AWS gets 40% more MoE RL rollout throughput by running DeepEP v2 over EFA.

The writeup covers porting DeepEP v2's dispatch and combine kernels from CUDA-specific RDMA to libfabric so expert-parallel all-to-all runs over EFA. Running GRPO on a GLM-5 744B-class MoE across 48 P5en instances, 16 training and 32 inference, aggregate rollout throughput rose 40% over the Slime NCCL baseline. The recipe is in the open-source Miles 0.1.0 framework.