Nvidia Megatron Bridge Bump Dependency

Bump Dependency

End-to-end workflow for shipping a dependency bump in Megatron Bridge. Optimised for the case where TE, MCore, or another GPU-heavy pin moves forward — which often surfaces flakes that have to be quarantined before the PR can land.

The pipeline is always: edit → relock → push → /ok to test → watchdog → quarantine on red → re-trigger → repeat until green.

When to reach for this skill

• Bumping a git-source pin in pyproject.toml override-dependencies (e.g. transformer-engine @ git+...@<ref>).
• Bumping the 3rdparty/Megatron-LM submodule.
• Any change that touches uv.lock and needs the full L0 + L1 matrix to prove out before merge.

For pure dep additions/removals without a CI loop, the build-and-dependency skill is enough.

Required context

Read first, then follow the steps below:

• @CONTRIBUTING.md — PR title/label policy, DCO sign-off
• @skills/nvidia-megatron-bridge-build-and-dependency/SKILL.md — uv lock mechanics, container choice
• @skills/nvidia-megatron-bridge-cicd/SKILL.md — how copy-pr-bot and /ok to test work
• @skills/nvidia-megatron-bridge-testing/SKILL.md — active/ vs flaky/ directory layout, git mv quarantine recipe

Step 1 — Worktree and edit

Create a worktree off main per @CLAUDE.md. Then, before any uv lock:

git submodule update --init 3rdparty/Megatron-LM

The submodule must be initialised in the worktree or uv lock errors with "not a Python project" on the MCore path.

Edit the pin. For TE the canonical knob is the override line in pyproject.toml:

override-dependencies = [
    ...
    "transformer-engine @ git+https://github.com/NVIDIA/TransformerEngine.git@<new-ref>",
    ...
]

Use a branch name (release_v2.15) only when you want to track a moving tip; use a full SHA for reproducibility. TE branches use release_vX.Y (underscore), not release/vX.Y. Verify with git ls-remote https://github.com/NVIDIA/TransformerEngine.git.

Step 2 — Regenerate the lockfile

Run uv lock inside the project container per @skills/nvidia-megatron-bridge-build-and-dependency/SKILL.md "Regenerating uv.lock". Then confirm only the intended packages moved:

git diff --stat pyproject.toml uv.lock

If the diff carries changes you didn't ask for (transitive movements you can't explain), stop and investigate before pushing. Note that override-dependencies carries CVE floors that float — unrelated packages bumping by a patch version is expected; accept those, don't revert them.

Step 3 — Commit and push

Sign-off + signed-commit + PR title format per @CONTRIBUTING.md and @skills/nvidia-megatron-bridge-cicd/SKILL.md "Commit and PR Workflow". For a bump:

git add pyproject.toml uv.lock
git commit -S -s -m "[build] chore: bump <package> to <ref>"
git push -u origin <branch-name>

A signed commit (-S) lets copy-pr-bot trigger CI without manual /ok to test for the first push — but you'll still post /ok to test on every subsequent SHA in this loop (Step 5).

Step 4 — Open the PR

Title and labels per @CONTRIBUTING.md. Two bump-specific requirements:

• Apply needs-more-tests — mandatory for a bump; expands the matrix from L0 to L0+L1.
• For a high-blast-radius bump (TE, MCore submodule, anything that touches CUDA kernels), also apply full-test-suite to pull L2 into the PR run. L2 covers VL models, checkpoint conversion, and heavy quantization which otherwise only run on schedule.

The PR body template — this is the durable record of the bump:

<details><summary>Claude summary</summary>

## What
- Bump <package> to <ref>.
- Regenerate `uv.lock`.

## Lockfile delta

Updated ->

## Test plan
- [ ] L0 CI green
- [ ] L1 CI green (label `needs-more-tests` applied)

## Quarantined tests (this bump)
_None yet — will be appended as flakes are identified during CI iteration._

</details>

To update the PR title or body later, use gh api -X PATCH "repos/NVIDIA-NeMo/Megatron-Bridge/pulls/<N>" -F "body=@/tmp/pr-body.md" — never gh pr edit.

Step 5 — Trigger CI on the exact SHA

Trigger mechanics live in @skills/nvidia-megatron-bridge-cicd/SKILL.md "How CI Is Triggered". For this loop the rule is simple: on every new SHA you push, post /ok to test $(git rev-parse HEAD) as a PR comment, even if your commits are signed. This guarantees the run targets the SHA you actually want exercised and re-fires anything that got cancelled or cached.

Step 6 — Attach the watchdog (always; never a cronjob)

For a bump PR you want a single live process that emits per-job state changes for the CICD NeMo workflow only. Other workflows (docs, wheel, copyright, install-test) are noise here — the gate that decides green-or-red for a bump is CICD NeMo.

Always attach a watchdog with the Monitor tool. Never schedule wakeups or cronjobs for this loop. A watchdog gives you:

• Sub-minute reaction time on every job transition.
• A single live process — no scattered scheduled-wakeup state to reason about.
• Natural early termination via TaskStop once the run is green.

Watchdog script

Save to /tmp/watchdog-<PR>.sh and chmod +x:

#!/usr/bin/env bash
# Watchdog: monitor "CICD NeMo" runs on pull-request/<PR> and emit
# per-job state changes. Stays alive across re-runs (new commits).
set -u
PR=<PR>
REPO=NVIDIA-NeMo/Megatron-Bridge
BRANCH="pull-request/$PR"

prev_run_id=""
declare -A prev_state

emit() { echo "[$(date -u +%H:%M:%SZ)] $*"; }

while true; do
  run_json=$(gh run list --repo "$REPO" --workflow "CICD NeMo" \
    --branch "$BRANCH" --limit 1 \
    --json databaseId,status,conclusion,headSha 2>/dev/null || echo "[]")
  run_id=$(echo "$run_json" | jq -r '.[0].databaseId // empty')
  run_status=$(echo "$run_json" | jq -r '.[0].status // empty')
  run_conclusion=$(echo "$run_json" | jq -r '.[0].conclusion // empty')
  run_sha=$(echo "$run_json" | jq -r '.[0].headSha // empty')

  if [[ -z "$run_id" ]]; then
    sleep 30; continue
  fi

  if [[ "$run_id" != "$prev_run_id" ]]; then
    emit "RUN ${run_id} STARTED sha=${run_sha:0:8} status=${run_status}"
    prev_run_id="$run_id"
    unset prev_state
    declare -A prev_state
  fi

  jobs_json=$(gh run view "$run_id" --repo "$REPO" --json jobs 2>/dev/null || echo "{}")
  while IFS=$'\t' read -r name status conclusion; do
    [[ -z "$name" ]] && continue
    cur="${status}/${conclusion}"
    if [[ "${prev_state[$name]:-}" != "$cur" ]]; then
      case "$status" in
        completed)
          emit "JOB ${name} -> ${conclusion}" ;;
        in_progress)
          if [[ -z "${prev_state[$name]:-}" || "${prev_state[$name]}" == "queued/" ]]; then
            emit "JOB ${name} -> in_progress"
          fi ;;
      esac
      prev_state[$name]="$cur"
    fi
  done < <(echo "$jobs_json" | jq -r '.jobs[]? | [.name, .status, (.conclusion // "")] | @tsv')

  if [[ "$run_status" == "completed" ]]; then
    emit "RUN ${run_id} COMPLETED conclusion=${run_conclusion}"
  fi

  sleep 60
done

Arming the watchdog

Monitor(
  description="CICD NeMo run state changes on PR <N>",
  command="bash /tmp/watchdog-<N>.sh",
  persistent=true,
  timeout_ms=3600000
)

persistent: true keeps it alive across re-runs (you'll push more commits when quarantining flakes). Stop it with TaskStop(<task-id>) once the run is green.

Why never a cronjob / scheduled wakeup

• Cronjobs run blind — they fire on a clock, not on an event. You'll either over-poll (cache miss every wake-up) or miss long stalls.
• Wakeups can't easily fan out to "tell me whenever a job transitions" — they only resume the agent on a fixed interval.
• A persistent Monitor surfaces every job edge in real time and exits cleanly when the work is done.

Step 7 — Quarantine on red, then iterate

When a JOB <name> -> failure event fires:

1. Triage the failure — is it the bump or a flake? Skim the logs:

   RUN_ID=<from "RUN ... STARTED" event>
   gh run view "$RUN_ID" --repo NVIDIA-NeMo/Megatron-Bridge --log-failed > /tmp/run.log
   wc -l /tmp/run.log
   tail -200 /tmp/run.log
   

   This is the bump-specific judgement call: only quarantine if the failure reproduces on main or is clearly unrelated infrastructure. If the failure is caused by the bump (real regression), stop quarantining — fix the underlying issue or revert the bump. Quarantining a real regression hides the very signal the bump PR exists to surface.

2. Move the launch script to flaky/ per @skills/nvidia-megatron-bridge-testing/SKILL.md "Moving a Test to Flaky". Map a CI job name to its launch script via:

   • prefix gb200_ → gb200/active/, otherwise h100/active/
   • the rest is the script's basename without .sh

3. Append to the PR body's Quarantined tests section with a one-line reason and a follow-up tracking link if you have one. This is the durable record of what this bump deferred — the section exists precisely so a reviewer can see at a glance which flakes were side-stepped to land the bump.

4. Commit, push, retrigger:

   git commit -S -s -m "[ci] chore: quarantine flaky <test> for <package> bump"
   git push
   gh pr comment <N> --repo NVIDIA-NeMo/Megatron-Bridge \
     --body "/ok to test $(git rev-parse HEAD)"
   

5. Update the PR body via gh api PATCH so the quarantine list stays current.

The watchdog is persistent — it picks up the new run automatically and emits RUN <id> STARTED for the new attempt. Loop back to step 1.

Step 8 — Stop when green

RUN <id> COMPLETED conclusion=success is the exit condition. Then:

gh pr checks <N> --repo NVIDIA-NeMo/Megatron-Bridge | awk '{print $2}' | sort | uniq -c
TaskStop(<watchdog-task-id>)
gh api -X PATCH "repos/NVIDIA-NeMo/Megatron-Bridge/pulls/<N>" -F "body=@/tmp/pr-body.md"

Common pitfalls

Anti-patterns

• Cron / scheduled wakeups for this loop. Always Monitor.
• Polling all workflows. Filter to CICD NeMo — the rest are noise for a bump.
• Quarantining a real regression to "make CI green." That defeats the purpose of the bump PR. Only quarantine if the failure reproduces on main or is clearly unrelated infrastructure.
• gh pr edit for title/body. Use gh api PATCH.
• HEREDOC in gh pr create --body. Always go through a tmpfile + --body-file.
• Bundling unrelated changes (feature work, refactors) into a bump PR. Bumps should stay surgical so CI failures attribute cleanly.