skills.rest
yo-steven
yo-stevenCOMMUNITY
Explore all skills by this author →
GitHub

perf-moe-hardware-configs

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

Updated May 23, 2026
Search Tags:#training optimization#parallelism#fp8#moe#gpu hardware#cuda graphs#throughput planning
One-click install
npx skills add https://github.com/yo-steven/skills-exploration-20260522 --skill perf-moe-hardware-configs-yo-steven
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: perf-moe-hardware-configs
Source: https://github.com/yo-steven/skills-exploration-20260522/tree/main/skills/Megatron-Bridge/perf-moe-hardware-configs
Command: npx skills add https://github.com/yo-steven/skills-exploration-20260522 --skill perf-moe-hardware-configs-yo-steven

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

It solves the problem of picking effective MoE training parallelism and tuning choices for a specific hardware platform, without relying on guesswork or fragile “one-size-fits-all” recipes.

Core Features & Use Cases

  • Hardware-to-config mapping: Provides representative MoE training playbooks by platform (H100, B200, GB200, GB300) and model family (DSV3, Qwen3, Qwen3-Next) to guide planning and run setup.
  • Throughput planning ranges: Summarizes rounded throughput bands (TFLOPS/GPU and MFU) to set realistic expectations before benchmarking.
  • Parallelism + tuning stacks: Recommends dispatcher types (DeepEP/HybridEP), PP/TP/EP/VPP shapes, routing defaults, recompute starting points, and common environment variables, plus CPU-side tuning notes.
  • Use case: If you’re preparing a new MoE training run for DSV3 on GB200, you can start from the GB200 HybridEP playbook, select a matching PP shape, enable CUDA-graph-friendly scopes, and choose a recompute level aligned to memory pressure.

Quick Start

Use the perf-moe-hardware-configs skill to generate a starting parallelism and tuning plan for “DSV3 on GB200” from the representative config families and cross-cutting patterns.

Frequently Asked Questions about perf-moe-hardware-configs

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

FAQPage Schema
How do I configure MoE training parallelism for H100 or B200 GPUs?▼

To configure MoE training parallelism for H100 or B200, select dispatcher types like DeepEP or HybridEP, define matching PP/TP/EP/VPP shapes, and apply routing defaults to optimize throughput.

What is the best way to plan throughput for DSV3 or Qwen3 MoE training?▼

Planning throughput for DSV3 or Qwen3 MoE training involves using rounded throughput bands like TFLOPS/GPU and MFU to set realistic performance expectations before benchmarking hardware runs.

Can I use CUDA graphs with MoE training on GB200 or GB300 platforms?▼

Yes, you can use CUDA graphs with MoE training on GB200 or GB300 by enabling CUDA-graph-friendly scopes within the platform playbook to optimize execution graphs.

When do I need to adjust recompute strategies for MoE training?▼

You need to adjust recompute strategies for MoE training when facing memory pressure, selecting a recompute level aligned with your current memory constraints to balance throughput and capacity.

Does MoE training tuning require specific environment variables or CPU-side setup?▼

MoE training tuning requires setting common environment variables and applying CPU-side tuning notes to ensure dispatcher and parallelism configurations perform optimally across target hardware.

Related Skills

Explore More Education & Research Skills→
NVIDIA-NeMoNVIDIA-NeMo
852

perf-moe-hardware-configs

Provide hardware-specific MoE training configurations and performance benchmarks.

Official
Advanced
EvolvingLMMs-LabEvolvingLMMs-Lab
1.2k

megatron-checkpoint-layout

Diagnose and convert Megatron checkpoint shard layouts across TP, PP, and EP dimensions.

Official
Advanced
sayalinvidiasayalinvidia

nemo-mbridge-perf-moe-vlm-training

Compare FSDP-first and 3D-parallel paths for MoE VLM training in Megatron Bridge.

Community
Advanced
blueskies1818blueskies1818
1

serving-llms-vllm

Deploy high-throughput LLM serving with OpenAI-compatible endpoints using vllm.

Community
Advanced