[Feature Request] Affinity of a model to a certain GPU

[charleswg]

I wonder if there's any reason why we cannot just set an affinity of a model to a GPU if we have multiple? That to me would simplify it as if I'm running multiple models, they are not both split into for example 2 GPUs. I understand setting it to auto to split into all GPU would simplify user experience but it'd be good to have that option so that model weights don't need to be splited between multiple GPUs unnecessary and no communication necessary between GPUs. Another option for example if I have two GPU with one having more VRAM than the other, it'd be wise to put the larger model on the GPU with more VRAM than just default to splitting.