Re-enable dim=4 in strided matmul test with proper cross-BLAS tolerances - #2949
Merged
Conversation
Contributor
|
Array API standard conformance tests for dpnp=0.21.0dev3=py314h509198e_21 ran successfully. |
Collaborator
|
coverage: 78.403% (-0.003%) from 78.406% — revert-gh-2912 into master |
antonwolfy
force-pushed
the
revert-gh-2912
branch
from
July 15, 2026 12:38
a41c3b1 to
d279e64
Compare
Contributor
|
View rendered docs @ https://intelpython.github.io/dpnp/index.html |
…test The revert of gh-2912 removed the widened tolerance on TestMatvec::test_axes, which reintroduced a CI failure on ubuntu-latest: dpnp computes matvec via oneMKL gemm_batch while NumPy uses OpenBLAS gemv, so the summation order differs and the float64 result deviates at the noise floor (~1.4e-14), exceeding the default 8e-15 tolerance. Restore factor=40 on test_axes (with an explanatory comment) while keeping the rest of the revert, i.e. re-enabling dim=4 in TestMatmul::test_strided1, whose factor=16 tolerance already absorbs any cross-BLAS discrepancy. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
antonwolfy
changed the title
antonwolfy
marked this pull request as ready for review
July 15, 2026 15:49
antonwolfy
requested review from
ndgrigorian and
vlad-perevezentsev
as code owners
July 15, 2026 15:49
antonwolfy
marked this pull request as draft
July 16, 2026 08:00
Re-enabling dim=4 in TestMatmul::test_strided1 reintroduced the CI failure gh-2912 had worked around by skipping it: for float32 the strided matmul deviates by ~1.87e-5, just above the factor=16 bound (~1.81e-5). The cause is benign cross-BLAS non-associativity -- dpnp copies the strided input to c-contiguous and runs oneMKL gemm_batch, while NumPy uses OpenBLAS, so the length-20 float32 contraction accumulates in a different order. Bump the tolerance to factor=24 (pass-bound ~2.27e-5) for float32 only; integer input is compared exactly, so the factor does not affect it. This keeps dim=4 covered without masking real regressions in the exact path. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
antonwolfy
changed the title
antonwolfy
marked this pull request as ready for review
July 23, 2026 16:19
vlad-perevezentsev
approved these changes
Jul 29, 2026
vlad-perevezentsev
left a comment
vlad-perevezentsev
left a comment
Contributor
There was a problem hiding this comment.
LGTM
Thank you @antonwolfy
github-actions Bot
added a commit
that referenced
this pull request
Jul 29, 2026
…ces (#2949) ## Summary gh-2912 introduced two temporary test workarounds after a CI failure on `ubuntu-latest`: - widened the tolerance on `TestMatvec::test_axes` to `factor=40`; - disabled the `dim=4` iteration in `TestMatmul::test_strided1`. Both failures share the same benign root cause, so this PR re-enables the disabled coverage and replaces the ad-hoc mitigations with documented, dtype-aware tolerances: - **`test_axes`** (matvec, float64) — keeps `factor=40`. dpnp computes `matvec` via oneMKL `gemm_batch` while NumPy uses OpenBLAS `gemv`; the summation order differs and the float64 result deviates at the noise floor (~1.4e-14), which exceeds the default `8e-15` tolerance. - **`test_strided1`** (matmul) — re-enables `dim=4`. dpnp copies the strided input to c-contiguous and runs it through oneMKL `gemm_batch`, again against NumPy's OpenBLAS. For float32, the length-20 contraction deviates by ~1.87e-5, just above the old `factor=16` bound (~1.81e-5), so the tolerance is bumped to `factor=24` **for float32 only** (pass-bound ~2.27e-5). Integer input is compared exactly and is unaffected. ## Root cause Neither discrepancy is a dpnp correctness bug. Both are expected floating-point non-associativity between two different BLAS backends (oneMKL vs OpenBLAS), surfaced when the conda-forge NumPy in CI advanced its bundled OpenBLAS. There is nothing to fix upstream, so documented tolerances are the correct handling and let us keep full test coverage (including `dim=4`). 9a16078
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
3 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
gh-2912 introduced two temporary test workarounds after a CI failure on
ubuntu-latest:TestMatvec::test_axestofactor=40;dim=4iteration inTestMatmul::test_strided1.Both failures share the same benign root cause, so this PR re-enables the disabled coverage and replaces the ad-hoc mitigations with documented, dtype-aware tolerances:
test_axes(matvec, float64) — keepsfactor=40. dpnp computesmatvecvia oneMKLgemm_batchwhile NumPy uses OpenBLASgemv; the summation order differs and the float64 result deviates at the noise floor (~1.4e-14), which exceeds the default8e-15tolerance.test_strided1(matmul) — re-enablesdim=4. dpnp copies the strided input to c-contiguous and runs it through oneMKLgemm_batch, again against NumPy's OpenBLAS. For float32, the length-20 contraction deviates by ~1.87e-5, just above the oldfactor=16bound (~1.81e-5), so the tolerance is bumped tofactor=24for float32 only (pass-bound ~2.27e-5). Integer input is compared exactly and is unaffected.Root cause
Neither discrepancy is a dpnp correctness bug. Both are expected floating-point non-associativity between two different BLAS backends (oneMKL vs OpenBLAS), surfaced when the conda-forge NumPy in CI advanced its bundled OpenBLAS. There is nothing to fix upstream, so documented tolerances are the correct handling and let us keep full test coverage (including
dim=4).