################################################
# Model
pretrained_model_name_or_path = '/root/personal_assistant/model/Shanghai_AI_Laboratory/internlm-chat-7b'
# Data
data_path = '/root/personal_assistant/data/personal_assistant.json'
prompt_template = PROMPT_TEMPLATE.internlm_chat
max_length = 512
pack_to_max_length = True
# Scheduler & Optimizer
batch_size = 2 # per_device
accumulative_counts = 16
dataloader_num_workers = 0
max_epochs = 3
optim_type = PagedAdamW32bit
lr = 2e-4
betas = (0.9, 0.999)
weight_decay = 0
max_norm = 1 # grad clip
# Evaluate the generation performance during the training
evaluation_freq = 90
SYSTEM = ''
evaluation_inputs = [ '请介绍一下你自己', '请做一下自我介绍' ]
#######################################################################
# PART 2 Model & Tokenizer #
#######################################################################
tokenizer = dict(
type=AutoTokenizer.from_pretrained,
pretrained_model_name_or_path=pretrained_model_name_or_path,
trust_remote_code=True,
padding_side='right')
model = dict(
type=SupervisedFinetune,
llm=dict(
type=AutoModelForCausalLM.from_pretrained,
pretrained_model_name_or_path=pretrained_model_name_or_path,
trust_remote_code=True,
torch_dtype=torch.float16,
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')),
lora=dict(
type=LoraConfig,
r=64,
lora_alpha=16,
lora_dropout=0.1,
bias='none',
task_type='CAUSAL_LM'))
#######################################################################
# PART 3 Dataset & Dataloader #
#######################################################################
train_dataset = dict(
type=process_hf_dataset,
dataset=dict(type=load_dataset, path='json', data_files=dict(train=data_path)),
tokenizer=tokenizer,
max_length=max_length,
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)
train_dataloader = dict(
batch_size=batch_size,
num_workers=dataloader_num_workers,
dataset=train_dataset,
sampler=dict(type=DefaultSampler, shuffle=True),
collate_fn=dict(type=default_collate_fn))
#######################################################################
# PART 4 Scheduler & Optimizer #
#######################################################################
# optimizer
optim_wrapper = dict(
type=AmpOptimWrapper,
optimizer=dict(
type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
accumulative_counts=accumulative_counts,
loss_scale='dynamic',
dtype='float16')
# learning policy
# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md # noqa: E501
param_scheduler = dict(
type=CosineAnnealingLR,
eta_min=0.0,
by_epoch=True,
T_max=max_epochs,
convert_to_iter_based=True)
# train, val, test setting
train_cfg = dict(by_epoch=True, max_epochs=max_epochs, val_interval=1)
#######################################################################
# PART 5 Runtime #
#######################################################################
# Log the dialogue periodically during the training process, optional
custom_hooks = [
dict(type=DatasetInfoHook, tokenizer=tokenizer),
dict(
type=EvaluateChatHook,
tokenizer=tokenizer,
every_n_iters=evaluation_freq,
evaluation_inputs=evaluation_inputs,
system=SYSTEM,
prompt_template=prompt_template)
]
# configure default hooks
default_hooks = dict(
# record the time of every iteration.
timer=dict(type=IterTimerHook),
# print log every 100 iterations.
logger=dict(type=LoggerHook, interval=10),
# enable the parameter scheduler.
param_scheduler=dict(type=ParamSchedulerHook),
# save checkpoint per epoch.
checkpoint=dict(type=CheckpointHook, interval=1),
# set sampler seed in distributed evrionment.
sampler_seed=dict(type=DistSamplerSeedHook),
)
# configure environment
env_cfg = dict(
# whether to enable cudnn benchmark
cudnn_benchmark=False,
# set multi process parameters
mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
# set distributed parameters
dist_cfg=dict(backend='nccl'),
)
# set visualizer
visualizer = None
# set log level
log_level = 'INFO'
# load from which checkpoint
load_from = None
# whether to resume training from the loaded checkpoint
resume = False
# Defaults to use random seed and disable `deterministic`
randomness = dict(seed=None, deterministic=False)
微调启动
xtuner train /root/personal_assistant/config/internlm_chat_7b_qlora_oasst1_e3_copy.py
微调后参数转换/合并
# 创建用于存放Hugging Face格式参数的hf文件夹
mkdir /root/personal_assistant/config/work_dirs/hf
export MKL_SERVICE_FORCE_INTEL=1
# 配置文件存放的位置
export CONFIG_NAME_OR_PATH=/root/personal_assistant/config/internlm_chat_7b_qlora_oasst1_e3_copy.py
# 模型训练后得到的pth格式参数存放的位置
export PTH=/root/personal_assistant/config/work_dirs/internlm_chat_7b_qlora_oasst1_e3_copy/epoch_3.pth
# pth文件转换为Hugging Face格式后参数存放的位置
export SAVE_PATH=/root/personal_assistant/config/work_dirs/hf
# 执行参数转换
xtuner convert pth_to_hf $CONFIG_NAME_OR_PATH $PTH $SAVE_PATH
Merge模型参数
export MKL_SERVICE_FORCE_INTEL=1
export MKL_THREADING_LAYER='GNU'
# 原始模型参数存放的位置
export NAME_OR_PATH_TO_LLM=/root/personal_assistant/model/Shanghai_AI_Laboratory/internlm-chat-7b
# Hugging Face格式参数存放的位置
export NAME_OR_PATH_TO_ADAPTER=/root/personal_assistant/config/work_dirs/hf
# 最终Merge后的参数存放的位置
mkdir /root/personal_assistant/config/work_dirs/hf_merge
export SAVE_PATH=/root/personal_assistant/config/work_dirs/hf_merge
# 执行参数Merge
xtuner convert merge \
$NAME_OR_PATH_TO_LLM \
$NAME_OR_PATH_TO_ADAPTER \
$SAVE_PATH \
--max-shard-size 2GB
网页DEMO
安装依赖
pip install streamlit==1.24.0
下载项目
# 创建code文件夹用于存放InternLM项目代码
mkdir /root/personal_assistant/code && cd /root/personal_assistant/code
git clone https://github.com/InternLM/InternLM.git
将 /root/code/InternLM/web_demo.py 中 29 行和 33 行的模型路径更换为Merge后存放参数的路径 /root/personal_assistant/config/work_dirs/hf_merge
启动
streamlit run /root/personal_assistant/code/InternLM/web_demo.py --server.address 127.0.0.1 --server.port 6006