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.