AITemplate VS rocm-gfx803

VoltaML is a relatively vanilla diffusers-based backend, so its not a hairy monster to hack like you may have seen with SAI-based UIs.

The AITTemplate code is a lightly modified version of Facebook's example, code, to get rid of small issues like VRAM spikes: https://github.com/facebookincubator/AITemplate/tree/main/ex...

InvokeAI is also diffusers based, but they seem to mess with the pipeline a bit more.

And anyway, all that may be a better reference for interesting features rather than a backend to try and adopt.

let me know if you discover any more ways to improve SD. I am currently looking into facebooks AITemplate : https://github.com/facebookincubator/AITemplate

Link to AITemplate itself: https://github.com/facebookincubator/AITemplate

That specific card has current support for rocm and that is supported by at least tensorflow and torch, plus many other less known/used libraries like cupy, although you are correct in the fact that support sucks in the long run, I have a GPU that is known to be useful and that has continued COMMUNITY support because AMD cut the support with rocm 4.0, thanks to Xuhuisheng for the patch to make the rx580 work with current rocm despite AMD lack of support, what open source can accomplish https://github.com/xuhuisheng/rocm-gfx803

You'll also need to compile pytorch and torchvision for gfx803, although I recommend you install the whl files from here inside your venv because it's a massive pain to compile them on non-Ubuntu (I tried)

I followed the guide from here: https://github.com/xuhuisheng/rocm-gfx803

There is an unofficial build of ROCm 5.2.0 + pytorch + torchvision with GFX8 support added back in. I have no idea if it works. Perhaps someone who knows Docker/Conda could get SD working with those files.