AMD GPU Support?
Is it possible to run this with an AMD GPU? I followed the install instructions and it's stuck at 0%. I installed the rocm version of pytorch and changed my device ID to match my 7900xtx and it's not generating anything.
i have it running using comfyui-Zlulda on a 6950
We currently don’t have any modern consumer-grade AMD GPUs available for testing. You may want to wait for further updates from the open-source community.
That's a shame, I would also like AMD support.
I have a succesful torch ROCM setup.
(ace_step) user@host:~/ACE-Step$ python3 -c 'import torch; print(torch.cuda.is_available())'
True
But when launching acestep, i get the following error:
OSError: libtorch_cuda.so: cannot open shared object file: No such file or directory
i have it running using comfyui-Zlulda on a 6950
How'd you get it working? Personally, I'm on Linux, but I guess I could install Windows to go the zluda route.
I don't have a Linux setup to test on currently, but for anyone looking to run on Windows, I can confirm that it works with comfyui-zluda on my 7800.
Edit:
I wanted to try this using ROCm on WSL. So, I've taken about a day or so to see if I could get it working. I followed the instruction here:
https://github.com/comfyanonymous/ComfyUI/issues/6434
(Updated for Ubuntu 24.04 with HWE and ROCm 6.4)
This will work for ComfyUI and I was able to do the basic text to image workflow with no problems.
However, running ACE-Step did not work because of the VAEDecodeAudio node. It gets stuck there with a memory error will just keep giving the same errors over and over and never finish. If anyone knows of a possible fix, I would be glad to try it out and let you know if it works.
In case it is helpful, the messages I am getting look like this:
MIOpen(HIP): Warning [IsEnoughWorkspace] [GetSolutionsFallback WTI] Solver , workspace required: 31653888, provided ptr: 0 size: 0
MIOpen(HIP): Warning [IsEnoughWorkspace] [EvaluateInvokers] Solver , workspace required: 31653888, provided ptr: 0 size: 0
Edit number 2:
There are some unofficial python packages for pytorch for ROCm available here: https://github.com/scottt/rocm-TheRock/releases/tag/v6.5.0rc-pytorch-gfx110x
I followed the instructions from here: https://ai.rncz.net/comfyui-with-rocm-on-windows-11/ to install comfyui on windows without installing the HIP SDK
This worked for me and I had a running version of comfyui which I tested by doing some basic text to image generation. When trying to run the ACE-Step workflow, I got the same error as running in the Windows Subsystem for Linux as I described in my first edit.
Since this is not a problem with the zluda version, it appears that this problem is something specific to the ROCm pytorch libraries where they think that memory cannot be allocated by VAEDecodeAudio even though memory is available.
In case anyone is still reading this thread, I will leave my installation around for a few days so I can try any suggestions or possible fixes anyone might have