Update rocm to 5.4.x

[hydrian]

Update docker build to use ROCm 5.4.x for better card support. The docker image is still using 5.1.x.

[winstonma]

I downloaded @l1na-forever's script and rebuild the whole thing (which install ROCm 5.4.2). And I found that I got the following error:

/var/lib/jenkins/pytorch/aten/src/ATen/native/hip/IndexKernel.hip:94: operator(): Device-side assertion `index >= -sizes[i] && index < sizes[i] && "index out of bounds"' failed.

I have to follow previous pytorch installation and install Pytorch v1.13.1 to start stable diffusion. The installation command is as follows:

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

@hydrian If you want to update the Pytorch, I can you can

• download this image, go into shell prompt and then run the above command, or
• download this repo, modify Dockerfile with the command above and build the image

[hydrian]

@winstonma I've moved on from the l1na-forever image. I've moved over to my own image. It currently runs a more recent version of SD and pytorch 2.x. If you interested in it, you can find it at https://github.com/hydrian/stable-diffusion-webui-rocm

[winstonma]

@hydrian Thanks. I tried but I have no idea why the docker version didn't work for me. The Stable Diffusion WebUI works when I press the Generate button nothing happens. I checked the rocm-smi GPU is running 0%.

By the way I am using AMD 6800U (built-in 680M GPU), running 2GB of dedicated memory. Even the memory is insufficient (I am asking the laptop manufacturer to modify the amount of dedicated memory) but on my Ubuntu 22.04 I can still start generating image.

Thus I have no idea how to troubleshoot the problem

[hydrian]

@winstonma With only 2gb of Vram, use the - - low-vram argument or something like it. I can't remember the exact argument name. SD in stock config needs at least 8GB of VRAM and 8 GB of system memory.

That wait when you click generate the first time per load would be the model being loaded into ram. If you don't have enough as system ram, then it will start using swap which is even slower so it may look like SD is hanging.