OpenXLA Microbenchmarking Presubmit Available
Julia Guo
We're excited to announce the initial rollout phase of our new XLA OSS Microbenchmarking Presubmit checks in Github Actions!
Background:
The OpenXLA project has lacked automated performance regression checks in presubmit. To address this, I've been building a robust microbenchmarking infrastructure to proactively detect and flag potential performance regressions before they are merged.
What's Happening Now: Optional Presubmit
Please enable the blocking performance presubmit by adding blocking_presubmit GitHub labels to your PR, example screenshot attached.
What to Expect:
The check will run a predefined benchmark (gemma3_1b_flax_call on an GPU/CPU).
It will compare the performance of your PR against an established baseline.
The initial threshold for flagging a regression is ~30% (this is a starting point and will be refined).
If a regression beyond the threshold is detected, the presubmit check will fail, effectively blocking the PR.
You can see an example of a successful run here, and a failed run with performance regression here.
Fallback Mechanisms:
To skip checks: Please remove blocking_presubmitin your PR.
Please feel free to reach out to me if you are interested in onboarding your own benchmarks.
I look forward to your feedbacks.
Best regards,
