easydiffusion
Speed with Multiple GPUs in easydiffusion
OS: Windows10 Two GPUs，3090 and 4090 Compared to Stable Diffusion with a GPU(4090) and "--precision full --no-half --xformers" , it's the same speed!
can you confirm the same thing I reported here: https://github.com/easydiffusion/easydiffusion/issues/1760
Where the overall speed gets split between two cards for no acceleration?
being same speed isn't really that bad of a thing, the fact that easy diffusion can't use split vram is so much worse. I have 2 12gb gpus. I just wish this could actually access all 24gb of ram instead of just 12gb. Koboldcpp can access 24gb of vram. Easy diffusion spawns 2 "python" processes in nvidia-smi, it just only uses all; the vram on one of them instead of 2. Despite having enough vram, I have to shut down all my other ai crap if I want to generate images just because this thing can't use split vram.
@Xeraster You can span an LLM across multiple GPUs because of how they work. LLMs function in a completely different way than the way stable diffusion does. Due to the underlying architecture of how works Stable Diffusion it's not possible to span across GPUs in the same way. That's why no SD UI does this. It's a fundamental issue with the architecture itself.
Merging into https://github.com/easydiffusion/easydiffusion/issues/1760
This is being worked on in an upcoming version (where each GPU will spawn a separate process). Right now, a single process is used and multiple threads are created (one per GPU), which results in Python's "Global Interpreter Lock (GIL)" becoming a chokepoint.
I'm working on using completely separate python processes, one per GPU.
The other python process noticed by Xeraster isn't meant for rendering, it's just a temporary python process. ED currently uses only one python process.
