release-cherry-pick-missing-reverts

作成者: pytorch

Find reverts that landed on pytorch/pytorch main but are missing from a release branch (release/X.Y) because the reverted commit was already shipped in a…

npx skills add https://github.com/pytorch/test-infra --skill release-cherry-pick-missing-reverts

Release: Cherry-pick missing reverts

When a commit ships in a release candidate (so it is on release/X.Y) and is later reverted on main, the revert does not automatically reach the release branch — the buggy commit is still present in release/X.Y. These are missing reverts: the revert must be cherry-picked onto the release branch.

This skill finds those missing reverts and opens one cherry-pick PR per revert against pytorch/pytorch:release/X.Y, each on its own branch in the user's fork. It never pushes to release/X.Y directly.

The detector is tools/analytics/github_analyze.py (run daily by the GitHub Analytics Daily workflow). It is the same script this skill lives next to in test-infra, but the cherry-picks operate on a pytorch/pytorch checkout.

Inputs

InputRequiredExampleNotes
Release branchyesrelease/2.13The branch to cherry-pick reverts onto.
Sourceone ofa GHA run URL, or "run the analyzer"Either a GitHub Analytics Daily run URL/ID whose log already has the analysis, or run the analyzer locally for fresh data.
pytorch/pytorch pathyes (for cherry-pick)~/pytorchA checkout with upstreampytorch/pytorch and origin → the user's fork.
Fork remotedefaults to originoriginWhere the cherry-pick branches are pushed.
Tracker issueoptional186934The [vX.Y.Z] Release Tracker issue. When given, post one cherry-pick nomination comment per opened PR (see Step 4).

If the release branch was not supplied, ask for it before doing anything — do not guess. If a tracker issue is given, the matching release/X.Y should agree with the tracker's version (e.g. issue [v2.13.0] Release Trackerrelease/2.13); confirm before posting comments.

When to use this skill

Use when the user asks to:

  • Find / list missing reverts for a release branch
  • Cherry-pick the reverts flagged by the analytics run to release/X.Y
  • Act on a GitHub Analytics Daily run that printed 🔴 WARNING: This is possibly a revert of a commit that was included in a release candidate

Background: what "missing revert" means and how it is flagged

analyze_reverts_missing_from_branch compares main against the release branch and, for every revert that is on main but not on the release branch, checks whether the reverted commit carries a release-candidate tag (v[0-9]+.[0-9]+.[0-9]+-rc[0-9]+). Three outcomes per revert (the analyzer prints the status lines with two spaces after the emoji, e.g. 🔴 WARNING; match on the emoji, not the exact spacing):

  • 🔴 WARNING ... — the reverted commit carries an RC tag, so it may be in release/X.Y. Treat as a missing-revert candidate, but verify before acting (see caveat below).
  • ✅ DETECTED: The reverted commit ... was cherry-picked to <branch> — the revert is already on the release branch. Skip.
  • 🟢 STATUS: ... may not be needed — the reverted commit was never in the release branch. Skip.

Caveat — the WARNING is a heuristic, not a guarantee. The tag regex matches an RC tag from any release line, not just the target. For release/2.13 a commit that only ever shipped in a v2.12.0-rcN (and was never on release/2.13) is still flagged 🔴. So a 🔴 WARNING does not prove the reverted commit is on the target branch — Step 3 must confirm with git merge-base --is-ancestor before cherry-picking, or it may try to revert code that isn't there (empty/conflicting cherry-pick).

Each flagged entry prints, in order:

Reverted GitHub Commit: <reverted_sha>          # the bad commit still in release/X.Y
🏷️  Tags matching ... : v2.13.0-rc1 ...          # proof it shipped in an RC
Commit Hash: <revert_sha>                        # the revert commit on main -> cherry-pick THIS
Author / Date / Title: Revert "<orig title> (#<PR>)"
🔴  WARNING: ...

The value to cherry-pick is Commit Hash (the revert commit on main), not the reverted commit.

Step 1 — Get the list of missing reverts

Option A — from a referenced run (when the user points at a GitHub Analytics Daily run):

# Find the github-analyze job id for the run, then fetch its log.
gh api repos/pytorch/test-infra/actions/runs/<RUN_ID>/jobs \
  -q '.jobs[] | select(.name=="github-analyze") | .id'
gh api repos/pytorch/test-infra/actions/jobs/<JOB_ID>/logs > /tmp/ghanalyze.log

Option B — run the analyzer locally (preferred for fresh data; the CI log can be stale). From a pytorch/pytorch checkout with upstreampytorch/pytorch:

git -C <pytorch> fetch upstream main release/X.Y --tags
python test-infra/tools/analytics/github_analyze.py \
  --repo-path <pytorch> --remote upstream \
  --branch release/X.Y --analyze-missing-reverts-from-branch | tee /tmp/ghanalyze.log

Step 2 — Parse the flagged reverts

Extract every entry whose status line contains the 🔴 (WARNING) emoji — match on the emoji rather than exact text/spacing, since the analyzer emits two spaces (🔴 WARNING). For each, capture:

  • revert_sha ← the Commit Hash: line (cherry-pick target)
  • reverted_sha ← the Reverted GitHub Commit: line
  • pr ← the #NNNN in the Title: line (the original PR that was reverted)
  • title ← the Title: text

A revert with ✅ DETECTED or 🟢 STATUS is not missing — skip it. Report the counts (flagged vs skipped) so nothing is silently dropped.

Reverts of Phabricator diffs print Reverted Phabricator Diff: instead of Reverted GitHub Commit:; the analyzer never resolves a GitHub SHA for them, so they can't get a 🔴 WARNING and won't appear here. They are out of scope for this skill (no GitHub commit to cherry-pick).

Step 3 — One cherry-pick PR per missing revert

Operate on the pytorch/pytorch checkout. For each flagged revert:

REL=release/X.Y
git -C <pytorch> fetch upstream "$REL" main --tags

# Confirm the reverted commit is actually on the target branch before reverting
# it (the 🔴 WARNING can fire on an RC tag from a different release line). If it
# is not an ancestor, skip and report "reverted commit not in <REL>".
git -C <pytorch> merge-base --is-ancestor <reverted_sha> "upstream/$REL" \
  || { echo "skip: <reverted_sha> not in $REL"; continue; }

BR="cherry-pick-revert-<PR>-${REL#release/}"     # e.g. cherry-pick-revert-185760-2.13
git -C <pytorch> checkout -B "$BR" "upstream/$REL"
git -C <pytorch> cherry-pick -x <revert_sha>
  • -x records (cherry picked from commit <revert_sha>) in the message.
  • On conflict: do not force-resolve blindly. Report the conflicting files for that revert, run git cherry-pick --abort, and leave it out of the PR batch (note it in the summary as "needs manual cherry-pick"). Continue with the others.

Push to the fork and open the PR against the release branch:

git -C <pytorch> push origin "$BR"
gh pr create --repo pytorch/pytorch --base "$REL" --head "<fork-owner>:$BR" \
  --title "[$REL] Revert \"<orig title> (#<PR>)\"" \
  --body "<see PR body below>"

Never push to release/X.Y itself, and never open the PR with --base main.

PR body

Cherry-pick of the main-branch revert <revert_sha> onto release/X.Y.

The reverted commit <reverted_sha> (#<PR>) shipped in a release candidate
(v X.Y.0-rcN) and is present in release/X.Y, but the revert only landed on
main. This cherry-picks the revert so release/X.Y matches main.

Detected by tools/analytics/github_analyze.py --analyze-missing-reverts-from-branch
(GitHub Analytics Daily). Cherry-pick (-x): (cherry picked from commit <revert_sha>).

This PR was authored with the assistance of an AI coding agent.

Step 4 — Nominate on the release tracker (if a tracker issue was given)

For each cherry-pick PR opened in Step 3, post one comment on the tracker issue in the tracker's standard nomination format. The landed trunk PR is the original PR that was reverted (the #<PR> from the revert title), and the release branch PR is the cherry-pick PR:

gh issue comment <tracker_issue> --repo pytorch/pytorch --body "Link to landed trunk PR (if applicable):
* https://github.com/pytorch/pytorch/pull/<PR>

Link to release branch PR:
* https://github.com/pytorch/pytorch/pull/<cherry_pick_PR>

Criteria Category:
* cherry-pick revert"

One comment per revert. Only comment for PRs actually opened in Step 3 — skip the ones that were skipped or hit a conflict. Like opening PRs, posting to the tracker is outward-facing: confirm first.

Step 5 — Summary

Report a table: PR, original title, revert_sha, and outcome — PR #<n> opened, skipped (already on <branch>), skipped (reverted commit not in <branch>), or conflict — needs manual cherry-pick. Include the new branch names, PR URLs, and (if a tracker issue was given) the tracker comment links.

Guardrails

  • Confirm before opening PRs or commenting. Creating multiple cherry-pick PRs and posting tracker comments are outward-facing actions — list what will be opened/posted and confirm first.
  • Never push to release/X.Y; only to fork branches, PRs target the release branch for review.
  • Skip non-missing reverts (✅ DETECTED / 🟢 STATUS).
  • Verify the reverted commit is on the target branch (git merge-base --is-ancestor) before cherry-picking — 🔴 WARNING can be a false positive for RC tags from another release line.
  • Stop on cherry-pick conflicts for that revert (abort + report); do not hand-resolve unless the user asks.
  • Requires gh authenticated for pytorch/pytorch and a pytorch checkout whose origin is the user's fork.

pytorchのその他のスキル

aoti-debug
pytorch
AOTInductor(AOTI)のエラーやクラッシュをデバッグします。AOTIのセグメンテーションフォールト、デバイス不一致エラー、定数読み込み失敗、またはランタイムエラーが発生した場合に使用します。
official
pt2-bug-basher
pytorch
PyTorch 2コンパイラスタックの障害(Dynamoのグラフ分割、Inductorのコード生成エラー、AOTAutogradのクラッシュ、精度不一致など)をデバッグします。以下の場合に使用…
official
r2-outage-toggle
pytorch
Disable or re-enable Cloudflare R2 (download-r2.pytorch.org) usage in manage_v2.py during R2 outages. Can toggle R2 off/on for nightly builds, prod/stable…
official
release-create-tracker-issue
pytorch
Generate (and optionally open) a PyTorch release tracker / cherry-pick tracking issue from a release announcement, like…
official
release-create-validation-issue
pytorch
Generate a PyTorch release validation checklist issue by pulling open/closed issues from a GitHub milestone and cherry-picks from a release tracker issue.…
official
release-go-live-binary-build-matrix
pytorch
Update tools/scripts/generate_binary_build_matrix.py when a PyTorch release goes live. Advances CURRENT_STABLE_VERSION to the new stable, promotes the…
official
release-update-docker-image-pin
pytorch
Pin (or re-pin) the Linux manywheel builder docker images used by the nightly/release binary build workflows in pytorch/pytorch to a fixed .ci/docker build.…
official
vllm-pytorch-ci-triage
pytorch
Triage a failing vLLM Buildkite CI build for a PyTorch version-bump PR, isolate new regressions vs. pre-existing failures on main by comparing against recent…
official