[enhancement]: Allow stable diffusion to use multi-GPU resources to speed up single image generation #8377
Description
Is there an existing issue for this?
- I have searched the existing issues
Contact Details
No response
What should this feature add?
Data Parallelism has taken off in multiple areas for stable diffusion. VRAM pooling across several GPUs is the most exiting in my opinion but may required significant code refactor. On the other hand, there is low hanging fruits that seems to be worth the effort like this one:
"In this code sample, i refactored the txt2img pipeline, but other pipelines such as img2img are similar concept. in a standard pipeline, diffusers need to generated the text guidance latents vector and unguidance latents for generate one single images, this logic is related to CFG_Scale settings, this step requires the text_encoder and unet to work together, but actually these 2 images can be generated in parallel from different GPU resources and pull then together into one GPU to sum the result."
https://github.com/hellojixian/StableDiffusionParallelPipeline
Hopefully, it may steer the developers towards supporting more "multiple GPUs" features in the future.
Alternatives
No response
Additional Content
No response