Reduce mac CI cost

[metascroy]

Skip macOS CoreML/MPS jobs on PRs that don't touch their relevant source. Non-PR events (push to main, schedule, workflow_dispatch, ciflow/* tag) bypass the gate so HUD signal on main is preserved. Linux/Windows jobs untouched.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/19708

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❗ 1 Active SEVs

There are 1 currently active SEVs. If your PR is affected, please view them below:

• Run pull request jobs on OSDC runners in shadow mode

❌ 6 Pending, 1 Unclassified Failure

As of commit 869c776 with merge base 3d86cc7 ():

UNCLASSIFIED FAILURE - DrCI could not classify the following job because the workflow did not run on the merge base. The failure may be pre-existing on trunk or introduced by this PR:

• Check Labels / Check labels (gh) (this job did not run on the merge base, so DrCI cannot tell whether the failure is pre-existing)
  RuntimeError: GraphQL query

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[metascroy]

@claude review this PR

[claude]

Claude finished @metascroy's task in 3m 37s —— View job

---

Review of PR #19708 — Reduce mac CI cost

• Read changed files and understand the CI gating logic
• Check path filters for correctness and completeness
• Review the changed-files reusable workflow pattern
• Post review feedback

---

Overall Assessment

The approach is sound: use a lightweight _get-changed-files.yml reusable workflow to emit the file list, then gate expensive macOS jobs with contains() checks. The github.event_name != 'pull_request' bypass correctly preserves signal on push/schedule/dispatch.

Detailed Findings

1. contains() substring semantics — works in your favor

contains() does substring matching on the space-separated file list, so contains(files, 'backends/apple/coreml') matches any file under that directory. This is the right approach for directory-prefix gating, and false positives (job runs when not strictly needed) are the safe direction.

One nit: contains(files, 'examples/models') in test-models-macos-coreml (trunk.yml:815) will also match files like examples/models_v2/... if such a path ever exists. Low risk in practice, but worth noting.

2. Missing paths that could affect gated jobs

• test-llama-runner-macos uses examples/models/llama/install_requirements.sh in its script, but that path isn't in its if: filter. A change solely to that file won't trigger the job. (Covered by examples/models/llama matching as a prefix — actually this IS covered since contains(files, 'examples/models/llama') would match examples/models/llama/install_requirements.sh. Scratch this, it's fine.)

• Core runtime changes (runtime/, exir/, extension/module/) could break any backend. The nightly schedule is the safety net here — a reasonable trade-off given the cost savings. Just confirming this is intentional.

3. test-huggingface-transformers-macos combined condition (trunk.yml:989-1000)

The nested parenthesization combining the fork check with the path gate is correctly structured:

if: |
  !github.event.pull_request.head.repo.fork &&
  (
    github.event_name != 'pull_request' ||
    contains(...)
  )

On non-PR events, github.event.pull_request.head.repo.fork is null, and !null is truthy in GitHub expressions, so the job runs. Correct.

4. Zephyr job refactoring (trunk.yml:66-224)

The matrix collapses from 5 entries to 3 (one per target), with the inlined script running both hello-executorch and mv2-ethosu samples sequentially per target. A few observations:

• This reduces job parallelism (5 → 3 jobs) but saves runner spin-up overhead and reduces total billed minutes.
• The test_zephyr.sh script is deleted and its logic inlined. This adds ~140 lines to the workflow but removes a layer of indirection.
• The MV2 section correctly skips cortex-m55 with continue, matching the old behavior where cortex-m55 only appeared in the hello-executorch matrix entries.
• Note: The PR title is "Reduce mac CI cost" but this Zephyr change affects a Linux job. Consider mentioning it in the PR description for reviewer clarity.

5. _get-changed-files.yml — Minor robustness note

The reusable workflow (line 30) fetches files via gh api ... --paginate --jq ... | tr '\n' ' '. For extremely large PRs (1000+ files), the output is a single long string in GITHUB_OUTPUT. GitHub's per-step output limit is 1MB — unlikely to be hit but worth a mental note.

6. PR description vs. actual changes

The PR description mentions changes to _unittest.yml, metal.yml, and mlx.yml, and adding nightly schedules (cron: 0 6, 0 8, 0 4). These don't appear in the current diff (only 3 files changed). Either these were dropped in a later commit or the description needs updating.

Summary

The gating logic is well-designed and correctly handles all trigger types. The contains() prefix-matching approach is a pragmatic choice that errs on the side of running jobs (safe). The Zephyr refactoring is a bonus cost reduction but might warrant a note in the description since it's unrelated to macOS.

No blocking issues found. The PR should meaningfully reduce macOS CI costs on PRs that don't touch Apple-specific code.

---