在 Intel GPU 上开始使用pytorch-xpu(WSL2)

发表于 更新于 本文字数： 2.8k 阅读时长 ≈ 3 分钟
在 Intel GPU 上开始使用pytorch

参考链接：

Getting Started on Intel GPU PyTorch 2.8 documentation

PyTorch Prerequisites for Intel® GPUs

Installing Client GPUs Intel® software for general purpose GPU capabilities documentation

相关配置

Intel GPUs Driver Installation

Installing Data Center GPU

Installing Data Center GPU: Rolling Releases —— Intel® software for general purpose GPU capabilities documentation

执行

1
2
3
4
5
6
sudo apt install -y \
    linux-headers-$(uname -r) \
    linux-modules-extra-$(uname -r) \
    flex bison \
    intel-fw-gpu intel-i915-dkms xpu-smi
sudo reboot

可能遇到的问题：

  1. wsl: Failed to translate ‘C:\Windows\system32’
    wsl: Failed to translate ‘C:\Windows’
    wsl: Failed to translate ‘C:\Windows\System32\Wbem’
    wsl: Failed to translate ‘C:\Windows\System32\WindowsPowerShell\v1.0'
    wsl: Failed to translate ‘C:\Windows\System32\OpenSSH'
    wsl: Failed to translate ‘C:\Program Files\dotnet'
    wsl: Failed to translate ‘D:\Program Files\Git\cmd’
    wsl: Failed to translate ‘D:\Program Files\Git\bin’
    wsl: Failed to translate ‘D:\Program Files\nodejs'
    wsl: Failed to translate ‘C:\Users\29655\AppData\Local\Microsoft\WindowsApps’
    wsl: Failed to translate ‘D:\Program Files\Microsoft VS Code\bin’
    wsl: Failed to translate ‘C:\Users\29655\AppData\Roaming\npm’
  • 解决方案：禁止wsl2访问windows的环境变量。
  1. E: Unable to locate package linux-headers-6.6.87.2-microsoft-standard-WSL2
    E: Couldn’t find any package by glob ‘linux-headers-6.6.87.2-microsoft-standard-WSL2’
    E: Unable to locate package linux-modules-extra-6.6.87.2-microsoft-standard-WSL2
    E: Couldn’t find any package by glob ‘linux-modules-extra-6.6.87.2-microsoft-standard-WSL2’

  • 解决方案：问Kimi

    1. 在 WSL2 Ubuntu 中运行：uname -r，输出例如6.6.87.2-microsoft-standard-WSL2。

    2. 安装依赖

    1
    2
    sudo apt update
    sudo apt install build-essential bc flex bison libssl-dev libelf-dev
    1. 下载并解压 WSL2 内核源码（含头文件）
    1
    2
    3
    cd /usr/src
    sudo git clone --depth 1 --branch linux-msft-wsl-6.6.87.2 \
      https://github.com/microsoft/WSL2-Linux-Kernel.git
    1. 创建头文件链接（供编译模块使用）
    1
    sudo ln -s /usr/src/WSL2-Linux-Kernel /lib/modules/$(uname -r)/build
    1. 验证是否成功
    1
    ls -l /lib/modules/$(uname -r)/build

    如果输出的是源码目录的链接，说明已配置好。

执行xpu-smi时可能遇到的问题：

1
2
3
4
xpu-smi: /home/daniel/miniconda3/envs/learning/lib/libstdc++.so.6: version `GLIBCXX_3.4.32' not found (required by xpu-smi)
xpu-smi: /home/daniel/miniconda3/envs/learning/lib/libstdc++.so.6: version `GLIBCXX_3.4.32' not found (required by /lib/x86_64-linux-gnu/libxpum.so.1)
xpu-smi: /home/daniel/miniconda3/envs/learning/lib/libstdc++.so.6: version `GLIBCXX_3.4.30' not found (required by /lib/x86_64-linux-gnu/libxpum.so.1)
xpu-smi: /home/daniel/miniconda3/envs/learning/lib/libstdc++.so.6: version `GLIBCXX_3.4.32' not found (required by /lib/x86_64-linux-gnu/libze_loader.so.1)

解决方案：

升级 Conda 环境中的 libstdc++ 。

1
conda install -c conda-forge libstdcxx-ng=13

Driver Installation for Client GPUs from Intel

Installing Client GPUs ——Intel® software for general purpose GPU capabilities documentation

Install pytorch-xpu binaries

Getting Started on Intel GPU PyTorch 2.8 documentation

1
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/xpu)