martial

That log looks very bare, as if nothing happened. The previous step completed (installing torch), then it goes into installing Comfy's requirements.txt which should run uv pip install -r and be a little bit more verbose, especially in case of error. Here it is like it was not able to start the command.

When you deleted the run folder you downloaded a fresh install of ComfyUI during the next startup, so we can expect this to be something else. Doing a quick gpt search seems to say that it might be a CPU issue and you might have to compile kornia from source. You can try to use the userscript (check the README for instructions) to attempt to run the build manually: /comfy/mnt/venv/bin/pip install kornia-rs --force-reinstall --no-cache-dir --no-binary :all: (also copied from the GPT answer)

The crash comes from ComfyUI itself so I would not be able to comment too much on what the issue is. When I see this kind of behavior, I rename the run folder to run.old and re-run the container which will re-create a new installation. In general this fixed things for me, knowing that the custom nodes and images are in basedir, everything in "run" should be able to be deleted.

The error I see in there is RuntimeError: Detected that PyTorch and TorchAudio were compiled with different CUDA versions. PyTorch has CUDA version 13.0 whereas TorchAudio has CUDA version 12.9. Please install the TorchAudio version that matches your PyTorch version. You said you have a Blackwell card. I would try with the CUDA 13.2 container, not the 13.0 (or are you trying 12.9)?

And you are not using TORCH_LOCK, correct? In your case, my belief is that Torch is trying to update to 2.12 but as noted above, they dropped support for CUDA 12.8 in Torch 2.12 Details: "Introducing CUDA 13.2 and Deprecating CUDA 12.8 (Release 2.12)" https://dev-discuss.pytorch.org/t/introducing-cuda-13-2-and-deprecating-cuda-12-8-release-2-12/3337 The two options I would recommend are either: 1) use TORCH_LOCK=torch==2.9.1+cu130 torchvision==0.24.1+cu130 torchaudio==2.9.1+cu130or a more recent Torch (as long as you get the right combination, I am not sure what it is honestly :) ) 2) change the base container from ubuntu24-cuda12.8 to a more recent one. Important note: latest is likely going to move from CUDA 12.8 to CUDA 12.9 because of that (likely this week end)

Is everyone having a problem on 7.3.1?

Understood, maybe other will encounter the same issue and report their solution. I do not have the issue or am able to reproduce it currently, so I can not help further.

In the Settings/nvidia-driver page of my Unraid, I see in the GPU Driver Support a note: Supported candidates: 580.159.03, 590.48.01, 595.71.05, 595.80, 610.43.02 Recommended Driver: 595.71.05 (open-source) Apparently I am running the expected "recommended" driver.

The thing is that it should not be fighting you at all ... I run 3x instances of the container with my primary on an RTX 4090 Linux Mint using the CUDA 13.2 base and it works fine. The container's part appear to be working: it installs ComfyUI, then starts it, it is ComfyUI that then crash (I did not see anything wrong in the logs that you shared with us). The 3090 supports the latest driver with CUDA 13.2 I believe. If you ssh into your Unraid --or use the >_ terminal-- and type nvidia-smi you will see by the top the max version of CUDA that the host (and therefore container runtime) can support, my Unraid runs Driver Version: 595.71.05 CUDA Version: 13.2 (ie I have not yet upgraded to the 610.43.02 driver that appears to be available) so I can use the 13.2 version (and do not use TORCH_LOCK) From the release notes, I do not see any expected issue but maybe others that have a more recent driver can say what they see? https://forums.developer.nvidia.com/t/610-release-feedback-discussion/371356

Importantly it also tells us that Comfy started so your installation is (from the view of the container's purpose) clean

That error just tells us that ComfyUI exited with an error. Everything in the second log is just normal ComfyUI starting with PyTorch 2.11.0+cu129 and Comfy 0.23.0 "when" it fails is shows the line that caused the error, here the call to Comfy which is normal. If you are comfortable with this, can you delete the container (do not delete the folder where your files are in /mnt/usr/appdata). Then move the run folder that should be in /mnt/user/appdata/comfyui-nvidia/mnt to "mnt.old" (not the basedir, it contains your generations and models) From that Community Apps, reinstall a new ComfyUI, use at least CUDA 12.9 and it should recreate the "mnt" folder with a fresh installation?

The image uploaded should have the right info (apparently it did not copy): Beyond that you should have nothing to do, the tool create a venv specific to the "ubuntu24-cuda12.9" at first start.

Keep in mind if you use the template that you can select the CUDA version you want to run against. I have made 12.8 the default but with the depreciation of 12.8 for PyTorch I might be forced to make 12.9 the default "latest" shortly. It is recommended to manually select which version of CUDA you want to run against (if your Unraid Nvidia driver is up to date you should be able to run 13.2 ). You can find it from the same "edit" page as "Repository": The list of possible values is in the table at https://github.com/mmartial/ComfyUI-Nvidia-Docker

I run the ubuntu24_cuda12.9 version on Unraid. It uses: torch==2.11.0+cu129 torchaudio==2.11.0+cu129 torchvision==0.26.0+cu129 From the "Docker" tab, click on the application icon and use "Edit" So I could add "Add another Path, Port, Variable, Label or Device" with the following parameters:

Good news is that cu132 finally has it own download page https://download.pytorch.org/whl/cu132 Bad news ... uv does not yet support it, so not moving yet