It’s KINDA being done in the ways we run MoE models. Strata makes more efficient use of limited VRAM if you have a decent amount of RAM, but you don’t even need to keep the entire model in RAM with it, it can keep the least frequently used experts on disk instead (it being an NVMe SSD hopefully, otherwise if you hit those experts, good luck)
It’s KINDA being done in the ways we run MoE models. Strata makes more efficient use of limited VRAM if you have a decent amount of RAM, but you don’t even need to keep the entire model in RAM with it, it can keep the least frequently used experts on disk instead (it being an NVMe SSD hopefully, otherwise if you hit those experts, good luck)