Skip to content

Make the INT8 parity test hardware-tolerant - #44

Merged
professor3333 merged 1 commit into
mainfrom
fix/int8-parity-test
Nov 6, 2026
Merged

professor3333 merged 1 commit into
mainfrom
fix/int8-parity-test

Conversation

@professor3333

Copy link
Copy Markdown
Owner

Summary

  • The v0.1.0 release workflow failed in test_onnx_export_and_int8_parity. Dynamic INT8 kernels on that runner moved one span of the tiny test model ((42, 61) → (42, 47)), and the same commit passed on the CI runner. Nothing was built or published.
  • FP32 ONNX export stays exact per document.
  • INT8 is now judged like the release gate in docs/targets.md (≤ 1 point typed-F1 loss on real data). On the 24-document tiny fixture it must keep typed F1 within 0.05 of the reference and leave at least 90% of documents identical.

Checks

  • ruff check, ruff format --check, pytest -m "not slow" pass.

The v0.1.0 release run failed test_onnx_export_and_int8_parity: on that
runner dynamic INT8 kernels moved one span of the tiny test model, while
the same commit passed on the CI runner. FP32 export stays exact; INT8
must keep typed F1 within 0.05 of the reference and leave at least 90%
of documents identical, in the spirit of the 1-point release gate.
@professor3333
professor3333 merged commit fe0e668 into main Nov 6, 2026
2 checks passed
@professor3333
professor3333 deleted the fix/int8-parity-test branch November 6, 2026 13:00
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant