feat: CI benchmark regression detection on PRs#2013
Merged
Conversation
Add a consolidated benchmark suite (`ci_benchmark`) and a GitHub Actions
workflow that automatically compares performance against the main branch
baseline on every PR.
Benchmark coverage (13 measurements, ~4 min on CI):
- BPE GPT-2: encode, batch, no-cache, batch-no-cache
- Llama-3: encode, batch, encode-fast, char-offsets, concurrent-4t
- Serialization: GPT-2 load, Llama-3 load, Llama-3 save
- Training: BPE small corpus
CI workflow:
- On push to main: run benchmarks, store baseline in gh-pages branch
- On PR: run benchmarks, compare vs baseline, post/update a single
PR comment with the delta table
- Alert threshold: 15% regression (warn, don't fail)
Uses benchmark-action/github-action-benchmark with criterion's bencher
output format for machine-readable results.
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
- pytest-benchmark based test suite covering: - BPE GPT-2: encode, encode_batch, multithreaded (4 workers) - Llama-3: encode, encode_batch, encode_fast, multithreaded, decode_batch - Async: async_encode_batch, async_encode_batch_fast - Serialization: from_file, to_str, from_str (roberta, llama3, albert) - Training: BPE small corpus - CI workflow: separate benchmark-python job with sccache + maturin - Re-enabled ci-benchmarks branch trigger for testing
- New benchmark-trigger.yml: maintainer comments '/benchmark' on a PR to dispatch the benchmark workflow on the PR's ref - Upload steps gated on github.event_name == 'push' so workflow_dispatch (PR runs) never overwrite the baseline - Trigger requires MEMBER/OWNER/COLLABORATOR association
Collaborator
Author
|
/benchmark |
1 similar comment
Collaborator
Author
|
/benchmark |
…en done - benchmark-trigger.yml creates a 'Benchmark Results' check on the PR head SHA - benchmarks.yml marks it in_progress at start, completed (success/failure) at end - The check body contains the comparison markdown table - Can be made a required check in branch protection rules
Collaborator
Author
|
/benchmark |
1 similar comment
Collaborator
Author
|
/benchmark |
Rust: - Push to main: runs with --save-baseline main, uploads criterion data (tar.gz) + bencher output to HF Hub - workflow_dispatch: downloads criterion baseline, runs with --baseline main for automatic criterion comparison - criterion HTML report uploaded as GitHub Actions artifact (30 day retention) Python: - bench_output.json uploaded as GitHub Actions artifact - Baseline stored/compared via HF Hub as before Both: - Artifacts downloadable from the workflow run page for manual inspection - Comparison tables posted to PR comments
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
2 participants
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.
Summary
Adds automated benchmark regression detection to CI. Every PR gets a comment showing how it compares against main.
New benchmark suite (
ci_benchmark.rs)13 measurements covering all key performance surfaces:
Runs in ~4 minutes on CI.
GitHub Actions workflow (
benchmarks.yml)gh-pagesbranchbenchmark-action/github-action-benchmarkwith criterion's bencher output formatLocal numbers (Apple M-series, sample_size=10)
Setup needed
The
gh-pagesbranch needs to be initialized once (can be empty) for the baseline storage to work. After the first push to main with this workflow, subsequent PRs will get comparison comments automatically.