Skip to content

macos-cpu

Apple Silicon, MPS rather than CUDA, and the most expensive runner in the matrix. The lane that catches "CUDA or nothing" assumptions.
OS / accelerator macOS / CPU
Install method manual -- venv + a ComfyUI checkout
Runner macos-latest (GitHub-hosted, 10x billing)
Install path attach
Config key [test.macos]
Also accepts macos, macos_cpu

Why this lane exists

It is the only lane where cuda_capable is false. A pack that reaches for torch.device("cuda") unconditionally, or calls torch.cuda.empty_cache() as if it were universal, fails here and passes everywhere else.

The ComfyUI-native answers are comfy.model_management.get_torch_device() and soft_empty_cache(), which dispatch across MPS, XPU, NPU, MLU and CUDA. syntax fails the first pattern statically and warnings reports the second.

Cost, and execution_light

At 10x billing this lane dominates a full matrix's runner budget. It is also memory-constrained: the full per-frame capture loop can peg the browser process and kill the Playwright IPC pipe on a 7 GB runner, which is exactly why execution_light exists (ADR-0011).

A pack that hits this lists execution_light instead of execution; for per-lane variation use skip_workflow:

[test.macos]
skip_workflow = true   # run the pipeline, not the workflows

Gotchas

  • Attach lane -- no installability claim.
  • macOS resolves pack requirements differently. That lane overrides the index-routed install path with plain uv, so extra_pip_indices may not reach your pack's own requirements there. Verify before depending on it.

See also