AWS gets 40% more MoE RL rollout throughput by running DeepEP v2 over EFA on EKS
AWS Machine Learning Blog·low signal
AWS ported DeepEP v2's dispatch and combine kernels from CUDA-specific RDMA to libfabric so expert-parallel all-to-all traffic runs over EFA. It then ran GRPO on a GLM-5 744B-class MoE across 48 P5en instances, 16 for training and 32 for inference. Aggregate rollout throughput rose 40% over the Slime-based NCCL baseline, and the recipe ships in the open-source Miles 0.1.0 framework.
AWS Machine Learning Bloglow signal