MoonshotAI/MoonEP

Health v2 · maintenance only

Flags: no_releases young

Adoption not part of the score

1101 stars · 122 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

MoonEP is an Expert Parallelism communication library for Mixture-of-Experts training that keeps token loads perfectly balanced across ranks using dynamically planned redundant experts. It features zero-copy token dispatch, static buffer shapes, and online GPU planning, outperforming DeepEP v2 under skewed routing.

Use cases

• train MoE models without load imbalance slowing down ranks
• avoid OOM from dynamic activation shapes in expert parallel training
• speed up dispatch/combine communication for expert parallelism
• benchmark expert parallelism communication against DeepEP
• run MoE training on H20 GPUs with EP=8

When to choose

• training large MoE models where router imbalance degrades throughput
• you need static memory shapes to avoid fragmentation and OOMs
• you want faster dispatch/combine than DeepEP under imbalance

When to avoid

• you don't use Mixture-of-Experts architectures
• you need non-NVIDIA hardware (Zhenwu PPU support is incomplete)
• you need a mature, widely-adopted library with broad community support

Facets

library · maturity active

machine-learning llm-training gpu-computing concurrency machine-learning deep-learning large-language-models gpu-computing microservices python expert-parallelism mixture-of-experts moe communication-library load-balancing training-infrastructure nvidia-gpu linux gpu

Member repositories

For agents

markdown · JSON · MCP: product_card(name="MoonshotAI/MoonEP")