UltraEP: Unleash MoE Training and Inference on Rack-Scale Nodes with Near-Optimal Load Balancing
UltraEP:在机架级节点上以近最优负载均衡释放MoE训练与推理
机构 * School of Computer Science, Peking University(北京大学计算机学院) ; Xiaohongshu Inc.(小红书公司) ; Shanghai AI Laboratory(上海人工智能实验室)
AI总结 提出UltraEP,首个基于精确负载的实时均衡器,通过协同设计规划求解与专家复制通信,在机架级节点上实现MoE训练和推理的微批次与逐层重均衡,达到94.3%的力均衡理想吞吐量。
Comments The authors have identified issues related to information disclosure in the current version of the manuscript and therefore request its withdrawal. A revised version may be prepared at a later date