【InternLM 实战营笔记】XTuner 大模型单卡低成本微调实战

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

开始微调

xtuner train ./internlm_chat_7b_qlora_oasst1_e3_copy.py

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

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

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

运行效果

在这里插入图片描述