nemo-mbridge-perf-moe-dispatcher-selection

nemo-mbridge-perf-moe-dispatcher-selection is an NVIDIA decision-support skill that helps engineers choose the right Mixture-of-Experts (MoE) token dispatcher — alltoall, DeepEP, or HybridEP — in Megatron-Bridge based on hardware, expert-parallelism (EP) degree, and optimization stage. It solves the problem of picking a dispatcher without expensive trial-and-error by distilling patterns observed across DSV3, Qwen3, Qwen3-Next, and MoE VLM bring-up work into quick decision tables (by hardware and by EP size) and model-family guidance.

The skill's main value is its rounded evidence summary. It stresses a backend availability gate: a dispatcher timing is meaningless until the container proves the selected backend package (DeepEP or HybridEP) is actually importable, and alltoall is the correctness fallback when they are missing. It records concrete smoke-run observations (for example a Qwen3 30B A3B run on H100 with EP=16 where DeepEP and HybridEP selected the flex backend but failed before the first iteration because the packages were not installed) and broad MFU trends on GB200/GB300 NVL72 systems, while repeatedly cautioning that dispatcher wins are workload-dependent and should be tested rather than assumed.

It targets performance and infrastructure engineers training large MoE models, or those tracing an MoE regression or crash to a dispatcher config change. The output is configuration guidance and analysis in Markdown with inline bash — no code execution, credentials, or destructive actions. NVSkills-Eval reported a 100% Security score.