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

onboard-gb200-1node-tests

Converts GB200 2-node GitHub MR functional tests into 1-node variants.

Updated May 23, 2026
Search Tags:#megatron-core#functional tests#gb200#github-mr#parallelism configuration#ci onboarding#test recipes
One-click install
npx skills add https://github.com/yo-steven/skills-exploration-20260522 --skill onboard-gb200-1node-tests
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: onboard-gb200-1node-tests
Source: https://github.com/yo-steven/skills-exploration-20260522/tree/main/skills/Megatron-Core/onboard-gb200-1node-tests
Command: npx skills add https://github.com/yo-steven/skills-exploration-20260522 --skill onboard-gb200-1node-tests

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

This Skill solves the problem of expanding GB200 GitHub MR test coverage by converting existing 2-node functional tests into correct 1-node (mr-github) variants without misconfiguring parallelism for a 4-GPU world.

Core Features & Use Cases

  • Discovers candidate tests by scanning gpt.yaml and moe.yaml for 2-node (mr-scoped) entries that need 1-node counterparts, while skipping scopes like nightly, weekly, and mr-broken.
  • Generates or updates _1node test cases by copying existing model_config.yaml directories and applying the right PP/EP reductions to fit 8 GPUs → 4 GPUs.
  • Produces working recipe and products entries by creating gpt-1node.yaml if missing and appending to moe-1node.yaml, then assigning appropriate mr-github/mr-github-slim scopes.

Quick Start

Use this Skill to onboard GB200 by generating _1node variants of all applicable mr-scoped GPT and MoE functional tests and wiring them into GitHub MR CI with the correct mr-github product scopes.

Frequently Asked Questions about onboard-gb200-1node-tests

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

FAQPage Schema
How do I convert GB200 2-node functional tests into 1-node variants?▼

You convert 2-node functional tests by copying existing model_config.yaml directories and halving data parallelism via PP/EP/world-size classification to correctly fit the 4-GPU world configuration.

What does halving DP via PP/EP/world-size classification mean for GB200 CI onboarding?▼

Halving DP via PP/EP/world-size classification adjusts parallelism configurations from an 8-GPU world (4 GPUs per node × 2 nodes) to a 4-GPU world, ensuring deterministic mr-github execution for CI onboarding.

How do I generate _1node model configs for megatron-core functional tests?▼

You generate _1node model configs by creating or updating gpt-1node.yaml and moe-1node.yaml files, then appending the corresponding entries while assigning appropriate mr-github or mr-github-slim scopes.

Does the GB200 1-node test conversion skip specific recipe scopes?▼

Yes, the conversion process skips nightly, weekly, and mr-broken scopes, targeting only the 2-node mr-scoped GPT and MoE entries that require new 1-node counterparts for coverage.

What are the limitations when migrating GitHub MR CI test recipes to a 4-GPU world?▼

Limitations include strictly skipping already-covered or non-target scopes and avoiding misconfiguring parallelism during the 8-GPU to 4-GPU reduction to maintain valid CI coverage.

Related Skills

Explore More Software Engineering Skills→
NVIDIANVIDIA
17.4k

onboard-gb200-1node-tests

Generate single-node GB200 test cases from 2-node MR recipes.

Official
Intermediate
kentojekentoje
2

gitlab-create-merge-request

Create a GitLab merge request using the first commit message as the title.

Community
Advanced
KappTech88KappTech88

training-llms-megatron

Optimize Megatron-Core LLM training workflows across multi-GPU clusters.

Community
Advanced
sayalinvidiasayalinvidia

mcore-onboard-gb200-1node-tests

Generate 1-node GB200 model_config.yaml variants and update MR test recipes.

Community
Intermediate