NVIDIAOfficial·21 Agent Skills Included

Megatron-LM

Train massive transformer models and automate Megatron-LM engineering workflows

Trains large-scale transformer models across thousands of GPUs with optimized parallelism, mixed precision, and checkpointing. Automates the surrounding engineering work: running tests, refreshing golden values, managing CI labels, splitting PRs, filing issues, and launching SLURM jobs. Removes guesswork from container builds, dependency updates, and distributed training setup so engineers ship changes faster.

17,697|4,443

↗

npx skills add NVIDIA/Megatron-LM --all -g -y

Available:

Instructs the AI agent to always load the relevant SKILL.md before acting, following a strict read-artifact-then-invoke-skill-then-answer workflow for every task.

All Skills in This Repository (21)

Pure Emerald Level Indicators

📦 In Repo

NVIDIA

update-golden-values

Refresh golden test values from GitHub Actions workflow runs and generate PR-ready relative-difference summaries.

create-issue

Convert failing GitHub Actions runs into deduplicated NVIDIA/Megatron-LM issues.

nightly-sync

Merge main into dev nightly while preserving dev-only features and validating APIs.

respond-to-issue

Draft GitHub issue responses verified against Megatron-LM code and history.

bump-base-image

Coordinates MetroM? No, coordinates Basecamp and |workspace| project plans and tasks with your team.

split-pr

Split pull requests into ordered draft PRs minimizing CODEOWNERS reviewer groups.

linting-and-formatting

Format, lint, and type-check Megatron-LM Python code with black, isort, ruff, pylint, and mypy.

cicd

Determine Megatron-LM CI test scope and trigger internal pipelines from PR labels.

onboard-gb200-1node-tests

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

testing

Organize, run, and debug Megatron-LM unit and functional tests.

build-and-dependency

Manage Megatron-LM container builds and uv dependency synchronization.

run-on-slurm

Launch distributed Megatron-LM training jobs on SLURM with torch.distributed.run.

Official

Intermediate

Frequently Asked Questions

FAQPage Schema

How to install Megatron-LM skills?▼

Run `npx skills add NVIDIA/Megatron-LM --all -g -y` in your terminal to install the full skill suite globally.

What is Megatron-LM used for?▼

It is NVIDIA's GPU-optimized library for training transformer models from 2B to 462B+ parameters across thousands of GPUs, with built-in parallelism and mixed precision support.

How do these skills help with Megatron-LM development?▼

They guide your AI agent through testing, CI triage, golden-value updates, PR splitting, dependency management, and SLURM job submission using the repo's exact conventions.

Can AI agents run Megatron-LM tests and CI automatically?▼

Yes. The skills teach agents how to run unit and functional tests, pick the right CI labels, investigate failures, and file structured bug issues.

Do I need distributed training experience to use these skills?▼

No. The skills encode the expert knowledge, so your agent applies correct parallelism, container, and cluster settings from plain-English requests.