---

name: nvidia-megatron-bridge-cicd description: CI/CD reference for Megatron Bridge — pipeline structure, commit and PR workflow, CI failure investigation, and common failure patterns. license: Apache-2.0 AND CC-BY-4.0 metadata: when_to_use: Investigating a CI failure, understanding the pipeline structure, writing a commit or PR, triggering CI, 'CI is red', 'how do I trigger CI', 'PR workflow', 'where are the logs', 'CI did not run', 'copy-pr-bot', '/ok to test'. author: NVIDIA

CI/CD

Commit and PR Workflow

• Never commit directly to main — always create a feature branch.
• Always sign commits: git commit -s -m "message".
• PR title format: [{areas}] {type}: {description} (e.g., [model] feat: Add Qwen3 model bridge). See @CONTRIBUTING.md for the full PR workflow, area/type labels, and DCO requirements.

How CI Is Triggered

The workflow is defined in @.github/workflows/cicd-main.yml and is triggered on push — not on pull_request. This is intentional: a bot called copy-pr-bot controls when CI runs.

Mechanism:

1. When a PR is opened, copy-pr-bot watches for a trust signal.
2. Trust is established in one of two ways:
   • All commits on the PR branch are GPG-signed by a verified NVIDIA contributor → bot triggers automatically.
   • An NVIDIAN posts /ok to test <commit-sha> as a PR comment → bot triggers manually for that SHA.
3. Once trusted, copy-pr-bot copies the PR's code into the remote branch pull-request/<number> and pushes it.
4. That push fires the workflow's push trigger on refs/heads/pull-request/<number>, launching CI.

Consequences:

• CI never runs on untrusted pushes — external contributors always need /ok to test.
• The running workflow branch is pull-request/<number>, not the author's feature branch.
• Pushing a new commit to a PR does not automatically re-trigger CI unless the commit is signed or /ok to test <new-sha> is posted.
• Concurrent runs for the same PR are cancelled automatically (concurrency group per PR number).

Pipeline Structure

pre-flight
  └── lint-check
        └── cicd-wait-in-queue       # queues workflows to avoid runner interleaving across PRs
              └── cicd-container-build
                    ├── unit-tests-core
                    ├── unit-tests-diffusion
                    └── functional-tests (L0 always; L1 with needs-more-tests label; L2 on schedule or full-test-suite label)

• Slack notifications are sent on completion for scheduled and nightly runs.

For functional test tier semantics and job-to-directory mapping, see the testing skill.

CI Failure Investigation

Locating the PR from a CI Branch

# Extract PR number from branch name (e.g. pull-request/1234)
PR_NUMBER=$(git rev-parse --abbrev-ref HEAD | grep -oP '(?<=pull-request/)\d+')

gh pr view "$PR_NUMBER" --repo NVIDIA-NeMo/Megatron-Bridge
gh pr diff "$PR_NUMBER" --repo NVIDIA-NeMo/Megatron-Bridge --name-only
gh pr checks "$PR_NUMBER" --repo NVIDIA-NeMo/Megatron-Bridge

Investigating a Failing Job

1. Get the PR number from the branch name (see above).
2. Review the changeset:

   gh pr diff "$PR_NUMBER" --repo NVIDIA-NeMo/Megatron-Bridge
   

3. Identify the failing job from gh pr checks output.
4. Fetch job logs:

   gh run list --repo NVIDIA-NeMo/Megatron-Bridge --branch "pull-request/$PR_NUMBER"
   gh run view <run_id> --repo NVIDIA-NeMo/Megatron-Bridge --log-failed > run.log
   

5. Scan logs in chunks — log files can exceed 10,000 lines, never load them whole:

   wc -l run.log
   tail -200 run.log          # start from the end
   sed -n '1,200p' run.log    # or scan forward in 200-line chunks
   

6. Cross-reference the changeset against the failing step.

Hugging Face Model Access In CI

Assume CI functional-test containers run with Hugging Face models offline (HF_HUB_OFFLINE=1) and a pre-populated HF_HOME. When reproducing or fixing CI failures involving HF models, mirror this locally by setting HF_HUB_OFFLINE=1 after warming the cache. Test fixtures must not depend on live Hub API calls such as list_repo_files() or uncached downloads during CI. For trust_remote_code=True toy checkpoints, copy custom Python modules from the already loaded local/cache source files or a local snapshot, not by listing the remote repo at test time.

Common Failure Patterns

Symptom	Likely Cause	Action
CI never started on a PR	Commits not GPG-signed and no /ok to test comment	Post /ok to test <full-sha> on the PR
Lint job fails	ruff or pre-commit violation	Run ruff check --fix + ruff format locally
Container build fails	Dependency conflict or stale uv.lock	Re-run uv lock inside Docker and commit updated lock
Unit tests fail	Code regression or missing import	Run failing test locally; check the PR diff
Functional test (L0) fails	Integration breakage	Check GPU runner logs; reproduce with L0_Launch_*.sh
HF model fixture passes locally but fails in CI with OfflineModeIsEnabled	Test made a live Hugging Face Hub API/download call; CI has HF_HUB_OFFLINE=1	Warm local cache, reproduce with HF_HUB_OFFLINE=1, and change the fixture to use cached/local artifacts only
cicd-wait-in-queue running long	Many PRs queued; automation serializes runners to avoid interleaving	Wait; or check queue depth in the Actions tab
MCore submodule mismatch	Pinned commit out of sync	Update 3rdparty/Megatron-LM submodule and re-lock
Stale checkpoint auto-resume	nemo_experiments/ from a previous run exists	rm -rf nemo_experiments before starting fresh
Port collision on Slurm (EADDRINUSE)	ntasks-per-node=8 with torchrun	Drop torchrun; use ntasks-per-node=8 with uv run python script.py