conda安装PyTorch后执行import torch触发Illegal instruction报错

Hey there, let's figure out why importing PyTorch throws an Illegal instruction error right after a fresh Anaconda install—this is a super common issue with older CPUs or systems that don't support advanced instruction sets. Here's how to diagnose and fix it:

Why This Happens #

The default PyTorch packages you get via conda install pytorch are pre-compiled to take advantage of modern CPU features like AVX2 or FMA (Advanced Vector Extensions). If your CPU doesn't support these instructions, or they're disabled in your BIOS/UEFI, Python will crash with that "Illegal instruction" message as soon as it tries to execute code that uses those unsupported features.

Step-by-Step Fixes #

1. Check Your CPU's Supported Instruction Sets #

First, confirm what your CPU can handle. On Linux, run this command in your terminal:

lscpu | grep Flags

Look for flags like avx2, fma, or avx512f in the output. If you don't see these, your CPU is too old to run the default PyTorch build.

2. Uninstall the Broken PyTorch Installation #

Remove the current PyTorch packages from your conda environment to start fresh:

conda uninstall pytorch torchvision torchaudio cpuonly --force

(Adjust the command if you installed a GPU-specific version—just include the relevant packages you have.)

3. Install a PyTorch Version Compatible With Your CPU #

You have two reliable options here:

Option A: Install a Pre-Compiled "Legacy" Build

For older CPUs, use a PyTorch version compiled without advanced instruction set dependencies. Try this conda command (works for CPU-only setups):

conda install pytorch=1.12.1 torchvision=0.13.1 torchaudio=0.12.1 cpuonly nomkl -c pytorch

The nomkl flag skips the Intel MKL library (which relies on advanced CPU features), and version 1.12.1 is a stable release that works well on older hardware.

Option B: Compile PyTorch From Source (For Latest Versions)

If you need the newest PyTorch features, compile it directly from source to match your CPU's capabilities. Here's the quick breakdown:

• First, install all required dependencies (like cmake, gcc, etc.) via conda or your system package manager.
• Clone the PyTorch repository:

  git clone --recursive https://github.com/pytorch/pytorch
  cd pytorch
  

• Set an environment variable to tell the compiler to optimize for your specific CPU:

  export CXXFLAGS="-march=native"
  

• Run the build commands (this will take a while, depending on your system):

  conda env update --file environment.yml --prune
  python setup.py install
  

4. Verify the Fix #

Open a new terminal window (to ensure environment changes take effect), activate your conda environment, and run:

python -c 'import torch; print("PyTorch imported successfully! Version:", torch.__version__)'

If this runs without crashing, you're good to go.

Quick Note for Laptop/Server Users #

If your CPU should support AVX2 but you're still getting the error, check your BIOS/UEFI settings—sometimes advanced instruction sets are disabled by default. Enable them if possible, then retry the default PyTorch install.

内容的提问来源于stack exchange，提问作者drevicko