How to run comfyui on a blackwell gpu (like 5090) in windows with triton, sageattention, flashattention and bitsandbytes

[aliabougazia]

first do these:

• make sure your gpu driver in windows is updated.
• install wsl ubuntu 24.04
• install docker in wsl (not windows).
• install nvidia container toolkit in wsl.
• pull the nvidia pytorch image https://catalog.ngc.nvidia.com/orgs/nvidia/containers/pytorch.
• run the image as bind mount with your comfyui folder in it.

Inside the running container:

• no need to install cuda in or out of container as it comes with cuda 12.8 already.
• install comfyui depencencies again in the container.
• you may have to uninstall the included torch/torchvision and install the latest of those for cuda12.8.
• you can now compile bitsandbytes as per their github repo and save it as a wheel for later use. you can also use flashattention which comes with the image, and the kijai precompiled wheels of sageattention and triton in the container.
• you can also save the running container as an image to run later instead of doing this everytime.

[alloystorm]

Got this partially working.

Yes you do need to install docker in WSL and can't use docker desktop for this. Otherwise you don't get to use cuda.
The nvidia image comes with pytorch 2.7 and cuda 12.8 but the tricky part is that when you install custom nodes it often revert to pytorch 2.6 again. Do you know anyway to avoid that?
At the moment it seem running on windows with nightly pytorch installed with cuda 12.8 is a better choice if you turn off xformers.

[aliabougazia]

Yes.

method 1: manually:

AFTER pip install requirements and BEFORE running comfyui, do:
$ pip install --upgrade --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128
It will upgrade torch in the container to the latest available torch. if a custom node undoes it, press ctrl+c to exit back to container bash, run this again, then re-run comfyui.

method 2: automated
use comfyui manager blacklist: https://github.com/ltdrdata/ComfyUI-Manager#paths
Comfy-Org/ComfyUI-Manager#980

Also now you need to install frontend in container:
$ pip install comfyui-frontend-package==1.10.17

let me know if this sorts the issue.

[magnuky]

Using docker will run slow if you store safetensors files on windows environment and not within the docker environment which is typical. It’s slower because the large safetensors files need to be converted from ntfs file format to Linux’s for ComfyUI within the docker to use.

After spending much time on setting up docker and ComfyUI, I decided to go back to installing on Windows. If you would like to manage the environments better like I do, use Conda (miniconda)

[MG007Raiser]

@magnuky does mini conda on Windows not have that same side effect of converting safetensors like docker?

[magnuky]

When you run comfyui on windows and storing safetensors files on Windows, both use ntfs, so no conversion needed. Miniconda, also on windows, simply manages the Python version and libraries for your comfyui. All in ntfs.