AWS 在 EKS 上用 EFA 和 DeepEP 扩展 MoE 强化学习，吞吐量提升 40%
原标题：Scaling MoE reinforcement learning on Amazon EKS with EFA and DeepEP with 40% more throughput
AI 摘要
AWS 机器学习博客介绍了一种在 Amazon EKS 上结合 EFA 与 DeepEP 扩展 MoE 强化学习训练的架构，声称吞吐量提升 40%。文章指出大规模 RLHF/GRPO 训练面临 rollout 生成与策略训练资源竞争、计算/内存/带宽平衡、以及从节点内 NVLink 到节点间低带宽通信转变三大挑战，其中专家并行（EP）的动态 all-to-all 通信是主要瓶颈。该方案通过 EKS 编排、EFA 加速网络和 DeepEP 优化专家并行通信来应对这些问题。
正文节选
Scaling MoE reinforcement learning on Amazon EKS with EFA and DeepEP with 40% more throughput When you post-train a Mixture-of-Experts (MoE) model with Reinforcement Learning from Human Feedback (RLHF) or Group Relative Policy Optimization (GRPO) at scale, three simultaneous challenges emerge. The first requires coordinating heterogeneous compute for rollout generation and policy training. Second, sustaining high-throughput communication across hundreds of accelerators. And third, dynamically or