Skip to content

필사 모드: AI 开发环境完全指南:从 GPU 服务器搭建到 Jupyter、VS Code、Docker

中文
0%
정확도 0%
💡 왼쪽 원문을 읽으면서 오른쪽에 따라 써보세요. Tab 키로 힌트를 받을 수 있습니다.

概述

AI 开发环境搭建,是项目成功的一半。没有恰当的环境,实验的可复现性、团队协作、快速迭代都无从谈起。本指南涵盖 AI/ML 开发者需要掌握的全部环境配置内容 — 从 GPU 服务器的初始设置,到日常开发工作流。

内容包括:在 Ubuntu 上安装 CUDA、管理 Python 环境、像专家一样使用 JupyterLab 和 VS Code,以及用 Docker 容器搭建可复现的实验环境,逐一分步说明。


1. AI 开发环境要求

1.1 硬件推荐

GPU(最优先)

在 AI 研究中,GPU 不是可选项,而是必需品。按用途推荐如下:

  • 入门/个人研究:NVIDIA RTX 4080/4090(16-24GB 显存)
  • 团队共享服务器:NVIDIA A100 40GB 或 80GB
  • 大规模 LLM 训练:NVIDIA H100 或 H200(80GB+ 显存)
  • 云端:AWS p3/p4d/p5、GCP A100/H100、Lambda Labs

CPU

GPU 训练时,CPU 主要负责数据预处理和加载。

  • 最低:8 核 16 线程(Intel Core i9 或 AMD Ryzen 9)
  • 推荐:16 核以上(AMD Threadripper、Intel Xeon)
  • DataLoader worker 数量 = CPU 核心数的一半较为合适

内存

  • 最低:32GB
  • 推荐:64GB(显存的 4 倍以上)
  • 大规模 NLP:128GB 以上(分词、数据加载)

存储

/home/user/NVMe SSD(操作系统、代码、环境)
/data/NVMe SSD 或高速 HDD(训练数据)
/models/HDDNAS(模型检查点)
  • 操作系统及软件包:NVMe SSD 500GB 以上
  • 数据集:NVMe SSD(I/O 密集型训练)或高性能 HDD
  • 检查点:HDD 4TB 以上(性价比高的大容量存储)

1.2 操作系统选择

强烈推荐 Ubuntu 22.04 LTS

# 查看 Ubuntu 版本
lsb_release -a

# 示例输出:
# Ubuntu 22.04.3 LTS (Jammy Jellyfish)

选择 Ubuntu 的理由如下:

  • NVIDIA 官方支持(第一时间提供最新驱动、CUDA、cuDNN 版本)
  • 丰富的 AI/ML 社区文档与软件包
  • 与 Docker、Kubernetes 兼容性最佳
  • 5 年 LTS 支持,服务器运行稳定

macOS(M1/M2/M3)

Apple Silicon 通过 Metal Performance Shaders 支持 GPU 加速。

# 确认 MPS 加速可用性(PyTorch)
python -c "import torch; print(torch.backends.mps.is_available())"

1.3 云端 vs 本地开发

项目本地服务器云端
初始成本高(硬件)
运营成本低(电费)高(按小时计费)
等待时间实例启动时间
扩展性有限无限
数据安全取决于策略
推荐用途持续性研究一次性大规模实验

2. NVIDIA GPU 驱动与 CUDA 安装

2.1 安装 nvidia-driver(Ubuntu)

# 1. 移除已有的 NVIDIA 软件包
sudo apt-get purge nvidia*
sudo apt autoremove

# 2. 查看可用驱动
ubuntu-drivers devices

# 3. 自动安装推荐驱动
sudo ubuntu-drivers autoinstall

# 或安装指定版本
sudo apt install nvidia-driver-535

# 4. 重启
sudo reboot

# 5. 确认安装
nvidia-smi

nvidia-smi 输出示例:

+-----------------------------------------------------------------------------+
| NVIDIA-SMI 535.154.05   Driver Version: 535.154.05   CUDA Version: 12.2    |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  NVIDIA A100-SXM4-40GB  Off| 00000000:00:04.0 Off |                    0 |
| N/A   34C    P0    56W / 400W |   1024MiB / 40960MiB |      0%      Default |
+-----------------------------------------------------------------------------+

2.2 安装 CUDA Toolkit

在 NVIDIA 官网生成适合你环境的安装命令。

# 安装 CUDA 12.2(以 Ubuntu 22.04 为例)
wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-keyring_1.1-1_all.deb
sudo dpkg -i cuda-keyring_1.1-1_all.deb
sudo apt-get update
sudo apt-get -y install cuda-toolkit-12-2

# 设置环境变量(追加到 ~/.bashrc 或 ~/.zshrc)
echo 'export PATH=/usr/local/cuda/bin:$PATH' >> ~/.bashrc
echo 'export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH' >> ~/.bashrc
source ~/.bashrc

# 确认 CUDA 版本
nvcc --version

2.3 安装 cuDNN

# 需要 NVIDIA Developer 账号
# 下载 cuDNN 后:

# 以 Ubuntu 22.04、CUDA 12.x 为例
sudo dpkg -i cudnn-local-repo-ubuntu2204-8.9.7.29_1.0-1_amd64.deb
sudo cp /var/cudnn-local-repo-ubuntu2204-8.9.7.29/cudnn-local-*-keyring.gpg /usr/share/keyrings/
sudo apt-get update
sudo apt-get install libcudnn8 libcudnn8-dev libcudnn8-samples

# 确认 cuDNN 版本
cat /usr/include/cudnn_version.h | grep CUDNN_MAJOR -A 2

2.4 验证安装

# verify_gpu.py
import subprocess
import sys

def check_nvidia_smi():
    result = subprocess.run(['nvidia-smi'], capture_output=True, text=True)
    if result.returncode == 0:
        print("nvidia-smi 运行正常")
        # 提取 GPU 信息
        lines = result.stdout.split('\n')
        for line in lines:
            if 'NVIDIA' in line and 'Driver' in line:
                print(f"  {line.strip()}")
    else:
        print("nvidia-smi 出错:", result.stderr)

def check_pytorch_cuda():
    try:
        import torch
        print(f"\nPyTorch 版本:{torch.__version__}")
        print(f"CUDA 可用:{torch.cuda.is_available()}")
        if torch.cuda.is_available():
            print(f"CUDA 版本:{torch.version.cuda}")
            print(f"cuDNN 版本:{torch.backends.cudnn.version()}")
            print(f"GPU 数量:{torch.cuda.device_count()}")
            for i in range(torch.cuda.device_count()):
                props = torch.cuda.get_device_properties(i)
                print(f"  GPU {i}: {props.name} ({props.total_memory / 1024**3:.1f}GB)")

            # 简单的张量运算测试
            x = torch.randn(1000, 1000).cuda()
            y = torch.randn(1000, 1000).cuda()
            z = x @ y
            print(f"GPU 张量运算测试:成功(shape: {z.shape})")
    except ImportError:
        print("未安装 PyTorch")

def check_tensorflow_gpu():
    try:
        import tensorflow as tf
        print(f"\nTensorFlow 版本:{tf.__version__}")
        gpus = tf.config.list_physical_devices('GPU')
        print(f"检测到的 GPU:{len(gpus)}个")
        for gpu in gpus:
            print(f"  {gpu}")
    except ImportError:
        print("未安装 TensorFlow")

if __name__ == "__main__":
    check_nvidia_smi()
    check_pytorch_cuda()
    check_tensorflow_gpu()
python verify_gpu.py

3. Python 环境管理

3.1 用 pyenv 管理 Python 版本

# 安装 pyenv
curl https://pyenv.run | bash

# 追加到 ~/.bashrc 或 ~/.zshrc
echo 'export PYENV_ROOT="$HOME/.pyenv"' >> ~/.bashrc
echo 'command -v pyenv >/dev/null || export PATH="$PYENV_ROOT/bin:$PATH"' >> ~/.bashrc
echo 'eval "$(pyenv init -)"' >> ~/.bashrc
source ~/.bashrc

# 查看可用的 Python 版本列表
pyenv install --list | grep -E "^\s+3\.(10|11|12)"

# 安装 Python
pyenv install 3.11.7
pyenv install 3.10.13

# 设置全局版本
pyenv global 3.11.7

# 设置项目专用版本(在该目录下执行)
cd my-project
pyenv local 3.10.13

# 确认当前 Python 版本
python --version
pyenv versions

3.2 conda 环境(GPU 库)

conda 在安装 RAPIDS、cuML 等 GPU 加速库时特别有用。

# 安装 Miniconda
wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
bash Miniconda3-latest-Linux-x86_64.sh -b
eval "$($HOME/miniconda3/bin/conda shell.bash hook)"
conda init

# 配置渠道
conda config --add channels conda-forge
conda config --add channels nvidia
conda config --set channel_priority strict

# 创建 AI/ML 环境
conda create -n aiml python=3.11 -y
conda activate aiml

# 安装带 CUDA 的 PyTorch
conda install pytorch torchvision torchaudio pytorch-cuda=12.1 -c pytorch -c nvidia

# RAPIDS(GPU 加速数据科学)
conda install -c rapidsai -c conda-forge -c nvidia rapids=23.10 cuda-version=12.0

# 导出与导入环境
conda env export > environment.yml
conda env create -f environment.yml

# 环境列表
conda env list

environment.yml 示例:

name: aiml
channels:
  - pytorch
  - nvidia
  - conda-forge
  - defaults
dependencies:
  - python=3.11
  - pytorch>=2.1
  - torchvision
  - cudatoolkit=12.1
  - numpy>=1.24
  - pandas>=2.0
  - scikit-learn>=1.3
  - pip:
      - transformers>=4.35
      - wandb>=0.16
      - pydantic>=2.0

3.3 用 poetry 管理依赖

# 安装 poetry
curl -sSL https://install.python-poetry.org | python3 -

# 创建项目
poetry new ml-research
cd ml-research

# 添加依赖
poetry add torch torchvision numpy pandas transformers
poetry add --group dev pytest black ruff mypy jupyter

# 指定版本或包含 extra
poetry add "torch[cuda]>=2.1"
poetry add "transformers[torch]>=4.35"

# 安装
poetry install

# 查看虚拟环境信息
poetry env info
poetry env list

3.4 uv(超高速包管理器)

# 安装 uv
curl -LsSf https://astral.sh/uv/install.sh | sh

# 比 pip 快 10-100 倍的包安装
uv pip install torch torchvision numpy pandas

# 创建虚拟环境
uv venv .venv --python 3.11
source .venv/bin/activate

# 从 requirements.txt 安装(速度极快)
uv pip install -r requirements.txt

# 管理 Python 版本
uv python install 3.11 3.12
uv python list

4. JupyterLab 高级配置

4.1 安装 JupyterLab 及扩展

# 安装 JupyterLab
pip install jupyterlab

# 安装常用扩展
pip install jupyterlab-git           # Git 集成
pip install jupyterlab-lsp           # 语言服务器协议(自动补全)
pip install python-lsp-server        # Python LSP
pip install jupyterlab-code-formatter  # 代码格式化工具
pip install black isort              # 格式化工具
pip install jupyterlab-vim           # Vim 键位绑定
pip install ipywidgets               # 交互式控件

# 查看扩展列表
jupyter labextension list

# 启动 JupyterLab
jupyter lab --ip=0.0.0.0 --port=8888 --no-browser

4.2 内核管理

# 查看内核列表
jupyter kernelspec list

# 将当前 conda/venv 环境注册为内核
pip install ipykernel
python -m ipykernel install --user --name myenv --display-name "Python (myenv)"

# 将指定 conda 环境注册为内核
conda activate aiml
conda install ipykernel
python -m ipykernel install --user --name aiml --display-name "Python (aiml GPU)"

# 删除内核
jupyter kernelspec remove old-kernel

# 修改自定义内核规格
# ~/.local/share/jupyter/kernels/aiml/kernel.json

kernel.json 示例:

{
  "argv": [
    "/home/user/miniconda3/envs/aiml/bin/python",
    "-m",
    "ipykernel_launcher",
    "-f",
    "{connection_file}"
  ],
  "display_name": "Python (aiml GPU)",
  "language": "python",
  "env": {
    "CUDA_VISIBLE_DEVICES": "0",
    "PYTHONPATH": "/home/user/projects"
  }
}

4.3 远程 Jupyter(SSH 隧道)

# 在服务器上启动 Jupyter(不使用 token)
jupyter lab --no-browser --port=8888 --ip=127.0.0.1

# 在本地创建 SSH 隧道
ssh -N -L 8888:localhost:8888 user@your-server.com

# 浏览器中访问
# http://localhost:8888

# 设置密码(更安全)
jupyter lab password
# 之后可在 localhost:8888 用密码登录

用于自动启动的 systemd 服务配置:

# /etc/systemd/system/jupyter.service
cat > /tmp/jupyter.service << 'EOF'
[Unit]
Description=Jupyter Lab
After=network.target

[Service]
Type=simple
User=your-username
WorkingDirectory=/home/your-username
ExecStart=/home/your-username/miniconda3/envs/aiml/bin/jupyter lab --no-browser --port=8888
Restart=on-failure

[Install]
WantedBy=multi-user.target
EOF

sudo mv /tmp/jupyter.service /etc/systemd/system/
sudo systemctl daemon-reload
sudo systemctl enable jupyter
sudo systemctl start jupyter

4.4 Jupyter Magic 命令

# 计时
%time result = expensive_function()
%timeit -n 100 result = fast_function()  # 重复 100 次

# 逐行性能分析
%load_ext line_profiler
%lprun -f my_function my_function(data)

# 内存性能分析
%load_ext memory_profiler
%memit result = memory_heavy_function()

# 将单元格内容保存为文件
%%writefile my_script.py
import numpy as np
# ...

# 将文件内容加载到单元格
%load my_script.py

# 执行 shell 命令
!nvidia-smi
!pip list | grep torch
files = !ls -la *.py

# 环境变量
%env CUDA_VISIBLE_DEVICES=0

# 自动重新加载(代码修改后立即生效)
%load_ext autoreload
%autoreload 2

# matplotlib 内嵌显示
%matplotlib inline

# 当前变量列表
%who
%whos

# 查看之前的输出结果
print(_)   # 最后一次结果
print(__)  # 倒数第二次结果

4.5 nbconvert(Notebook 转换)

# 将 notebook 转换为脚本
jupyter nbconvert --to script notebook.ipynb

# 转换为 HTML(便于分享)
jupyter nbconvert --to html notebook.ipynb

# 转换为 PDF(需要 LaTeX)
jupyter nbconvert --to pdf notebook.ipynb

# 转换的同时执行 notebook,输出 HTML
jupyter nbconvert --to html --execute notebook.ipynb

# 在命令行中执行 notebook
jupyter nbconvert --to notebook --execute --inplace notebook.ipynb

# 参数化执行(papermill)
pip install papermill
papermill input.ipynb output.ipynb -p lr 0.001 -p epochs 100

5. AI 场景下的 VS Code

5.1 安装必备扩展

# 在 VS Code 命令行中安装扩展
code --install-extension ms-python.python           # Python
code --install-extension ms-toolsai.jupyter         # Jupyter
code --install-extension ms-python.pylance          # Pylance(LSP)
code --install-extension charliermarsh.ruff         # Ruff 检查工具
code --install-extension github.copilot             # GitHub Copilot
code --install-extension github.copilot-chat        # Copilot Chat
code --install-extension ms-vscode-remote.remote-ssh  # Remote SSH
code --install-extension ms-vscode-remote.remote-containers  # Dev Containers
code --install-extension eamodio.gitlens            # GitLens
code --install-extension njpwerner.autodocstring     # 自动生成 docstring
code --install-extension tabnine.tabnine-vscode     # Tabnine AI

5.2 VS Code 设置(settings.json)

{
  "python.defaultInterpreterPath": "${workspaceFolder}/.venv/bin/python",
  "python.formatting.provider": "none",
  "[python]": {
    "editor.defaultFormatter": "charliermarsh.ruff",
    "editor.formatOnSave": true,
    "editor.codeActionsOnSave": {
      "source.fixAll.ruff": true,
      "source.organizeImports.ruff": true
    }
  },
  "pylance.typeCheckingMode": "basic",
  "python.analysis.autoImportCompletions": true,
  "python.analysis.indexing": true,
  "python.analysis.packageIndexDepths": [
    { "name": "torch", "depth": 5 },
    { "name": "transformers", "depth": 5 }
  ],
  "jupyter.askForKernelRestart": false,
  "jupyter.interactiveWindow.creationMode": "perFile",
  "editor.inlineSuggest.enabled": true,
  "github.copilot.enable": {
    "*": true,
    "yaml": true,
    "plaintext": false
  },
  "files.exclude": {
    "**/__pycache__": true,
    "**/*.pyc": true,
    ".mypy_cache": true,
    ".ruff_cache": true
  },
  "terminal.integrated.env.linux": {
    "CUDA_VISIBLE_DEVICES": "0"
  }
}

5.3 Remote SSH 配置

# 本地 ~/.ssh/config
Host ml-server
    HostName 192.168.1.100
    User your-username
    IdentityFile ~/.ssh/id_rsa
    ForwardAgent yes
    ServerAliveInterval 60
    ServerAliveCountMax 3

Host gpu-cloud
    HostName gpu-server.example.com
    User ubuntu
    IdentityFile ~/.ssh/cloud-key.pem
    Port 22

在 VS Code 中使用 Remote SSH:

  1. Ctrl+Shift+P(Mac 上为 Cmd+Shift+P)
  2. 选择 "Remote-SSH: Connect to Host"
  3. 选择已配置的主机名

远程服务器上自动选择 Python 环境:

{
  "remote.SSH.defaultExtensions": ["ms-python.python", "ms-toolsai.jupyter", "charliermarsh.ruff"]
}

5.4 Dev Containers

.devcontainer/devcontainer.json 示例:

{
  "name": "AI Development",
  "build": {
    "dockerfile": "Dockerfile",
    "args": {
      "CUDA_VERSION": "12.1.1",
      "PYTHON_VERSION": "3.11"
    }
  },
  "runArgs": ["--gpus", "all", "--shm-size", "8g"],
  "mounts": [
    "source=/data,target=/data,type=bind",
    "source=${localEnv:HOME}/.cache,target=/root/.cache,type=bind"
  ],
  "customizations": {
    "vscode": {
      "extensions": [
        "ms-python.python",
        "ms-toolsai.jupyter",
        "charliermarsh.ruff",
        "github.copilot"
      ],
      "settings": {
        "python.defaultInterpreterPath": "/usr/local/bin/python"
      }
    }
  },
  "postCreateCommand": "pip install -e '.[dev]'",
  "remoteUser": "root"
}

6. GPU Docker 容器

6.1 安装 NVIDIA Container Toolkit

# 安装 NVIDIA Container Toolkit
distribution=$(. /etc/os-release;echo $ID$VERSION_ID)
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/$distribution/libnvidia-container.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
sudo nvidia-ctk runtime configure --runtime=docker
sudo systemctl restart docker

# 测试 GPU 访问
docker run --rm --gpus all nvidia/cuda:12.2.0-base-ubuntu22.04 nvidia-smi

6.2 AI 项目用 Dockerfile

# Dockerfile
FROM nvidia/cuda:12.1.1-cudnn8-devel-ubuntu22.04

# 环境变量
ENV DEBIAN_FRONTEND=noninteractive
ENV PYTHONUNBUFFERED=1
ENV PYTHONDONTWRITEBYTECODE=1
ENV PIP_NO_CACHE_DIR=1

# 系统软件包
RUN apt-get update && apt-get install -y \
    python3.11 \
    python3.11-dev \
    python3.11-venv \
    python3-pip \
    git \
    wget \
    curl \
    vim \
    htop \
    tmux \
    && rm -rf /var/lib/apt/lists/*

# 设置默认 Python
RUN update-alternatives --install /usr/bin/python python /usr/bin/python3.11 1
RUN update-alternatives --install /usr/bin/pip pip /usr/bin/pip3 1

# 升级 pip
RUN pip install --upgrade pip setuptools wheel

# 工作目录
WORKDIR /workspace

# Python 依赖(优化分层缓存)
COPY requirements.txt .
RUN pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
RUN pip install -r requirements.txt

# 复制源码
COPY . .

# 安装包
RUN pip install -e ".[dev]"

# 创建非 root 用户(安全考虑)
RUN useradd -m -u 1000 researcher
RUN chown -R researcher:researcher /workspace
USER researcher

# 暴露端口
EXPOSE 8888 6006

CMD ["jupyter", "lab", "--ip=0.0.0.0", "--port=8888", "--no-browser"]

6.3 用 Docker Compose 搭建服务栈

# docker-compose.yml
version: '3.8'

services:
  # 训练容器
  trainer:
    build:
      context: .
      dockerfile: Dockerfile
    runtime: nvidia
    environment:
      - NVIDIA_VISIBLE_DEVICES=0
      - CUDA_VISIBLE_DEVICES=0
      - WANDB_API_KEY=${WANDB_API_KEY}
    volumes:
      - ./src:/workspace/src
      - ./configs:/workspace/configs
      - /data:/data:ro
      - model-checkpoints:/workspace/checkpoints
      - ~/.cache/huggingface:/root/.cache/huggingface
    shm_size: '8g'
    command: python train.py

  # Jupyter notebook 服务器
  notebook:
    build:
      context: .
      dockerfile: Dockerfile
    runtime: nvidia
    environment:
      - NVIDIA_VISIBLE_DEVICES=1
    ports:
      - '8888:8888'
    volumes:
      - ./notebooks:/workspace/notebooks
      - ./src:/workspace/src
      - /data:/data:ro
    command: jupyter lab --ip=0.0.0.0 --no-browser --allow-root

  # TensorBoard
  tensorboard:
    image: tensorflow/tensorflow:latest
    ports:
      - '6006:6006'
    volumes:
      - model-checkpoints:/workspace/checkpoints:ro
    command: tensorboard --logdir=/workspace/checkpoints --host=0.0.0.0

  # MLflow 追踪服务器
  mlflow:
    image: ghcr.io/mlflow/mlflow:v2.8.0
    ports:
      - '5000:5000'
    volumes:
      - mlflow-data:/mlflow
    command: mlflow server --host=0.0.0.0 --port=5000 --default-artifact-root=/mlflow/artifacts

volumes:
  model-checkpoints:
  mlflow-data:
# 启动完整服务栈
docker-compose up -d

# 查看日志
docker-compose logs -f trainer

# 重启指定服务
docker-compose restart notebook

# 查看 GPU 使用情况(容器内)
docker-compose exec trainer nvidia-smi

# 清理
docker-compose down --volumes

6.4 GPU 共享策略

# 分配指定 GPU
docker run --gpus '"device=0,1"' myimage  # 使用 GPU 0、1
docker run --gpus '"device=2"' myimage     # 仅使用 GPU 2

# MIG(Multi-Instance GPU)配置(A100/H100)
sudo nvidia-smi mig -i 0 --create-gpu-instance 3g.20gb
sudo nvidia-smi mig -i 0 --create-compute-instance 3g.20gb

# 为容器分配 MIG 实例
docker run --gpus '"MIG-GPU-xxxxxxxx-xx-xx-xxxx-xxxxxxxxxxxx/x/x"' myimage

# GPU 内存分数分配(time-slicing)
# 在 /etc/nvidia-container-runtime/config.toml 中设置

7. 远程开发环境

7.1 基于 SSH 的远程开发

# 生成 SSH 密钥(本地)
ssh-keygen -t ed25519 -C "your-email@example.com"

# 将公钥复制到服务器
ssh-copy-id -i ~/.ssh/id_ed25519.pub user@server.com

# 或手动复制
cat ~/.ssh/id_ed25519.pub | ssh user@server.com "mkdir -p ~/.ssh && cat >> ~/.ssh/authorized_keys"

# SSH 连接优化(~/.ssh/config)
Host ml-server
    HostName server.example.com
    User researcher
    IdentityFile ~/.ssh/id_ed25519
    ControlMaster auto
    ControlPath ~/.ssh/cm_%r@%h:%p
    ControlPersist 10m
    ServerAliveInterval 30
    Compression yes

7.2 tmux(保持会话)

# 安装 tmux
sudo apt install tmux

# 启动新会话
tmux new-session -s training

# 重新连接会话
tmux attach-session -t training

# 会话列表
tmux list-sessions

# 常用快捷键(Prefix:Ctrl+B)
# Ctrl+B d    :分离会话(SSH 断开后依然运行)
# Ctrl+B c    :新建窗口
# Ctrl+B n/p  :下一个/上一个窗口
# Ctrl+B %    :垂直分屏
# Ctrl+B "    :水平分屏
# Ctrl+B z    :最大化/还原面板

~/.tmux.conf 配置:

# ~/.tmux.conf

# 启用鼠标支持
set -g mouse on

# 增加历史缓冲区
set -g history-limit 50000

# 窗口编号从 1 开始
set -g base-index 1
setw -g pane-base-index 1

# 颜色支持
set -g default-terminal "screen-256color"
set -ga terminal-overrides ",xterm-256color:Tc"

# 自定义状态栏
set -g status-bg colour235
set -g status-fg colour136
set -g status-right '#[fg=colour166]%d %b #[fg=colour136]%H:%M '

# 修改 Prefix 键(Ctrl+A,screen 风格)
set -g prefix C-a
unbind C-b
bind C-a send-prefix

# 快速切换面板
bind -n M-Left select-pane -L
bind -n M-Right select-pane -R
bind -n M-Up select-pane -U
bind -n M-Down select-pane -D

7.3 文件同步(rsync、sshfs)

# 用 rsync 同步代码
# 本地 -> 服务器
rsync -avz --exclude '__pycache__' --exclude '.git' \
    ./my-project/ user@server:/home/user/my-project/

# 服务器 -> 本地(下载结果)
rsync -avz user@server:/home/user/my-project/outputs/ ./outputs/

# 自动同步脚本
cat > sync.sh << 'EOF'
#!/bin/bash
rsync -avz --exclude '__pycache__' \
    --exclude '.git' \
    --exclude '.venv' \
    --exclude '*.pyc' \
    --delete \
    ./ user@ml-server:/home/user/project/
echo "同步完成:$(date)"
EOF
chmod +x sync.sh

# 用 sshfs 挂载远程文件系统
sudo apt install sshfs
mkdir -p ~/remote-server
sshfs user@server:/home/user ~/remote-server -o reconnect,ServerAliveInterval=15

# 卸载挂载
fusermount -u ~/remote-server

8. 监控工具

8.1 nvidia-smi watch

# 每 1 秒刷新一次 GPU 状态
watch -n 1 nvidia-smi

# 仅查看 GPU 内存使用量
nvidia-smi --query-gpu=index,name,memory.used,memory.total,utilization.gpu \
    --format=csv -l 1

# 记录 CSV 日志
nvidia-smi --query-gpu=timestamp,index,utilization.gpu,memory.used,temperature.gpu \
    --format=csv -l 1 > gpu_log.csv

8.2 nvtop

# 安装 nvtop(htop 的 GPU 版本)
sudo apt install nvtop

# 运行
nvtop

8.3 用 Python 监控 GPU

# gpu_monitor.py
import subprocess
import time
import json
from dataclasses import dataclass
from typing import List

@dataclass
class GPUStats:
    index: int
    name: str
    temperature: float
    utilization: float
    memory_used: int
    memory_total: int
    power_draw: float

def get_gpu_stats() -> List[GPUStats]:
    """通过 nvidia-smi 查询 GPU 统计信息"""
    cmd = [
        'nvidia-smi',
        '--query-gpu=index,name,temperature.gpu,utilization.gpu,memory.used,memory.total,power.draw',
        '--format=csv,noheader,nounits'
    ]
    result = subprocess.run(cmd, capture_output=True, text=True)

    gpus = []
    for line in result.stdout.strip().split('\n'):
        parts = [p.strip() for p in line.split(',')]
        gpus.append(GPUStats(
            index=int(parts[0]),
            name=parts[1],
            temperature=float(parts[2]),
            utilization=float(parts[3]),
            memory_used=int(parts[4]),
            memory_total=int(parts[5]),
            power_draw=float(parts[6]) if parts[6] != 'N/A' else 0.0
        ))
    return gpus

def monitor_training(interval: int = 10, duration: int = 3600):
    """训练过程中监控 GPU"""
    print(f"开始 GPU 监控(间隔:{interval}秒)")
    history = []

    start_time = time.time()
    while time.time() - start_time < duration:
        stats = get_gpu_stats()
        timestamp = time.time() - start_time

        for gpu in stats:
            mem_pct = gpu.memory_used / gpu.memory_total * 100
            print(
                f"[{timestamp:.0f}s] GPU {gpu.index}: "
                f"利用率={gpu.utilization:.0f}% "
                f"内存={gpu.memory_used}/{gpu.memory_total}MB ({mem_pct:.0f}%) "
                f"温度={gpu.temperature:.0f}C "
                f"功率={gpu.power_draw:.0f}W"
            )
            history.append({
                'timestamp': timestamp,
                'gpu_index': gpu.index,
                'utilization': gpu.utilization,
                'memory_used': gpu.memory_used,
                'temperature': gpu.temperature
            })

        time.sleep(interval)

    return history

if __name__ == "__main__":
    history = monitor_training(interval=5, duration=60)
    print(f"\n共收集 {len(history)} 个测量值")

8.4 系统监控

# 安装并使用 htop
sudo apt install htop
htop

# 用 iotop 监控 I/O
sudo apt install iotop
sudo iotop -o  # 仅显示有 I/O 活动的进程

# 查看磁盘 I/O
iostat -x 1 5

# 内存使用量
free -h
vmstat 1 10

# 网络监控
sudo apt install nethogs
sudo nethogs

# 综合监控面板(glances)
pip install glances
glances

9. 代码质量工具

9.1 black + ruff 配置

# 安装
pip install black ruff pre-commit

# 运行 black
black .
black --line-length 88 src/

# 运行 ruff
ruff check .
ruff check --fix .  # 自动修复
ruff format .       # 格式化(可替代 black)

pyproject.toml 配置:

[tool.black]
line-length = 88
target-version = ['py310', 'py311']
include = '\.pyi?$'

[tool.ruff]
line-length = 88
target-version = "py310"

[tool.ruff.lint]
select = [
    "E",   # pycodestyle errors
    "W",   # pycodestyle warnings
    "F",   # pyflakes
    "I",   # isort
    "N",   # pep8-naming
    "UP",  # pyupgrade
    "B",   # flake8-bugbear
]
ignore = ["E501", "B008"]

[tool.ruff.lint.isort]
known-first-party = ["my_package"]

9.2 pre-commit hooks

# .pre-commit-config.yaml
repos:
  - repo: https://github.com/pre-commit/pre-commit-hooks
    rev: v4.5.0
    hooks:
      - id: trailing-whitespace
      - id: end-of-file-fixer
      - id: check-yaml
      - id: check-json
      - id: check-merge-conflict
      - id: detect-private-key
      - id: check-added-large-files
        args: ['--maxkb=10000']

  - repo: https://github.com/psf/black
    rev: 23.11.0
    hooks:
      - id: black

  - repo: https://github.com/astral-sh/ruff-pre-commit
    rev: v0.1.6
    hooks:
      - id: ruff
        args: [--fix]

  - repo: https://github.com/pre-commit/mirrors-mypy
    rev: v1.7.1
    hooks:
      - id: mypy
        additional_dependencies:
          - types-requests
          - pydantic
# 安装并启用 pre-commit
pre-commit install

# 对全部文件执行
pre-commit run --all-files

# 执行指定 hook
pre-commit run black --all-files

9.3 带覆盖率的 pytest

# 安装
pip install pytest pytest-cov pytest-xdist

# 基本执行
pytest tests/ -v

# 包含覆盖率
pytest tests/ --cov=src --cov-report=html --cov-report=term-missing

# 并行执行
pytest tests/ -n auto  # 按 CPU 数量并行

# 显示耗时最长的测试
pytest tests/ --durations=10

# 仅执行指定标记
pytest tests/ -m "not slow"

pytest.inipyproject.toml

[tool.pytest.ini_options]
testpaths = ["tests"]
python_files = ["test_*.py", "*_test.py"]
python_classes = ["Test*"]
python_functions = ["test_*"]
markers = [
    "slow: marks tests as slow",
    "gpu: marks tests requiring GPU",
    "integration: integration tests"
]
addopts = [
    "-v",
    "--tb=short",
    "--cov=src",
    "--cov-report=html",
    "--cov-fail-under=80"
]

10. Weights & Biases 配置

10.1 安装与初始化 W&B

# 安装
pip install wandb

# 登录(设置 API key)
wandb login
# 或通过环境变量设置
export WANDB_API_KEY=your-api-key-here

10.2 W&B 基本用法

import wandb
import numpy as np
import torch

# 初始化实验
run = wandb.init(
    project="bert-sentiment-analysis",
    name="run-001-baseline",
    config={
        "learning_rate": 2e-5,
        "batch_size": 32,
        "epochs": 10,
        "model": "bert-base-uncased",
        "optimizer": "adamw"
    },
    tags=["baseline", "bert", "nlp"],
    notes="基础 BERT 微调实验"
)

# 访问配置
config = wandb.config
print(f"学习率:{config.learning_rate}")

# 记录指标
for epoch in range(config.epochs):
    train_loss = 2.0 - epoch * 0.15 + np.random.randn() * 0.1
    val_loss = 2.2 - epoch * 0.12 + np.random.randn() * 0.1
    val_acc = 0.5 + epoch * 0.04 + np.random.randn() * 0.01

    wandb.log({
        "epoch": epoch,
        "train/loss": train_loss,
        "val/loss": val_loss,
        "val/accuracy": val_acc,
        "learning_rate": config.learning_rate * (0.95 ** epoch)
    })

    # 梯度直方图(PyTorch)
    # wandb.log({"gradients": wandb.Histogram(model.fc.weight.grad)})

# 记录可视化图表
import matplotlib.pyplot as plt

fig, ax = plt.subplots()
x = np.linspace(0, 10, 100)
ax.plot(x, np.sin(x), label='预测')
ax.plot(x, np.cos(x), label='实际')
ax.legend()
wandb.log({"prediction_plot": wandb.Image(fig)})
plt.close()

# 结束实验
wandb.finish()

10.3 Artifact 管理

import wandb
import os

# 保存 artifact(模型检查点)
def save_model_artifact(model_path: str, run, metadata: dict = None):
    artifact = wandb.Artifact(
        name="model-checkpoint",
        type="model",
        metadata=metadata or {}
    )
    artifact.add_file(model_path)
    run.log_artifact(artifact)
    print(f"模型 artifact 已保存:{model_path}")

# 保存 artifact(数据集)
def save_dataset_artifact(data_dir: str, run):
    artifact = wandb.Artifact(
        name="training-dataset",
        type="dataset",
        description="预处理后的训练数据集"
    )
    artifact.add_dir(data_dir)
    run.log_artifact(artifact)

# 加载 artifact
def load_model_artifact(artifact_name: str, version: str = "latest"):
    run = wandb.init(project="my-project", job_type="inference")
    artifact = run.use_artifact(f"{artifact_name}:{version}")
    artifact_dir = artifact.download()
    print(f"artifact 下载位置:{artifact_dir}")
    return artifact_dir

# 使用示例
run = wandb.init(project="bert-training")

# 模拟保存临时模型文件
with open("/tmp/model.pt", "w") as f:
    f.write("model_weights")

save_model_artifact(
    "/tmp/model.pt",
    run,
    metadata={"accuracy": 0.94, "epoch": 10, "val_loss": 0.28}
)

wandb.finish()

10.4 用于超参数优化的 Sweep

import wandb
import numpy as np

# Sweep 配置
sweep_config = {
    "method": "bayes",  # random, grid, bayes
    "metric": {
        "name": "val/accuracy",
        "goal": "maximize"
    },
    "parameters": {
        "learning_rate": {
            "distribution": "log_uniform_values",
            "min": 1e-5,
            "max": 1e-3
        },
        "batch_size": {
            "values": [16, 32, 64, 128]
        },
        "hidden_size": {
            "values": [256, 512, 768, 1024]
        },
        "dropout": {
            "distribution": "uniform",
            "min": 0.1,
            "max": 0.5
        },
        "num_layers": {
            "values": [2, 4, 6, 8]
        }
    },
    "early_terminate": {
        "type": "hyperband",
        "min_iter": 3
    }
}

def train_sweep():
    """Sweep 调用的训练函数"""
    run = wandb.init()
    config = wandb.config

    # 按配置模拟训练模型
    best_val_acc = 0.0

    for epoch in range(10):
        # 实际场景中会真正训练模型
        train_loss = 1.0 / (config.learning_rate * 1000) * np.random.uniform(0.8, 1.2)
        val_acc = min(0.99, config.hidden_size / 2000 + np.random.uniform(0, 0.1))

        if val_acc > best_val_acc:
            best_val_acc = val_acc

        wandb.log({
            "epoch": epoch,
            "train/loss": train_loss,
            "val/accuracy": val_acc
        })

    wandb.finish()

# 创建并运行 Sweep
sweep_id = wandb.sweep(sweep_config, project="hyperparameter-search")
print(f"Sweep ID:{sweep_id}")

# 运行 agent(尝试 N 次)
wandb.agent(sweep_id, function=train_sweep, count=20)

如需在多台服务器上分布式运行 Sweep,可使用如下方式:

# 服务器 1
wandb agent username/project/sweep-id

# 服务器 2(同时运行)
wandb agent username/project/sweep-id

结语

AI 开发环境搭建,前期投入的时间越多,后期的生产力提升就越明显。本指南涉及的核心要点总结如下:

必备优先级:

  1. CUDA 环境搭建:正确安装 GPU 驱动和 CUDA Toolkit 是一切的基础。
  2. Python 环境隔离:用 pyenv + conda/poetry 组合,按项目分离环境。
  3. JupyterLab + VS Code:探索性分析用 Jupyter,正式开发用 VS Code。
  4. Docker 容器化:保证可复现的实验环境,简化团队协作。
  5. W&B 实验追踪:追踪全部实验,获得可复现性与洞察。

值得长期投入的方向:

  • 用 pre-commit hooks 和 CI/CD 自动化代码质量管控。
  • 熟练掌握远程开发环境(Remote SSH + tmux),随时随地高效工作。
  • 通过 GPU 监控,及早发现训练过程中的问题。

参考资料

현재 단락 (1/810)

AI 开发环境搭建,是项目成功的一半。没有恰当的环境,实验的可复现性、团队协作、快速迭代都无从谈起。本指南涵盖 AI/ML 开发者需要掌握的全部环境配置内容 — 从 GPU 服务器的初始设置,到日常开...

작성 글자: 0원문 글자: 24,665작성 단락: 0/810