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

nemo-mbridge-perf-moe-hardware-configs

Identify optimal MoE training configurations by hardware platform and model family.

Updated Apr 16, 2026
Search Tags:#hardware#training#throughput#configs#dispatcher#moe#cuda-graphs
One-click install
npx skills add https://github.com/sayalinvidia/sayali-skills-test --skill nemo-mbridge-perf-moe-hardware-configs
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: nemo-mbridge-perf-moe-hardware-configs
Source: https://github.com/sayalinvidia/sayali-skills-test/tree/main/skills/nemo-mbridge-perf-moe-hardware-configs
Command: npx skills add https://github.com/sayalinvidia/sayali-skills-test --skill nemo-mbridge-perf-moe-hardware-configs

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

Helps engineers select and tune MoE training configurations tailored to specific hardware platforms and model families, summarizing throughput bands, parallelism patterns, and tuning stacks to save time and optimize performance.

Core Features & Use Cases

  • Hardware platform playbooks for MoE workloads across DSV3, Qwen3, and Qwen3-Next.
  • Representative config families with dispatcher, VPP, and CUDA graph guidance to improve throughput and efficiency.
  • Cross-cutting patterns and environment recommendations (CUDA knobs, CPU tuning) to accelerate production-ready training setups.

Quick Start

Refer to the Quick Platform Playbook to tailor MoE training configurations for your hardware and model.

Frequently Asked Questions about nemo-mbridge-perf-moe-hardware-configs

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

FAQPage Schema
What are the optimal MoE training configurations for DSV3 on H100?▼

Optimal MoE training configurations for DSV3 on H100 specify dispatcher types, VPP settings, and CUDA Graph scopes to maximize throughput. Use canonical playbooks to select environment knobs and avoid unmeasured values for production-ready setups.

How do I tune MoE training configs for Qwen3 235B on GB200?▼

Tune MoE training configs for Qwen3 235B on GB200 by applying recommended dispatcher settings, VPP, and CUDA Graph scopes. Validate target-stack environments and adjust CPU-side tuning notes rather than copying tracker rows directly.

Can I use the same MoE config patterns across H100, GB200, and GB300?▼

MoE config patterns differ across H100, GB200, and GB300 platforms. Apply hardware-specific playbooks to match dispatcher, VPP, and CUDA Graph scopes to your target stack, ensuring configurations are validated rather than copied blindly.

What CUDA Graph scopes and environment knobs improve MoE throughput?▼

Specified CUDA Graph scopes and recommended environment knobs improve MoE throughput by optimizing dispatcher execution and VPP. Apply cross-cutting patterns from canonical playbooks and include CPU-side tuning to accelerate training efficiency.

Why should I avoid copying MoE config tracker rows without validation?▼

Copying MoE config tracker rows without validation risks applying unmeasured values to incompatible target stacks. Canonical playbooks provide validated dispatcher, VPP, and CUDA Graph settings to ensure safe CPU-side tuning and hardware-specific performance.

Related Skills

Explore More Software Engineering Skills→
NVIDIANVIDIA
2.8k

moe-hardware-configs

Document MoE training configurations across H100, B200, GB200, and GB300 platforms.

Official
Advanced
yo-stevenyo-steven

perf-moe-hardware-configs

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

Community
Intermediate
sayalinvidiasayalinvidia

nemo-mbridge-perf-moe-comm-overlap

Enable expert-parallel communication overlap for Megatron Bridge MoE training.

Community
Advanced
icyzhicyzh
1

huggingface-accelerate

Simplify distributed PyTorch training across GPUs, TPUs, and multi-node clusters.

Community
Advanced