PIPEMORPH: Treating Pipeline Parallel Training as a Runtime Control Problem

Distributed training systems often optimize the steady state. They choose a pipeline schedule, minimize bubbles on a healthy cluster, and assume the execution order will remain useful for the next ite