SpecServe

“These SLO-oriented speculation techniques have two key problems: (i) they are designed for non-latency critical scenario of batch sizes that make decoding closer to compute intensive "knee" of the GPU, and (ii) they employ analytical modeling to predict model execution time, as they cater to dense models. Single-batch MoE serving is highly memory bound, rendering OI-centric heuristics uneffective. Moreover, analytically modeling MoE execution time would not work, as the verification time varies depending from request-to-request and even across iterations.”