Megatron-Bridge

NVIDIA Megatron-Bridge + UCCL-EP for MoE training with expert-parallel all-to-all over EFA

View on GitHub

Overview

NVIDIA Megatron-Bridge enables Mixture-of-Experts (MoE) training on Amazon EKS with UCCL-EP carrying the expert-parallel all-to-all over AWS EFA — without patching Megatron-Core.

Key Features

  • EFA-native expert parallelism — No InfiniBand required
  • UCCL — Unified Communication Collectives Library for EFA
  • Model-agnostic — Works with any MoE architecture on Megatron-Core
  • No source patches — Plugs in via top-level deep_ep module
  • DeepEP replacement — NVIDIA DeepEP uses NVSHMEM + InfiniBand verbs; this provides EFA-native alternatives