skills.rest
sayalinvidia
sayalinvidiaCOMMUNITY
Explore all skills by this author →
GitHub

nemo-mbridge-perf-moe-dispatcher-selection

Select the optimal MoE dispatcher for hardware and expert-parallelism configurations.

Updated Apr 16, 2026
Search Tags:#deepep#megatron-bridge#nvlink#hybridep#alltoall#moe-dispatcher#dispatcher-selection
One-click install
npx skills add https://github.com/sayalinvidia/sayali-skills-test --skill nemo-mbridge-perf-moe-dispatcher-selection
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: nemo-mbridge-perf-moe-dispatcher-selection
Source: https://github.com/sayalinvidia/sayali-skills-test/tree/main/skills/nemo-mbridge-perf-moe-dispatcher-selection
Command: npx skills add https://github.com/sayalinvidia/sayali-skills-test --skill nemo-mbridge-perf-moe-dispatcher-selection

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

Helps engineers select the optimal MoE token dispatcher (alltoall, DeepEP, or HybridEP) for Megatron Bridge workloads, reducing trial-and-error and aligning with hardware and EP configurations.

Core Features & Use Cases

  • Hardware-aware dispatcher guidance: Recommends the best dispatcher per hardware (H100, GB200, GB300 NVL72) and EP degree.
  • Back-end and tuning guidance: Documents the required flex backend settings (deepep, hybridep) and default SM tuning knobs, with notes on environment limitations.
  • Use cases: Useful when bringing up MoE models, tracing regressions, or selecting configurations for Qwen3/DSV3 families.

Quick Start

Run the MoE dispatcher selector for your hardware and EP size to obtain the recommended dispatcher and default tuning knobs.

Frequently Asked Questions about nemo-mbridge-perf-moe-dispatcher-selection

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I choose the best MoE dispatcher for my hardware?▼

MoE dispatcher selection depends on your specific hardware platform and expert-parallelism configuration. This Skill recommends the optimal dispatcher among alltoall, DeepEP, or HybridEP for target hardware like H100, GB200, and GB300 NVL72.

What is the difference between DeepEP and HybridEP for MoE workloads?▼

DeepEP and HybridEP are distinct MoE dispatcher backends for Megatron Bridge. HybridEP combines dispatch strategies, while DeepEP offers specific routing optimized for certain NVLink topologies and expert-parallelism degrees.

How do I configure flex backend settings for MoE models like DSV3?▼

Configuring flex backend settings for DSV3 requires specifying the correct dispatcher backend, such as deepep or hybridep. This Skill outputs the required flex backend settings alongside default SM tuning knobs for your hardware.

Can I use the alltoall dispatcher for Qwen3 variants on NVL72?▼

Using the alltoall dispatcher for Qwen3 variants on NVL72 depends on your specific EP degree. This Skill evaluates whether alltoall, DeepEP, or HybridEP is optimal by checking NVL72 memory constraints and expert-parallelism configurations.

What SM tuning knobs are required when bringing up MoE models?▼

Required SM tuning knobs for bringing up MoE models depend on the selected dispatcher and hardware. This Skill provides the default SM tuning values necessary for your chosen configuration to ensure optimal performance.

What happens if I am missing backend packages for DeepEP or HybridEP?▼

Missing backend packages for DeepEP or HybridEP act as environment limitations that prevent using those dispatchers. This Skill identifies these constraints, allowing you to fall back to an available dispatcher like alltoall.

Related Skills

Explore More Software Engineering Skills→
NVIDIANVIDIA
2.8k

moe-dispatcher-selection

Select optimal MoE token dispatchers based on hardware, EP degree, and workload.

Official
Advanced
NVIDIANVIDIA
2.8k

moe-optimization-workflow

Diagnose and optimize MoE training bottlenecks across memory, compute, and communication.

Official
Advanced
sayalinvidiasayalinvidia

nemo-mbridge-perf-moe-long-context

Guide long-context MoE training with CP heuristics and selective recompute for Megatron Bridge.

Community
Advanced
NVIDIA-NeMoNVIDIA-NeMo
852

perf-moe-dispatcher-selection

Select optimal MoE token dispatchers based on hardware and model size.

Official
Advanced