The nightly benchmark run in tile-ai/TileOPs publishes here, and the docs site renders the newest of it.
| Branch | Holds | Written by |
|---|---|---|
snapshots |
One commit per run: bench_results.xml, test_results.xml, meta.json |
the nightly's publish-bench-data job |
main |
What this repository says about itself | people |
Nothing publishes over main, and nothing but the nightly writes snapshots:
a commit there is a claim about a measurement that happened.
| File | What it is |
|---|---|
bench_results.xml |
The benchmark run, as pytest wrote it. One testcase per workload; every measurement is a property named <implementation>_<metric>. |
test_results.xml |
The correctness run over the same commit: which ops passed, and the worst absolute error each reported. |
meta.json |
What produced the numbers: the commit, the run, the GPU and its settings, the runner image and its digest, and every installed package. |
git log --format='%ad %H' --date=short origin/snapshots # every run, newest first
git show <commit>:meta.json # what it ran on
git show <commit>:bench_results.xml > bench.xml # what it measuredReproducing one of those numbers takes three things from meta.json: commit
(the TileOPs source), environment.image_digest (the container, by digest —
the tag can have been pushed again since), and environment.gpu with the power
cap and clocks recorded beside it.