Share MoE model in RAM across multiple-GPU

[hepengfe]

Is it possible to share MoE model stored in RAM across multiple-GPUs?
Suppose there are 4 GPUs, it can handle 4 inference requests at the same time. All GPU can load MoE tensors from RAM for inference.

[fernandaspets]

Same question for mixed bag of 7 gpus and model Kimi K2