How to Install PyTorch in CI (CPU Wheels)

The default PyTorch wheel can pull multi-GB CUDA payloads you do not need on CPU CI. The fix is the CPU wheel index.

PyTorch publishes platform/accelerator-specific wheels through its own index. On CPU-only CI, installing from the CPU index avoids gigabytes of CUDA libraries and the timeouts they cause.

Why it fails in CI

The default install pulls CUDA-enabled torch (huge) on a CPU runner → slow/timeout.
A Python version with no matching torch wheel → no candidate.
Pinned torch/torchvision versions that are incompatible with each other.

Install it reliably

Install from the CPU index URL with matching torch/torchvision versions. Pin a supported Python first via setup-python.

Terminal

# CPU-only wheels (no CUDA payload)
pip install torch torchvision \
  --index-url https://download.pytorch.org/whl/cpu

# verify CPU build
python -c "import torch; print(torch.__version__, torch.cuda.is_available())"

Cache & speed

Cache ~/.cache/pip keyed on requirements. Even CPU wheels are large; the download is where faster/larger managed runners help and transient failures auto-retry.

.github/workflows/ci.yml

- uses: actions/cache@v4
  with:
    path: ~/.cache/pip
    key: pip-${{ runner.os }}-${{ hashFiles('**/requirements*.txt') }}

Common errors

Multi-GB download / timeout → use --index-url .../whl/cpu for CPU-only CI.
No matching distribution found for torch → Python too new or wrong index; pin Python.
torchvision ... compiled against a different version of torch → pin a matching pair.

Key takeaways

Install from the PyTorch CPU index to skip multi-GB CUDA wheels.
Pin torch and torchvision to a compatible pair.
Cache pip; even CPU wheels are large.