行业知识
【InternLM 实战营笔记】XTuner 大模型单卡低成本微调实战
2026-10-20 08:32:14
来源：互联网转载
XTuner概述
一个大��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 XTunerpip 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-oasst1xtuner 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-oasst1vim 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 hfexport MKL_SERVICE_FORCE_INTEL=1export MKL_THREADING_LAYER=GNUxtuner 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_assistantmkdir /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 XTunerpip 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 = 10000data = [ { "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_Laboratorycp -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/configxtuner 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 torchfrom bitsandbytes.optim import PagedAdamW32bitfrom datasets import load_datasetfrom mmengine.dataset import DefaultSamplerfrom mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook, LoggerHook, ParamSchedulerHook)from mmengine.optim import AmpOptimWrapper, CosineAnnealingLRfrom peft import LoraConfigfrom transformers import (AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig)from xtuner.dataset import process_hf_datasetfrom xtuner.dataset.collate_fns import default_collate_fnfrom xtuner.dataset.map_fns import oasst1_map_fn, template_map_fn_factoryfrom xtuner.engine import DatasetInfoHook, EvaluateChatHookfrom xtuner.model import SupervisedFinetunefrom xtuner.utils import PROMPT_TEMPLATE######################################################################## PART 1 Settings ######################################################################### Modelpretrained_model_name_or_path = '/root/personal_assistant/model/Shanghai_AI_Laboratory/internlm-chat-7b'# Datadata_path = '/root/personal_assistant/data/personal_assistant.json'prompt_template = PROMPT_TEMPLATE.internlm_chatmax_length = 512pack_to_max_length = True# Scheduler & Optimizerbatch_size = 2 # per_deviceaccumulative_counts = 16dataloader_num_workers = 0max_epochs = 3optim_type = PagedAdamW32bitlr = 2e-4betas = (0.9, 0.999)weight_decay = 0max_norm = 1 # grad clip# Evaluate the generation performance during the trainingevaluation_freq = 90SYSTEM = ''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 ######################################################################### optimizeroptim_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: E501param_scheduler = dict( type=CosineAnnealingLR, eta_min=0.0, by_epoch=True, T_max=max_epochs, convert_to_iter_based=True)# train, val, test settingtrain_cfg = dict(by_epoch=True, max_epochs=max_epochs, val_interval=1)######################################################################## PART 5 Runtime ######################################################################### Log the dialogue periodically during the training process, optionalcustom_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 hooksdefault_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 environmentenv_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 visualizervisualizer = None# set log levellog_level = 'INFO'# load from which checkpointload_from = None# whether to resume training from the loaded checkpointresume = 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/hfexport 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/codegit 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
