Skip to content

windows-portable-cuda

The portable bundle with a GPU, installed fresh. The closest CI gets to what a typical Windows ComfyUI user actually has.
OS / accelerator Windows / CUDA
Install method portable -- the official ComfyUI bundle
Runner [self-hosted, windows, cuda]
Install path fresh
Config key [test.windows_portable_cuda]
Dispatch only yes

What a green cell proves

That your pack installs into an embedded Python with no venv and no compiler, on a real GPU, and that its CUDA workflows execute there. Of the ten lanes, this is the one whose environment most closely matches the median Windows user.

Gotchas

  • The embedded Python is the hard part, not the GPU. Most failures here are the portable constraints from windows-portable-cpu -- no .git, no compiler, python_embeded paths -- surfacing on the fresh path where the install actually runs.
  • Prebuilt wheels matter. With no compiler available, a CUDA dependency without a Windows wheel for the resolved torch simply cannot install. See cuda-wheels.

See also