learned-scheduler

This paper presents a measured study on learned request routing for disaggregated LLM serving, evaluating a calibrated router that uses admission-time features to assign requests, achieving higher mean goodput compared to traditional methods on a real vLLM/NIXL cluster.