Preprint
Jul 2026
ELDR: Expert-Locality-Aware Decode Routing for PD-Disaggregated MoE Serving
From a request's prefill expert activations, ELDR builds an expert signature predicting the experts it will activate during generation, which reduces median TPOT by 5.9-13.9% over the strongest of four load-balancing baselines across three MoE models and two workloads.
Sang-Jun Choi, Sukmin Cho, Yifan Xiong et al.
· 1 citation
· ⚡1