Nvidia 50 series cards...anyone have it working?

[Sandmanu]

I'm running Linux (Ubuntu 25) and the Swarm Installer/Manual Install directions drop me into the Land of All Errors with a 5070ti card. It started with the wrong Python error (resolved by venv and a new Python installation there (3.11.9) that got me further along the install process), then came cuda and pytorch errors, which seemingly can't be resolved by latest stable or overnights.

Yes, I know it's at the ComfyUI stage. Swarm is fully installing now (after the mentioned changes) and starting, but throws the cuda/torch errors (no kernel image available, etc etc). I know that Swarm is at the mercy of Pytorch and other drivers here, but has anyone successfully installed it with a 50 series card?

If so, what changes did you have to make for these newer cards?

[mcmonkey4eva]

Yes, Swarm works with 50xx on both Windows and Linux.

The hardest part is getting the drivers to work. Make sure eg nvidia-smi is returning proper results.

You miiight? need to manually reinstall torch https://github.com/mcmonkeyprojects/SwarmUI/blob/master/docs/Troubleshooting.md#common-error-messages to the 12.8 nightly --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128

[Sandmanu]

did that. Reinstalled SwarmUI. No errors until image generation and then we get :

ComfyUI execution error: CUDA error: no kernel image is available for execution on the device
CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1
Compile with TORCH_USE_CUDA_DSA to enable device-side assertions.

So of course, the 5oxx series has its versions of Cuda, the old versions of Python and Pytorch required by Swarm have their own preferences and that all adds up to a nonfunctioning environment, at least with Ubuntu 25.04 and my 5070ti card. I don't want to run this in Windows. Do you have it running properly in Linux with a 50xx card?

[mcmonkey4eva]

Yes, it works on Linux with blackwell cards.

[Sandmanu]

with which versions of cuda or which Nvidia drivers?

[AshtakaOOf]

Send the output of nvidia-smi ?
Note: nvidia-smi requires you to have cuda installed.

[AshtakaOOf]

Now I'm going to have to ask if you're using a local Linux environment.

I've been doing AI image generation on Linux since late 2022 if that answers your question.

[Sandmanu]

(SwarmUI)/dlbackend/ComfyUI/venv is the navigation structure it created running the ./install-linux.sh script as per your instructions for "Installing on Linux". And it doesn't matter if I initiate from the active venv (and it certainly won't install if I don't) or use the cd/ SwarmUI ./launch-linux.sh to get into the interface. Both load and land the same way, both throw the cuda errors.

So which cuda versions are working with this?

[Sandmanu]

I've been doing AI image generation on Linux since late 2022 if that answers your question.
It doesn't, since I asked if you were using a local version, not a Runpod or Collab version that someone else is maintaining the backend on.

[AshtakaOOf]

It doesn't, since I asked if you were using a local version, not a Runpod or Collab version that someone else is maintaining the backend on.

Yes I'm doing things on my local machine.
No I have never touched runpod.
Please don't assume I have no experience with what I'm talking about.

[AshtakaOOf]

(SwarmUI)/dlbackend/ComfyUI/venv is the navigation structure it created running the ./install-linux.sh script as per your instructions for "Installing on Linux". And it doesn't matter if I initiate from the active venv (and it certainly won't install if I don't) or use the cd/ SwarmUI ./launch-linux.sh to get into the interface. Both load and land the same way, both throw the cuda errors.

Yes it will install even if you don't initiate the venv.

So which cuda versions are working with this?

Please activate said venv (at dlbackend/comfy/venv) and paste the following command.

pip install --pre -U torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128

This should install torch nightly with cuda 12.8 over torch 2.7.1.
See if you still get the error when generating.