MoonshotAI Open-Sources MoonEP, an Expert-Parallelism Library Whose Comm Time Stays Flat as MoE Routing Skews

GitHub·medium signalXBlueskyLinkedInCopy link

MoonshotAI/MoonEP (created 2026-07-24, 956 stars, 102 forks, MIT) guarantees every rank receives exactly S × K tokens regardless of router skew by planning a small set of redundant experts online from current router outputs and prefetching them, then reducing their gradients back to home ranks in the backward pass. Benchmarked against DeepEP v2 on H20 with 8 EP ranks, MoonEP's communication time stays almost flat as maxvio grows while DeepEP v2 degrades steadily, and training iteration time stays flat at every imbalance level. The zero-copy path writes tokens straight into expert-grouped positions on remote ranks with no permute in or out, and static shapes eliminate per-layer MoE host synchronization — and the OOM failures DeepEP hits under high imbalance.