perf-moe-comm-overlap

To overlap MoE expert-parallel communication with expert compute, set cfg.comm_overlap.overlap_moe_expert_parallel_comm to True. You must also set cfg.model.moe_shared_expert_overlap to False when using the dispatch-overlap path to ensure correct routing behavior.

Enable MoE communication overlap when profiling traces show visible time spent in expert-parallel dispatch or combine steps under EP>1. This tuning step reduces throughput regressions after keeping dispatcher, routing, and recompute decisions stable.

Using MoE communication overlap requires expert_model_parallel_size and num_moe_experts greater than one. The moe_token_dispatcher_type must be set to alltoall or flex, and the model must use supported BF16 or FP16 precision.

Yes, you can add delay_wgrad_compute to increase overlap potential when constraints allow it. Using this optional delayed expert wgrad scheduling requires careful pairing with CUDA-graph and recompute configurations to avoid destabilizing the training process.

Yes, the communication overlap path supports DeepEP-style backends by coordinating the flex dispatcher. It activates compatible MoE dispatcher behavior, enabling overlapping of expert-parallel dispatch and combining operations with expert compute.