1 paper
Pengfei Xia, Tuo Hao, Shengwei Li +5
An MoE appliance for trillion-parameter models at low concurrency must host terabytes of weights on one node and serve prefill and decode with fixed resources. Combining operator a…