Comfy_UI:Attempting to use hipBLASLt on a unsupported architecture!

Thank you for the update. I’ve already tried running it on python 3.10.15 using pyenv. I’ve also tested using Fedora, Ubuntu, and recently Arch containers, both on python 3.10.15 and 3.13.2. No luck. I haven’t tried 3.12 though. Also If you have a suggestion for any python+container environment setup I’d be more than happy to give it a try.
Furthermore:
Is there a definitive way to rule out possible hardware issue?
And
if I find a way to upload a 4Gb core dump would it help? jaglinux above, suggested that I backtrace the issue using the core dump but I don’t know how to.

As Tim mentioned on the Github issue, the next thing to check is ensuring that you’re on Linux kernel 6.11. The one you’re using (6.12) is a minor version ahead of ROCm’s most recently supported Linux kernel.

Even if you ran a container, the fact is that the container leverages the amdgpu-dkms drivers from the host system. If amdgpu-dkms is not compatible with the host system linux drivers (and it’s not yet fully functional for 6.12), then you’ll run into problems.

To rule this out, you really need to try on an OS with a supported Linux kernel. See System requirements (Linux) — ROCm installation (Linux) for some details

I’ve installed Ubuntu 22.04 on a fresh hard drive:
Linux sy-lu 6.11.0-19-generic #19~24.04.1-Ubuntu SMP PREEMPT_DYNAMIC Mon Feb 17 11:51:52 UTC 2 x86_64 x86_64 x86_64 GNU/Linux
With Python 3.12.3, and the following pytorch verrsion:

Name: torch
Version: 2.8.0.dev20250314+rocm6.3
Summary: Tensors and Dynamic neural networks in Python with strong GPU acceleration
Home-page: https://pytorch.org/
Author: PyTorch Team
Author-email: packages@pytorch.org
License: BSD-3-Clause
Location: /home/sersys/pyproj/torch/lib/python3.12/site-packages
Requires: filelock, fsspec, jinja2, networkx, pytorch-triton-rocm, setuptools, sympy, typing-extensions
Required-by: kornia, spandrel, torchaudio, torchsde, torchvision

Comfy_ui version: v0.3.26

This works and I can generate images.

1 Like

So apparently there’s no way to use Rocm properly on the 6.12/6.13 kernels, at least at the time of writing. I’d like to thank @fluidnumerics_joe for his help and a big shout out to @tcgu-amd (tcgu-amd (Tim Gu) · GitHub) at github for his invaluable support as well.

1 Like

Rocm 6.3.4 works OK with 6.12 kernel (even without amdgpu-dkms).

I had the same problem (also on Radeon RX 7900 XT). The cause turned out to be the pytorch version. How to Install ComfyUI on Linux | ComfyUI Wiki gives commands to install 6.0 or 6.1. And on 6.1 I got segmentation error (on 6.12 and 6.11 kernel). After update to 6.3 everything works ok.

kernel: 6.12.21-amd64 #1 SMP PREEMPT_DYNAMIC Debian 6.12.21-1 (2025-03-30) x86_64 GNU/Linux
amdgpu: from kernel
torch: 2.8.0.dev20250420+rocm6.3
rocm: 6.3.4.60304-76~24.04