SeqMoE: Toward Full-Load Performance via Predictive and Graph-Compatible MoE Offloading
This work is the first to recast expert activation prediction as sequence modeling, enabling accurate multi-step, multi-layer forecasts that provide a long and reliable window for downstream decisions, and derives general runtime principles encompassing compute-transparent expert placement and synchronization-free orch...