[CI] Bazel: Phase 1 — wire stage-b-test-1-gpu-small into pr-test.yml (demo, stage 2 of 3) - #3
Draft
Kangyan-Zhou wants to merge 2 commits into
Draft
[CI] Bazel: Phase 1 — wire stage-b-test-1-gpu-small into pr-test.yml (demo, stage 2 of 3)#3Kangyan-Zhou wants to merge 2 commits into
Kangyan-Zhou wants to merge 2 commits into
Conversation
Phase-1 demo: wires `bazel test` into the existing stage-b-test-1-gpu-small
pr-test.yml job alongside the existing run_suite.py invocation. The Bazel
step is gated to matrix.partition == 0 so the full suite runs once per
workflow (vs run_suite.py's 8-way est_time partitioning), and continues-on-
error so a Bazel-side failure cannot break the existing pipeline while
we're verifying parity.
Edit at .github/workflows/pr-test.yml stage-b-test-1-gpu-small job adds
two steps after the existing "Run test":
- "Install bazelisk" — fetches bazelisk to /usr/local/bin/bazel if the
runner doesn't already have it.
- "Run test (Bazel, parallel-shipping verification)" — invokes
`bazel test //test/... --test_tag_filters=sgl-suite-stage-b-test-1-gpu-small`
with --jobs=1 --local_test_jobs=1 to avoid GPU contention on the
1-gpu-5090 runner.
Both steps drop the partition gate + continue-on-error once the Bazel
path has been validated against run_suite.py for one or two release
cycles.
Validated on H100 (lmsysorg/sglang:dev container kangyan_bazel_test):
- yaml parses cleanly via PyYAML.
- bazel test //test/... --test_tag_filters=sgl-suite-stage-b-test-1-gpu-small
dispatches the 9 tagged tests across the 3 wired BUILDs, runs them all
with --cache_test_results=no, returns "5 pass and 4 fail" — the 4
failures are real environment gaps (model-loading and fork-test paths
needing the full ci_install_dependency.sh venv), not Bazel issues.
Depends on feat/bazel-tag-codegen for the sgl-suite-* tags. The Bazel
side never ran without that layer's registry.
Next layer (feat/bazel-manifest-job) replaces partition matrix +
auto_partition with a compute-test-manifest job that does dep-graph-aware
selection and est_time-aware bin-packing.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Three fixes from the code-reviewer agent on c317514: - Bazelisk install: gate on `bazel version` succeeding instead of `command -v bazel`. The 1-gpu-5090 runner is self-hosted with shared /usr/local/bin; a half-written binary from a prior aborted job would pass the `command -v` check but fail at runtime. Also writes the download to a temp file and `mv`s into place atomically so a partial curl can't poison the path. - Architecture detection: `uname -m` with x86_64/arm64 fallback. Currently lock-in to amd64 was fine for 1-gpu-5090, but the one-line guard future-proofs the pool against an arm64 runner being added. - Coredump capture for Bazel-side failures: `continue-on-error: true` on the Bazel step means a Bazel-only failure does NOT mark the job failed, so the existing `if: failure()` upload-cuda-coredumps never fires for those runs. Even if it did, both upload calls share `artifact-suffix: ${{ matrix.partition }}` (= "0") and would collide. Add a Bazel-specific upload with `if: matrix.partition == 0 && always()` and suffix `0-bazel` so dumps are deterministically captured during the parallel-shipping window. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Demo stack — stage 2 of 3
This stage adds a parallel-shipping bazel test step to the existing stage-b-test-1-gpu-small job. The Bazel run is gated to matrix.partition == 0 and continue-on-error, so it can never break the existing pipeline while we verify parity with run_suite.py.
Stack:
Edit
Two new steps in the stage-b-test-1-gpu-small job, both gated to matrix.partition == 0:
Plus a Bazel-specific upload-cuda-coredumps step with always() guard + distinct artifact suffix to avoid collision with the existing run_suite.py coredump upload.
Validation on H100
Review history (in chat)
code-reviewer agent flagged 2 important findings on commit c317514:
Both fixed in aed1282:
🤖 Generated with Claude Code