amd gpu on ubuntu,no hip gpus are available

[PennyFranklin]

my sd-webui shows zhese errors at launch interface. And though the webui can run pictures, it's working by the cpu,8~9s/it.by the way,my gpu is 7900xtx.
ubuntu22.04.2 rocm5.4.2 pytorch2.0.0 pytorchvision0.15.1

[ClashSAN]

try export HSA_OVERRIDE_GFX_VERSION=10.3.0?
(the rocm version more tested is 5.2)

[PennyFranklin]

should I put this code to the "webuiuser. sh "?

[DutchMike64]

There seems to be an issue with rocm 5.2 not being available anymore, i tried with changing the webui.sh script to include 5.4.2 which is the current one. But that doesn't work as well. After much trying i gave up and went to EasyUI. I don't know my way around linux that much sadly. Automatic1111 script does work in windows 11 but much slower because of no Rocm.,,,,

Please AutomaticMan update the scripts so it's working in linux again for AMD users.

[PennyFranklin]

how do it work on EasyUi? if AMD' speed is ok that I'll have a try for that

[DutchMike64]

I am using EasyUI at the moment and that works very well, i get about 5/6 iterations per second on a 6700XT for the usual benchmark with 512x512. Everything works fine on linux mint, models and lora from huggingface and civitai.

[PennyFranklin]

well, i gave up to try on my 7900xtx temporarily because i find a automatic1111 package that works quite efficient on my NVIDIA 2060, and it work at the speed of about 5it per second with 512x
512, and it can produce 1080p works! https://gitlab.com/Nanhai2007/novelai-webui-aki，but it use the language of Chinese. I hear that the AMD is about to bring rocm to Windows in their GPU driver , so I'll wait for that (i have enough of ubuntu)

[DutchMike64]

Try Mint or kubuntu (KDE on Ubuntu) if you don't like Gnome. Plus you can run google colab for Stable Diffusion and you'll get about 6it per second, while you can still do other stuff as colab is a server :-)

[Sakura-Luna]

Try uninstalling and reinstalling PyTorch.

pip uninstall torch
pip install --upgrade torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm5.4.2

[PennyFranklin]

oh I tried it many times but not effect

[iDeNoh]

The 7900xtx isn't supported by ROCM yet.

[PennyFranklin]

o,quite sad to hear:(

[iDeNoh]

I would imagine we should see movement on that front no later than april

[PennyFranklin]

I'll wait for that, but before I gonna use my 2060 to run on win11, about 5it/s with 512x512, not bad

[elen07zz]

Install an older version
pip install torch==1.13.1+rocm5.2 torchvision==0.14.1+rocm5.2 torchaudio==0.13.1 --extra-index-url https://download.pytorch.org/whl/rocm5.2

I was having a similar problem but with invokeai.

[collsni]

Did not work for me on a 7900xtx and 7950x

I've only been able to get my dual boot windows os to work with https://github.com/nod-ai/SHARK

~8-9 iterations/sec on windows

(venv) [sassypants@arch stable-diffusion-webui]$ vk_pro python launch.py

Python 3.10.10 (main, Mar 5 2023, 22:26:53) [GCC 12.2.1 20230201]
Commit hash: 22bcc7b
Installing requirements for Web UI
Launching Web UI with arguments: --skip-torch-cuda-test --autolaunch
No module 'xformers'. Proceeding without it.
Loading weights [6ce0161689] from /home/sassypants/Downloads/stable-diffusion-webui/models/Stable-diffusion/v1-5-pruned-emaonly.safetensors
Creating model from config: /home/sassypants/Downloads/stable-diffusion-webui/configs/v1-inference.yaml
LatentDiffusion: Running in eps-prediction mode
DiffusionWrapper has 859.52 M params.
/usr/bin/vk_pro: line 42: 22973 Segmentation fault (core dumped) "$@"

(venv) [ian@arch stable-diffusion-webui]$ pip list | grep 'torch'
open-clip-torch 2.7.0
pytorch-lightning 1.9.4
torch 1.13.1+rocm5.2
torchaudio 0.13.1+rocm5.2
torchdiffeq 0.2.3
torchmetrics 0.11.4
torchsde 0.2.5
torchvision 0.14.1+rocm5.2

[notice] A new release of pip available: 22.3.1 -> 23.0.1
[notice] To update, run: pip install --upgrade pip

[collsni]

Pretty sure this is what is up with the 7900xtx

ROCm/ROCm#1880 (comment)

[Sakura-Luna]

Someone found a docker image that contains ROCm 5.5 RC.

[PennyFranklin]

not a official release?

[Sakura-Luna]

According to the information on Docker, it is official.

[jahu00]

Being a release candidate it's likely to have bugs. Plus building PyTorch can be a challenge. Official release of both ROCm and PyTorch would make things easier. Hopefully, now that they are at the stage where they have a RC version, they will release 5.5 soon (though I heard rumors about AMD wanting to release ROCm on Windows first).

If 7900xtx had ROCm support, I would be very tempted to replace my RX6700.

[chibiskuld]

I'm having this problem when I shouldn't be.
My issues isn't the fact it's not working, but that I don't want it to do this at all.
I have an AMD APU which is using a Radeon Integrated GPU for it's primary graphics and the machine has two tesla P40's in it.
But trying to get it to start, it gets stuck on trying to use HIP when I don't want it to even check for this at all.

Any ideas?

[chibiskuld]

Deleting the section that's pretty much if not NVIDIA from webui.sh seems to have bypassed this for me.