> no high quantisation library supports two GPUs for example, and there seems to be no interest in implementating it by the guys behind the best inference tools

I'm curious to hear what you're trying to run, because I haven't used any software that is not compatible with multiple GPUs.