---

name: testing-ci-and-maintenance description: "Megatron-LM repository testing, CI, golden-value, formatting, PR, and maintenance workflows." metadata: disco-role: operating disable-model-invocation: true license: NOASSERTION

testing-ci-and-maintenance

Use this sub-skill when the task is about Megatron-LM repository development rather than package-user training:

• adding, selecting, or running unit tests and functional tests;
• editing CI recipe YAML, H100 or GB200 scopes, golden values, or CI labels;
• triaging CI logs, artifacts, pytest failures, golden drift, lint failures, or container/dependency CI failures;
• updating base image pins, dependency lock files, API compatibility checks, or nightly main-to-dev sync PRs;
• preparing contribution/PR steps, splitting a PR by CODEOWNERS ownership, or responding to repo issues.

Route away from this sub-skill for:

• package installation or environment setup for Megatron-LM users: use install-and-environment;
• research training command construction, data preprocessing, SLURM launch templates, or checkpoint training flows: use training-cli-and-data and, when relevant, checkpointing-and-conversion;
• model API or parallelism strategy questions: use core-models-and-parallelism.

Mandatory read order

1. For test layout, pytest commands, recipe YAML fields, and H100/GB200 scope conventions, read references/testing-reference.md.
2. For CI labels, logs/artifacts, golden-value refresh, internal CI trigger safety, base-image/dependency maintenance, and nightly sync, read references/ci-maintenance-reference.md.
3. For PR creation, signing, draft PRs, forks, CODEOWNERS/final review, issues, and PR splitting, read references/contribution-reference.md.
4. For failure diagnosis, read references/troubleshooting.md before proposing fixes.

Bundled scripts

• scripts/check_golden_values.py is a copied/adapted finite-value checker for golden JSON files. Prefer this bundled checker when validating an arbitrary golden JSON fixture outside the repo CI path.
• scripts/summarize_recipe_scopes.py is a non-mutating YAML summarizer for recipe scope/platform/environment coverage and broken/disabled markers.

Operating rules

• Use Run tests for lightweight full-scope PR validation and Run functional tests when numerics, new functional cases, re-enabled cases, or golden values are involved.
• Treat container::lts as opt-in only. Do not add the label, edit LTS pins, or run LTS-specific maintenance unless the user explicitly asks for LTS validation or an LTS bump.
• Unit tests are distributed by default; do not run plain pytest tests/unit_tests as a CI-parity command when GPU/distributed context matters.
• Do not hand-edit generated lockfiles or golden JSONs. Regenerate uv.lock with the package manager inside the project container; refresh goldens from CI artifacts and summarize relative differences.
• Never run the internal GitLab CI trigger without a dry run first; the real trigger force-pushes the current branch to an internal pull-request/<branch> ref.
• Draft PRs are mandatory. Contributors push branches to forks and open draft PRs against the upstream repository; do not push contributor branches directly to upstream.