Ubuntu 20.04 Desktop数据盘模型制作，以 ChatGLM3-6B 为示例

环境配置

gcc & make

apt-get install -y gcc
apt-get install -y make

NVIDIA驱动

wget https://us.download.nvidia.cn/XFree86/Linux-x86_64/535.54.03/NVIDIA-Linux-x86_64-535.54.03.run
chmod +x NVIDIA-Linux-x86_64-535.54.03.run
./NVIDIA-Linux-x86_64-535.54.03.run --silent --no-questions

CUDA安装

wget https://developer.download.nvidia.com/compute/cuda/12.2.0/local_installers/cuda_12.2.0_535.54.03_linux.run
sudo sh cuda_12.2.0_535.54.03_linux.run

验证CUDA是否安装成功

ls /usr/local/cuda

如果显示

bin                EULA.txt  gds-12.2  libnvvp                  nsight-systems-2023.2.3  README  targets
compute-sanitizer  extras    include   nsight-compute-2023.2.0  nvml                     share   tools
DOCS               gds       lib64     nsightee_plugins         nvvm                     src     version.json

则说明已安装，接下来把路径添加进环境变量，由于刚安装完成添加进的是临时环境变量，所以很多情况下会发生开机后无法使用nvcc -V或nvcc --version指令查找CUDA是否安装的情况

vim ~/.bashrc

添加 CUDA 到 PATH 环境变量

export PATH=/usr/local/cuda/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH

使更改生效

source ~/.bashrc

再次检查 nvcc

nvcc -V

或

nvcc --version

nvidia-fabricmanager(A800以前的显卡不支持nvswitch，所以跳过此步骤)

wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64/nvidia-fabricmanager-535_535.54.03-1_amd64.deb
sudo dpkg -i ./nvidia-fabricmanager-535_535.54.03-1_amd64.deb

# 验证fabricmanager安装
nvidia-smi -pm 1
systemctl enable nvidia-fabricmanager
systemctl start nvidia-fabricmanager
systemctl status nvidia-fabricmanager
nvidia-smi topo -m

禁用更新

# 将后面的值全部修改为0，然后重启
vi /etc/apt/apt.conf.d/10periodic
vi /etc/apt/apt.conf.d/20auto-upgrades

pytorch安装

# 选择https://pytorch.org/
pip install -i https://pypi.tuna.tsinghua.edu.cn/simple some-package
pip3 install torch torchvision torchaudio

# 验证
python3
Python 3.8.10 (default, Nov 22 2023, 10:22:35)
[GCC 9.4.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import torch
>>> torch.cuda.is_available()
True

安装miniconda3

mkdir -p /useropt182/miniconda3
wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -O /useropt182/miniconda3/miniconda.sh
bash /useropt182/miniconda3/miniconda.sh -b -u -p /useropt182/miniconda3
rm -rf /useropt182/miniconda3/miniconda.sh
/useropt182/miniconda3/bin/conda init bash
source ~/.bashrc

下载模型源码 & 安装模型依赖

创建虚拟环境，一个模型一个环境

conda create --prefix=/useropt182/ChatGLM3-6B/env python=3.11
conda activate /useropt182/ChatGLM3-6B/env

下载源码

sudo git clone https://github.com/THUDM/ChatGLM3.git
pip config set global.index-url https://pypi.tuna.tsinghua.edu.cn/simple
pip3 install -r /useropt182/ChatGLM3/requirements.txt
pip3 install -r /useropt182/ChatGLM3/composite_demo/requirements.txt

下载模型

pip3 install 'huggingface_hub<0.22.0'

mirror

export HF_ENDPOINT=https://hf-mirror.com

单线程下载

export HF_HUB_ENABLE_HF_TRANSFER=0
#构建web可视化页面的工具
pip install streamlit
#多运行几次，确保不出现下载进度条为止
huggingface-cli download THUDM/chatglm3-6b --local-dir /useropt182/ChatGLM3-6B_hugginface

模型缓存 ，迁移到/useropt182/

mv /root/.cache /useropt182/
ln -s /useropt182/.cache  /root/

制作模型一健启动脚本/useropt182/main.sh

vim /useropt182/main.sh

# >>> conda initialize >>>
Model_Folder="useropt182"
S_Dir=`ls / | grep user | xargs -I {} echo "/{}"`
echo $S_Dir
T_Dir="/$Model_Folder"
ln -s $S_Dir $T_Dir

# !! Contents within this block are managed by 'conda init' !!
__conda_setup="$('/$Model_Folder/miniconda3/bin/conda' 'shell.bash' 'hook' 2> /dev/null)"
if [ $? -eq 0 ]; then
    eval "$__conda_setup"
else
    if [ -f "/$Model_Folder/miniconda3/etc/profile.d/conda.sh" ]; then
        . "/$Model_Folder/miniconda3/etc/profile.d/conda.sh"
    else
        export PATH="/$Model_Folder/miniconda3/bin:$PATH"
    fi
fi
unset __conda_setup
# <<< conda initialize <<<

# <<< Create a symlink<<< 
target_folder="/root/.cache"
file_to_move="/root/.cachebak"
new_location="/$Model_Folder/.cache"
symlink_name="/root/"

if [ ! -L "$target_folder" ]; then
    if [ -d "$target_folder" ]; then
        mv "$target_folder" "$file_to_move"
        echo "File $target_folder moved to $file_to_move"
    else
        echo "File $target_folder does not exist."
        #exit 1
    fi

    ln -s "$new_location" "$symlink_name"
    echo "Symlink $symlink_name created pointing to $new_location"
else
    echo "$target_folder is already a symlink."
fi
# <<< Create a symlink<<< 


conda activate /$Model_Folder/ChatGLM3-6B/env
export HF_ENDPOINT=https://hf-mirror.com
export MODEL_PATH=/$Model_Folder/ChatGLM3-6B_hugginface
streamlit run /$Model_Folder/ChatGLM3/composite_demo/main.py

运行模型

chmod +x /useropt182/main.sh
/useropt182/main.sh

验证

1. 用一个可用区，基础应用（镜像） ubuntu22.04_v3， 不挂载数据盘，创建实例
2. 将/useropt182 从制作镜像的实例中解绑
3. 数据盘/useropt182挂载到新购买的实例中
4. 直接运行/useropt182/main.sh ，ChatGLM3-6B模型运行正常，制作模型盘完成