I also just came across this:

https://huggingface.co/spaces/gemma-challenge/gemma-dashboar...

Agents collaborating to speed up gemma-4-E4B-it inference (tokens per second) on a fixed GPU.