> There is not way it's sending experts to the GPU per token.

Right, it seems like either experts are stable across sequential tokens fairly often, or there's more than 4 experts in memory and it's stable within the in-memory experts for sequential tokens fairly often, like the poster said.