硬件资源说明
本教程基于GPU 3090的服务器
| 资源类型 | 型号 | 核心指标 |
| CPU | Intel(R) Xeon(R) Bronze 3204 CPU @ 1.90GHz | 12核 |
| 内存 | / | 125Gi |
| GPU | NVIDIA GeForce RTX 3090 | 24G显存 |
注意：接下来的部分命令需要使用科学上网，需要事先配置好。
安装docker
参考docker的官方文档：Install | Docker Docs
这里简单列举一下常见系统的安装命令
对于ubuntu系统：
sudo apt install docker.io
sudo systemctl enable docker对于centos系统：
sudo yum install -y yum-utils
sudo yum-config-manager --add-repo https://download.docker.com/linux/centos/docker-ce.repo
sudo yum install docker-ce docker-ce-cli containerd.io docker-buildx-plugin docker-compose-plugin
sudo systemctl start docker
sudo systemctl enable docker这里默认安装的docker版本为24.0.7，docker版本的差异不会影响整个测试的流程，不过还是建议使用和本教程相同的版本
安装NVIDIA Container Toolkit
安装该工具主要是为了docker可以挂载Nvidia GPU设备。
参考：Installing the NVIDIA Container Toolkit — NVIDIA Container Toolkit 1.16.2 documentation
安装方式有很多，这里仅列举比较常用的ubuntu和centos安装方式：
在线安装
对于ubuntu系统
sudo apt install curl -y
curl -fsSL https://nvidia.github.io/libnvidia-container/gpgkey | sudo gpg --dearmor -o /usr/share/keyrings/nvidia-container-toolkit-keyring.gpg && curl -s -L https://nvidia.github.io/libnvidia-container/stable/deb/nvidia-container-toolkit.list | sed 's#deb https://#deb [signed-by=/usr/share/keyrings/nvidia-container-toolkit-keyring.gpg] https://#g' | sudo tee /etc/apt/sources.list.d/nvidia-container-toolkit.list
sudo apt-get update
sudo apt-get install -y nvidia-container-toolkit对于Centos系统
sudo yum install curl -y
curl -s -L https://nvidia.github.io/libnvidia-container/stable/rpm/nvidia-container-toolkit.repo | sudo tee /etc/yum.repos.d/nvidia-container-toolkit.repo
sudo yum install -y nvidia-container-toolkit离线安装
参考开源项目：libnvidia-container/stable/ubuntu18.04/amd64 at gh-pages · NVIDIA/libnvidia-container · GitHub
对于Ubuntu系统
下载以下安装包：
# 分别打开如下链接
https://github.com/NVIDIA/libnvidia-container/blob/gh-pages/stable/ubuntu18.04/amd64/libnvidia-container1_1.9.0-1_amd64.deb
https://github.com/NVIDIA/libnvidia-container/blob/gh-pages/stable/ubuntu18.04/amd64/libnvidia-container-tools_1.9.0-1_amd64.deb
https://github.com/NVIDIA/libnvidia-container/blob/gh-pages/stable/ubuntu18.04/amd64/nvidia-container-toolkit_1.9.0-1_amd64.deb
https://github.com/NVIDIA/libnvidia-container/blob/gh-pages/stable/ubuntu18.04/amd64/nvidia-docker2_2.10.0-1_all.deb点击下载
下载好的安装包执行如下命令即可安装：
dpkg -i libnvidia-container1_1.9.0-1_amd64.deb
dpkg -i libnvidia-container-tools_1.9.0-1_amd64.deb
dpkg -i nvidia-container-toolkit_1.9.0-1_amd64.deb
dpkg -i nvidia-docker2_2.10.0-1_all.deb安装完成后需要重启docker以生效
systemctl restart docker对于Centos系统
参考ubuntu的方式寻找适合自己系统的rpm包进行安装
这里假设是Centos7，则在如下页面寻找libnvidia-container/stable/centos7/x86_64 at gh-pages · NVIDIA/libnvidia-container · GitHub
https://github.com/NVIDIA/libnvidia-container/blob/gh-pages/stable/centos7/x86_64/libnvidia-container-devel-1.9.0-1.x86_64.rpm
https://github.com/NVIDIA/libnvidia-container/blob/gh-pages/stable/centos7/x86_64/libnvidia-container-tools-1.9.0-1.x86_64.rpm
https://github.com/NVIDIA/libnvidia-container/blob/gh-pages/stable/centos7/x86_64/nvidia-container-toolkit-1.9.0-1.x86_64.rpm
https://github.com/NVIDIA/libnvidia-container/blob/gh-pages/stable/centos7/x86_64/nvidia-docker2-2.10.0-1.noarch.rpm下载好后安装命令为
rpm -ivh libnvidia-container-devel-1.9.0-1.x86_64.rpm
rpm -ivh libnvidia-container-tools-1.9.0-1.x86_64.rpm
rpm -ivh nvidia-container-toolkit-1.9.0-1.x86_64.rpm
rpm -ivh nvidia-docker2-2.10.0-1.noarch.rpm安装完成后需要重启docker以生效
systemctl restart docker镜像制作
下载镜像
这里在Nvidia NGC上下载一个比较干净的镜像，模型开发需要的pip包我们自己来装「比如xtuner」。
镜像仓库：dockerhub
对于NVIDIA的GPU，很多基础的镜像名字（编译安装好各类软件包的环境）可以在NGC上找到，网址为：Data Science, Machine Learning, AI, HPC Containers | NVIDIA NGC
下载什么版本镜像来匹配自己的cuda版本呢？可以通过执行nvidia-smi来查看，我这台服务器是cuda 12.4，如下图所示，那我们最好也是下载cuda12.x对应的docker镜像（不需要严格对齐版本）。
找到对应的镜像名字：
这里选择的cuda版本和机器安装的cuda版本一致，都是12.x；此外选择一个自己常用的操作系统版本，比如这里我选择的是ubuntu22.04，值得一提的是，这里的系统版本和自己机器上的系统版本没有必然联系，机器上只要可以运行docker即可。
执行下载镜像的命令，这个过程的耗时和各自的网络速度有关：
# 刚刚复制的镜像名字放在docker pull后面即可
docker pull nvcr.io/nvidia/pytorch:23.10-py3
如果镜像无法下载，需要配置科学上网。
启动镜像
启动镜像时有一些参数如下。执行如下命令后，就已经启动了镜像并进入了容器内部，接下来执行的所有命令都是在容器内部进行的。
# --gpus=all 表示将机器上的所有GPU挂载到容器内
# -v /data2/:/datasets表示将本地的/data2 目录挂载到容器的/datasets目录
docker run -it --gpus=all -v /data2/:/datasets nvcr.io/nvidia/pytorch:23.10-py3 bash
开始安装环境
现在在容器内，开始执行环境安装命令：
apt-get update
apt-get install libaio-dev -y
pip install xtuner[all]==0.1.18 -i https://mirrors.ustc.edu.cn/pypi/web/simple
pip3 install "fschat[model_worker,webui]" -i https://mirrors.ustc.edu.cn/pypi/web/simple
pip3 install jupyterlab -i https://mirrors.ustc.edu.cn/pypi/web/simple
pip install opencv-python==4.8.0.74 -i https://mirrors.ustc.edu.cn/pypi/web/simple
pip uninstall transformer_engine flash-attn
pip install datasets==2.18.0 安装过程如下「这个过程大约持续10～20分钟，主要看机器的网速」：
保存镜像
将制作好镜像进行保存，上面的容器内部截图中显示了一个id，使用这个容器id即可保存安装好的环境。
#xtuner-fastchat-cuda122:23.10-py3表示给镜像取一个新的名字「命名形式为xxx:yyy」
docker commit dd129233d4f5 xtuner-fastchat-cuda122:23.10-py3Tips：比较高级的用法是使用Dockerfile进行镜像制作，参考：Docker Dockerfile | 菜鸟教程
运行微调测试
启动环境
# 用新的镜像启动容器
# 参数--shm-size 32G表示容器运行环境添加了32G的共享内存
# 参数-p 8888:8888表示将容器内部的8888端口和主机上的8888端口进行映射
docker run -it --shm-size 32G --gpus=all --network=host -v /mnt/nfs_share/data2/:/code/ -v /mnt/nfs_share/data2/:/dataset xtuner-fastchat-cuda122:23.10-py3 bash# 进入容器后执行如下命令开启jupyter-lab
cd /code
jupyter-lab --allow-root --ip='*' --NotebookApp.token='' --NotebookApp.password='' &在自己的浏览器中新建一个页面，输入http://127.0.0.1:8888/lab即可进入jupyterlab界面
运行测试
我们还是以微调qwen0.5B作为例子，展示下代码无缝迁移：
step1: 模型下载
export HF_ENDPOINT=https://hf-mirror.com
huggingface-cli download Qwen/Qwen1.5-0.5B-Chat --resume-download --local-dir-use-symlinks False --local-dir /dataset/Qwen1.5-0.5B-Chat
数据下载
export HF_ENDPOINT=https://hf-mirror.com
huggingface-cli download tatsu-lab/alpaca --repo-type dataset --revision main --local-dir-use-symlinks False --local-dir /dataset/datasets/tatsu-lab___alpaca/data/算法文件（qwen1_5_0_5b_chat_full_alpaca_e3.py）
# Copyright (c) OpenMMLab. All rights reserved.
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, LinearLR
from torch.optim import AdamW
from transformers import AutoModelForCausalLM, AutoTokenizerfrom xtuner.dataset import process_hf_dataset
from xtuner.dataset.collate_fns import default_collate_fn
from xtuner.dataset.map_fns import alpaca_map_fn, template_map_fn_factory
from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,VarlenAttnArgsToMessageHubHook)
from xtuner.engine.runner import TrainLoop
from xtuner.model import SupervisedFinetune
from xtuner.parallel.sequence import SequenceParallelSampler
from xtuner.utils import PROMPT_TEMPLATE, SYSTEM_TEMPLATE#######################################################################
# PART 1 Settings #
#######################################################################
# Model
pretrained_model_name_or_path = '/dataset/Qwen1.5-0.5B-Chat/'
use_varlen_attn = False# Data
alpaca_en_path = '/dataset/datasets/tatsu-lab___alpaca/data/'
prompt_template = PROMPT_TEMPLATE.qwen_chat
max_length = 2048
pack_to_max_length = False# parallel
sequence_parallel_size = 1# Scheduler & Optimizer
batch_size = 1 # per_device
accumulative_counts = 16
accumulative_counts *= sequence_parallel_size
dataloader_num_workers = 0
max_epochs = 3
optim_type = AdamW
lr = 2e-5
betas = (0.9, 0.999)
weight_decay = 0
max_norm = 1 # grad clip
warmup_ratio = 0.03# Save
save_steps = 500
save_total_limit = 2 # Maximum checkpoints to keep (-1 means unlimited)# Evaluate the generation performance during the training
evaluation_freq = 500
SYSTEM = SYSTEM_TEMPLATE.alpaca
evaluation_inputs = ['请给我介绍五个上海的景点', 'Please tell me five scenic spots in Shanghai'
]#######################################################################
# 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,use_varlen_attn=use_varlen_attn,llm=dict(type=AutoModelForCausalLM.from_pretrained,pretrained_model_name_or_path=pretrained_model_name_or_path,trust_remote_code=True))#######################################################################
# PART 3 Dataset & Dataloader #
#######################################################################
alpaca_en = dict(type=process_hf_dataset,dataset=dict(type=load_dataset, path=alpaca_en_path),tokenizer=tokenizer,max_length=max_length,dataset_map_fn=alpaca_map_fn,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)sampler = SequenceParallelSampler \if sequence_parallel_size > 1 else DefaultSamplertrain_dataloader = dict(batch_size=batch_size,num_workers=dataloader_num_workers,dataset=alpaca_en,sampler=dict(type=sampler, shuffle=True),collate_fn=dict(type=default_collate_fn, use_varlen_attn=use_varlen_attn))#######################################################################
# 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=LinearLR,start_factor=1e-5,by_epoch=True,begin=0,end=warmup_ratio * max_epochs,convert_to_iter_based=True),dict(type=CosineAnnealingLR,eta_min=0.0,by_epoch=True,begin=warmup_ratio * max_epochs,end=max_epochs,convert_to_iter_based=True)
]# train, val, test setting
train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)#######################################################################
# 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)
]if use_varlen_attn:custom_hooks += [dict(type=VarlenAttnArgsToMessageHubHook)]# configure default hooks
default_hooks = dict(# record the time of every iteration.timer=dict(type=IterTimerHook),# print log every 10 iterations.logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=1),# enable the parameter scheduler.param_scheduler=dict(type=ParamSchedulerHook),# save checkpoint per `save_steps`.checkpoint=dict(type=CheckpointHook,by_epoch=False,interval=save_steps,max_keep_ckpts=save_total_limit),# set sampler seed in distributed evrionment.sampler_seed=dict(type=DistSamplerSeedHook),
)# configure environment
env_cfg = dict(# whether to enable cudnn benchmarkcudnn_benchmark=False,# set multi process parametersmp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),# set distributed parametersdist_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)# set log processor
log_processor = dict(by_epoch=False)
算法上传到自己本地启动的notebook中：
开始微调
mkdir -p /userhome/xtuner-workdir
NPROC_PER_NODE=1 xtuner train qwen1_5_0_5b_chat_full_alpaca_e3_copy.py --work-dir /userhome/xtuner-workdir --deepspeed deepspeed_zero3_offload如果显存不够可以考虑使用lora或者qlora
3090的显存24G，全量参数微调不够用，需要开启lora训练
运行模型部署
打开4个终端，分别按顺序执行如下命令
多机多卡训练教程以及视频 mp4 下载链接：点击这里下载
python -m fastchat.serve.controller --host 0.0.0.0
python -m fastchat.serve.model_worker --model-path /dataset/Qwen1.5-0.5B-Chat/ --host 0.0.0.0 --num-gpus 1 --max-gpu-memory 20GiB
python -m fastchat.serve.openai_api_server --host 0.0.0.0
curl http://localhost:8000/v1/completions \-H "Content-Type: application/json" \-d '{"model": "Qwen1.5-0.5B-Chat","prompt": "Once upon a time","max_tokens": 41,"temperature": 0.5}'
调用结果返回正常：
