Presubmit Critical Path Gating (GitHub Workflow)

[haochengd-google]

This PR is a follow-up to comment in PR 290. The CL has been approved and submitted (reference numbers, 947929859 and 948702917), and this is the PR1 submission. Two workflows have been tested in this test branch test 1 and test 2.

[!Note] Please note that only push will trigger Gating in the current setting. Please let me know if we also would like to trigger Gating when there is a pull_request.

Files

CI workflows

• .github/workflows/run_benchmarks.yml — presubmit gate workflow (runs both benchmarks via BAP on PRs)
• .github/workflows/update_d2h_h2d_baseline.yml — manually-triggerable record workflow that rewrites the baselines file

Build infrastructure (needed to build the workload in CI)

• .bazelrc — adds the oss and ci build configs (toolchain/flags, RBE remote cache, and the torch_tpu module override)

[linamy85]

There was a recent fix to recover performance on KvCacheManager, which might improve the baseline. Do we need to update the baseline?

[haochengd-google]

Thanks for reminding me! I have observed this, I will submit another CL to update the baseline.

[linamy85]

What workflow is this?
And is this triggered on presubmit? (I'm thinking that the benchmark should be launched safely and avoid abuse by external users - e.g. when approvals are granted, or if it's generated by copybara)

[haochengd-google]

Thanks for flagging this — good to double-check.

What it is: this is the presubmit gating workflow from the PR-290 split (PR 1). It runs the H2D/D2H and H2H transfer benchmarks through BAP on a self-hosted TPU runner to catch perf regressions.

Triggers: it only runs on push to main (plus a nightly schedule and manual workflow_dispatch) — there's no pull_request trigger. Since main is copybara-managed, push only fires on copybara exports that were already reviewed internally, so external PRs can't launch the self-hosted TPU runner.

If you'd like a pull_request trigger too, happy to add one with a guard (e.g. author_association / approval-required) so untrusted PRs can't run it. Let me know what you'd prefer!

[linamy85]

Got it, I think we might soon need the capability to accept external PRs. Having such guard (approval-required) should be helpful, but let's follow-up in separate CL, and first enable VM-based presubmit guard for internal developers.

[linamy85]

Also just to double check, are you saying that the push means when critique cl is created, the external presubmit will be triggered? Or is it post-submit only?

[haochengd-google]

It is supposed to trigger a presubmit. I will create a CL to test once it is merged, thanks!