StabilityMatrix icon indicating copy to clipboard operation
StabilityMatrix copied to clipboard

ComfyUI packages always default to --cpu option, even when installed with Cuda option on linux

Open drmbt opened this issue 1 year ago • 4 comments

What happened?

installing a new ComfyUI package when running StabilityMatrix on WSL2, even when specifying cuda in advanced install options (which should be the default anyhow), I always need to uncheck the --cpu option from the settings for that package. This isn't an issue on windows

Steps to reproduce

install a new package, with Cuda option selected from Advanced dropdown. launch the package and realize GPU is unrecognized wave fist in the air go uncheck --cpu option from package Launch Options

Relevant logs

No response

Version

Latest Stable v2.12.3

What Operating System are you using?

Linux

drmbt avatar Nov 14 '24 00:11 drmbt

We're using lspci to get GPU info on Linux devices, and it seems wsl doesn't properly report GPUs in that. We'll be improving GPU detection in the next update as well as adding an override option to choose your "default" GPU.

mohnjiles avatar Nov 14 '24 03:11 mohnjiles

We're using lspci to get GPU info on Linux devices, and it seems wsl doesn't properly report GPUs in that. We'll be improving GPU detection in the next update as well as adding an override option to choose your "default" GPU.

any updates on this? Right now every new WSL package i install, I need to manually activate venv, uninstall torch torchvision torchaudio, and then re install torch, which is becoming quite the PITA.

Also, on Linux, comfy is pretty stable these days on modern torch 2.5.1, do we need to keep enforcing old pytorch everywhere? Can we get an option to choose the pytorch version when installing the package?

drmbt avatar Dec 22 '24 21:12 drmbt

GPU selector coming with 2.13.0 which should be in the next couple days. Though for the torch stuff, you can already change the preferred torch index for that specific install, or if you want, override the dependencies entirely - image

We aren't specifying a torch version for Comfy though, perhaps its due to the extra index url, will take a look.

mohnjiles avatar Dec 22 '24 21:12 mohnjiles

Is it possible to select stable cu118 torch CUDA index for older GPUs (https://download.pytorch.org/whl/cu118) ?

nouai avatar May 03 '25 20:05 nouai