作者位次 #4/12OSDI 2026Oral· Operating Systems & Distributed Systems
UEP: Portable Expert-Parallel Communication
Ziming Mao · Yihan Zhang · Chihan Cui · Zhen Huang · Kaichao You · Zhongjie Chen · Zhiying Xu · Zhenyu Gu · Scott Shenker · Costin Raiciu · Yang Zhou · Ion Stoica
摘要
Modern Mixture-of-Experts (MoE) workloads rely on expert parallelism (EP) to achieve high GPU efficiency. State-of-the-art EP communication libraries, such as DeepEP, rely on GPU-initiated RDMA communication. Although performant, they have poor portability across heterogeneous GPU and NIC hardwar...