skills.rest
NVIDIA
NVIDIAOFFICIAL
Explore all skills by this author →
WebsiteGitHub

moe-dispatcher-selection

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

2.8k|332|Updated Feb 25, 2026
Search Tags:#performance-tuning#dispatcher#qwen3#moe#nvlink#ep#dsv3
One-click install
npx skills add https://github.com/NVIDIA/skills --skill moe-dispatcher-selection
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: moe-dispatcher-selection
Source: https://github.com/NVIDIA/skills/tree/main/skills/Megatron-Bridge/perf-techniques/moe-dispatcher-selection
Command: npx skills add https://github.com/NVIDIA/skills --skill moe-dispatcher-selection

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

Efficiently selecting the optimal MoE token dispatcher (AlltoAll, DeepEP, or HybridEP) for a given hardware platform, EP degree, and workload, to maximize throughput and memory efficiency.

Core Features & Use Cases

  • Guidance by hardware and EP degree to pick among AlltoAll, DeepEP, and HybridEP based on workload patterns from DSV3, Qwen3, Qwen3-Next, and VLM bring-up work.
  • Includes tuning recommendations and NVLink-domain considerations for H100, GB200, and GB300 systems to plan experiments and deployments.
  • Use Case: data scientists and engineers tuning MoE workloads for large-scale language models and vision models across multi-GPU setups.

Quick Start

Provide a single instruction to evaluate hardware, EP degree, and workload characteristics to receive a recommended MoE dispatcher and baseline tuning options.

Frequently Asked Questions about 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?▼

Choosing the best MoE dispatcher requires evaluating your hardware platform, EP degree, and workload characteristics to select among AlltoAll, DeepEP, and HybridEP. This ensures maximum throughput and memory efficiency for your deployment.

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

AlltoAll and DeepEP differ in their hardware-aware token dispatching mechanisms for cross-node and intra-node communication. Selecting between them depends on your specific hardware platform, EP degree, and workload patterns to maximize efficiency.

Does the MoE dispatcher selection support Qwen3 and DSV3 on H100 systems?▼

Yes, the MoE dispatcher selection supports Qwen3, Qwen3-Next, and DSV3 workloads on H100 systems. It provides tailored tuning guidance and NVLink-domain considerations across H100, GB200, and GB300 platforms.

How do I tune MoE token dispatching for multi-GPU setups?▼

To tune MoE token dispatching for multi-GPU setups, evaluate your EP degree and NVLink-domain considerations to receive baseline tuning options. This guides cross-node and intra-node MoE workload configurations for large-scale language models.

When should I use HybridEP instead of AlltoAll for large-scale language models?▼

Use HybridEP instead of AlltoAll when your specific hardware platform, EP degree, and workload patterns require specialized cross-node and intra-node token routing. Evaluating these factors ensures you select the optimal dispatcher for maximum throughput.

Related Skills

Explore More Software Engineering Skills→
yo-stevenyo-steven

perf-expert-parallel-overlap

Overlap expert-parallel token dispatch and combine communication with expert feed-forward compute in Megatron-Bridge MoE models.

Community
Advanced
sayalinvidiasayalinvidia

nemo-mbridge-perf-parallelism-strategies

Select and tune Megatron Bridge parallelism configurations for model and hardware.

Community
Advanced
yo-stevenyo-steven

perf-moe-hardware-configs

Select MoE training parallelism and tuning configurations for H100, B200, GB200, and GB300.

Community
Intermediate
NVIDIA-NeMoNVIDIA-NeMo
852

perf-moe-dispatcher-selection

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

Official
Advanced