Multi-GPU workflow for video gen?

[stopandgo-44]

Good morning everyone, first of all, really appreciate the work and effort on this set of tools to helping people get into local LLM/visual gen.

So, I'm thinking of getting into video generation, and I was reading through the recent patch notes. There's a section where offloading to CPU is recommended for certain tasks:

"It's recommended to use --sdoffloadcpu and --sdvaecpu if you don't have enough VRAM. The VAE buffer can really be huge."

Here's the thing, what if I have two identical GPUs? Can I instead offload anything to the second GPU? I know you can't straight up use dual GPUs to generate any one frame any faster, but surely I can hand off the next parts of the processing pipeline to the second GPU? That way the CPU doesn't have to be the fall-back. How to achieve this? I have two AMD GPUs in my system so I will be trying this out later tomorrow and see what I can do! Any help or pointers are appreciated to take maximum advantage of both GPUs.

[LostRuins]

Unfortunately not, stable-diffusion.cpp does not support running any part of the inference of different GPUs. The only possible way would be some kind of Nvidia SLI which is no longer available.