moe-layout tool improves Megatron-LM pipeline layouts for MoE MTP

An engineer developed moe-layout, a plain Python tool that uses two short profiling runs to simulate and determine optimal pipeline parallelism layouts for Mixture-of-Experts models with multi-token prediction. Tests across hardware configurations showed the tool's chosen layouts ran significantly faster than Megatron-LM defaults and count-based rules, achieving step time predictions within 2.7 percent accuracy. Open article →