Moonshot open-sources MoonEP, a balanced MoE communication layer for GPUs and PPUs
deliprao · x · 2026-07-28
Moonshot’s MoonEP is an expert-parallelism communication library that keeps token loads balanced across ranks using dynamic redundant experts.
The repo describes online planning, a near-optimal GPU planning kernel with negligible overhead, and zero-copy/static-shape optimizations. It is designed to keep every rank receiving exactly S × K tokens even when routing is skewed, and it already targets NVIDIA GPUs while planning support for Alibaba’s TPU-like hardware.
Related event: MoonshotAI open-sources MoonEP for perfectly balanced expert parallelism(2 posts)→
More from Infra
- Embedding databases are hurting search, says a reply in the thread — davidmanheim · 2026-07-28
- Apple reclaims the top market-cap spot, passing Nvidia at $4.938T — mitchdeg · 2026-07-28
- Sponsored GPU contacts are now public, after the list was empty a few months ago — ctjlewis · 2026-07-28
- DeepInfra adds Claude Opus 5 with 1M context and $5/$25 pricing — gharik · 2026-07-28
- Bittensor gets an MCP-style front door for agents to route storage, inference and training — markjeffrey · 2026-07-28
- Post says China is at least a decade away from chips hyperscalers would buy — inductionheads · 2026-07-28