Repository navigation
BUG: segfaults in macOS arm64 wheel builds #31733
Description
Four jobs are segfaulting: cp314/cp315 x openblas/accelerate. Traceback, from this CI log:
.......................................Fatal Python error: Segmentation fault
Thread 0x0000000170d7b000 (most recent call first):
File "/private/var/folders/dq/2t71w9pd3jx50vh3kp15_dt40000gn/T/cibw-run-qnqf97vd/cp314-macosx_arm64/venv-test-arm64/lib/python3.14/site-packages/execnet/gateway_base.py", line 534 in read
File "/private/var/folders/dq/2t71w9pd3jx50vh3kp15_dt40000gn/T/cibw-run-qnqf97vd/cp314-macosx_arm64/venv-test-arm64/lib/python3.14/site-packages/execnet/gateway_base.py", line 567 in from_io
File "/private/var/folders/dq/2t71w9pd3jx50vh3kp15_dt40000gn/T/cibw-run-qnqf97vd/cp314-macosx_arm64/venv-test-arm64/lib/python3.14/site-packages/execnet/gateway_base.py", line 1160 in _thread_receiver
File "/private/var/folders/dq/2t71w9pd3jx50vh3kp15_dt40000gn/T/cibw-run-qnqf97vd/cp314-macosx_arm64/venv-test-arm64/lib/python3.14/site-packages/execnet/gateway_base.py", line 341 in run
File "/private/var/folders/dq/2t71w9pd3jx50vh3kp15_dt40000gn/T/cibw-run-qnqf97vd/cp314-macosx_arm64/venv-test-arm64/lib/python3.14/site-packages/execnet/gateway_base.py", line 411 in _perform_spawn
Current thread 0x00000001ed8590c0 (most recent call first):
File "/private/var/folders/dq/2t71w9pd3jx50vh3kp15_dt40000gn/T/cibw-run-qnqf97vd/cp314-macosx_arm64/venv-test-arm64/lib/python3.14/site-packages/numpy/_core/tests/test_dtype.py", line 903 in test_tuple_recursion
File "/private/var/folders/dq/2t71w9pd3jx50vh3kp15_dt40000gn/T/cibw-run-qnqf97vd/cp314-macosx_arm64/venv-test-arm64/lib/python3.14/site-packages/numpy/testing/_private/utils.py", line 2869 in wrapper
File "/private/var/folders/dq/2t71w9pd3jx50vh3kp15_dt40000gn/T/cibw-run-qnqf97vd/cp314-macosx_arm64/venv-test-arm64/lib/python3.14/site-packages/_pytest/python.py", line 194 in pytest_pyfunc_call
These jobs were passing 3 days ago (Jun 21), so it's due to a recent PR. It was passing on the commit from gh-31609, so the most likely candidate is gh-31701.
Activity
rgommers commentedon Jun 24, 2026
gh-31699 was also merged in this window and is also a candidate.
seberg commentedon Jun 24, 2026
Ouch... this is probably the PR making DTypes heap-types. On the CI there only the x86 mac failed and thus I only skipped the test for those.
The test does some crazy stuff with deep recursion, should be a Python issue (see link above) though and the only thing to do is probably to just skip it for all MacOS runs (or all with Python 3.14+, but that'll be the same soon enough)...
Now that it seems to reproduce on more hardware, maybe it's worth a quick look, but this happened before (even if by "passing") there was no really related change in NumPy.
rgommers commentedon Jun 24, 2026
Ah missed that one (gh-31364), sorting in the GitHub UI is hard ... that must be it indeed. Skipping for now seems fine indeed.