Moonshot AI Open-Sources MoonEP: A Perfectly Balanced Expert Parallelism Library for MoE Training

MarkTechPost · 2 months ago · 30 ● 33 sources

Moonshot AI open-sourced MoonEP, a library for efficient distributed Mixture-of-Experts training that guarantees every GPU rank receives exactly S × K tokens regardless of router imbalance. The library achieved a claimed 2.5× improvement in scaling efficiency for their Kimi K3 model, a 2.8-trillion-parameter MoE system, by using redundantly planned experts and static buffer shapes to eliminate communication overhead and memory fragmentation. MoonEP's approach removes per-layer host synchronization and makes communication latency nearly immune to router skew, allowing frameworks to scale MoE training more reliably across many GPUs.