Mixture-of-Experts (MoE) language models send each token through a few of many expert subnetworks. Serving methods make these models faster by running even f…
机构:USC
来源:arXiv 2611.06147 | AI4Papers 论文推荐平台