cuda-torch-install-on-cpu-runner

Flags standard Ubuntu x64 GitHub-hosted jobs that install PyTorch directly without a CPU-only index, pulling the CUDA build and its NVIDIA dependencies onto a runner that has no GPU.

Why it matters

The default PyPI PyTorch build for Linux x64 bundles CUDA. For torch 2.5.1 on CPython 3.11 Linux x64, the wheel is about 900 MB and pulls roughly a dozen nvidia-* packages, for a total of roughly 2.8 GB before extraction overhead (triton and other optional paths add more). A standard Ubuntu x64 runner has no GPU, so none of that CUDA code can run, yet the download, disk usage, and install time are paid on every job run and inflate any pip cache or container image built from it.

The CPU-only build is a small fraction of that size and installs no NVIDIA dependencies.

What it flags

This rule only fires when all of the following are true:

• the job runs on a standard Ubuntu x64 label (ubuntu-latest, ubuntu-20.04, ubuntu-22.04, ubuntu-24.04, and later)
• when runs-on references a matrix axis, that axis has a static list and every resolved value is a standard Ubuntu x64 label
• the job is not configured with a job container
• a step directly installs torch, torchvision, or torchaudio with pip install, uv pip install, uv add, or poetry add

It does not fire when:

• the job already selects a CPU build (--index-url .../whl/cpu, --find-links .../cpu, a +cpu version, --torch-backend=cpu|auto, or UV_TORCH_BACKEND=cpu|auto)
• CUDA build intent is visible in the workflow, such as TORCH_CUDA_ARCH_LIST, CUDA_HOME, CUDA_PATH, CUDA_COMPUTE_CAP, LIBTORCH_USE_PYTORCH, nvcc, a cuda-toolkit action, or a whl/cu index
• the runner is not a standard Ubuntu x64 label, for example a GPU-labeled or self-hosted runner, Windows, macOS, or arm64
• torch is only installed indirectly, for example through pip install -r requirements.txt or pip install .

torch.cuda.is_available() and step names mentioning CUDA are intentionally not treated as CUDA build intent, because CPU jobs often gate GPU-specific tests that way.

A --extra-index-url .../whl/cpu is intentionally not treated as CPU intent. pip does not rank indexes, so it does not guarantee the CPU build is selected.

CPU index/backend environment settings are read from the effective workflow, job, and step environment, including step-level UV_TORCH_BACKEND and PIP_INDEX_URL. A setting on another step does not apply to this install.

Suggested action

Install the CPU-only PyTorch build on CPU runners, for example with --index-url https://download.pytorch.org/whl/cpu or a matching CPU index or find-links for the pinned version. A +cpu version pin alone is not enough without the CPU index or find-links.

Verification

Compare dependency install time and cache or container image size before and after switching to the CPU-only build on CPU runners.

Notes

TensorFlow is out of scope. The official guidance installs tensorflow for CPU and tensorflow[and-cuda] for GPU, so the PyTorch CUDA-bundle heuristic does not transfer. A separate tensorflow-cpu sizing rule would be a different change.