Support nvidia blackwell (5090)

[ImamJanjua]

Hi I wanted to know if there are any plans to support new gpus?

[Zundrium]

I hope so too, sounds like all we need to do is update the base image to support CUDA 12.8?

Edit: Oh, comfy-cli doesn't support 12.8 yet, so it makes perfect sense.

[TimPietrusky]

@ImamJanjua @Zundrium i have seen ComfyUI running on 5090, so it is working in theory. we just need to check what we need to change in our Dockerfile (or create one just for 5090), so that you can use it there as well.

[kodxana]

@Zundrium @ImamJanjua You might want to give a try: https://hub.docker.com/r/madiator2011/worker-comfyui/tags

[TimPietruskyRunPod]

Blackwell (5090, 5080, B100, B200) is supported by Runpod GPUs and runs the worker fine. The 5.8.5 base image ships with PyTorch 2.10.0+cu128 (CUDA 12.8) and ComfyUI 0.17.2, which is sufficient. The hub config was previously locked to `allowedCudaVersions: ["12.7", "12.6"]` which excluded Blackwell hosts (they run 12.8+), so deploying from the hub onto a 5090 would land on an incompatible host and crash-loop with `nvidia-container-cli: requirement error: unsatisfied condition: cuda>=12.6`. That's fixed in PR #222 which expands the allow-list to include 12.8/12.9/13.0.

There's also a separate `base-cuda12-8-1` target in `release.yml` for users who want a cu128-specific image (different PyTorch wheel index).

I tested today on an RTX 4090 and it worked. Don't have a 5090 to test on, but the only known blocker is the CUDA allow-list (#222).

Once #222 lands and a release goes out, please retest. Closing speculatively.