Skip to content

windows-cuda

Windows, GPU, fresh install. The lane that proves a CUDA pack installs on the platform most ComfyUI users are actually on.
OS / accelerator Windows / CUDA
Install method manual -- venv + a ComfyUI checkout
Runner [self-hosted, windows, cuda]
Install path fresh
Config key [test.windows_cuda]
Dispatch only yes

What a green cell proves

The full fresh sequence on Windows: venv, torch triple from the CUDA index, ComfyUI clone, your requirements, your install.py, server boot, and your cuda workflow list executed on a real GPU.

This is the combination most of your users run, and the one where a wheel that exists on Linux but not Windows shows up.

How it runs

In a Windows container with process isolation and GPU device mapping, via comfy-test docker run. Windows containers cannot use --device under Hyper-V isolation, so process isolation is what makes GPU passthrough possible here -- which in turn means the container shares the host kernel, so the host's driver version and the image must agree.

Gotchas

  • Driver/image coupling. A host driver upgrade can break the image until it is rebuilt. This is the usual cause of a lane that was green last week.
  • Long paths. Windows' 260-character limit bites deep dependency trees inside a container more often than on a developer machine.
  • Windows Defender. Real-time scanning on the workspace slows installs significantly; comfy-test's docker path manages an exclusion for it.

See also