MoE is mostly used to enable load balancing since it makes it possible to put experts on different GPUs. This isn't so easy to do with a monolithic, but sparse

MoE is mostly used to enable load balancing since it makes it possible to put experts on different GPUs. This isn't so easy to do with a monolithic, but sparse layer.