I spun up the conda-forge BLAS testing again now that the 3.14 & 3.14t builds are in (for investigating an unrelated problem with llvm-openmp v21.1 on windows, actually), and noticed that 2 out of 4 linux-64 + 3.14t jobs crashed in test_huge_list_error.
Specifically, the crash occurred for BLAS flavours blis & MKL, while openblas & netlib had no failures. Due to the nature of threading bugs, this could be flaky though, or be dependent on totally different things than the BLAS flavour... In any case, it's the only failure in the whole test suite, which is actually a pretty good sign IMO. :)
│ =================================== FAILURES ===================================
│ ________________________ _core/tests/test_shape_base.py ________________________
│ [gw0] linux -- Python 3.14.0 $SRC_DIR_run_env/bin/python
│ worker 'gw0' crashed while running '_core/tests/test_shape_base.py::TestConcatenate::test_huge_list_error'
│ =============================== warnings summary ===============================
The last aspect that made me think "threading" is the name of the test. If the list is indeed huge, then there's a bigger window (relatively speaking) to hit some race conditions, and even more so if error paths are involved.
Note that this test did not fail on any of the other python implementations from 3.10-3.14, including 3.13t. It also doesn't seem to affect osx/win (though not all runs have finished yet).
CC @ngoldbaum
I spun up the conda-forge BLAS testing again now that the 3.14 & 3.14t builds are in (for investigating an unrelated problem with llvm-openmp v21.1 on windows, actually), and noticed that 2 out of 4
linux-64 + 3.14tjobs crashed intest_huge_list_error.Specifically, the crash occurred for BLAS flavours blis & MKL, while openblas & netlib had no failures. Due to the nature of threading bugs, this could be flaky though, or be dependent on totally different things than the BLAS flavour... In any case, it's the only failure in the whole test suite, which is actually a pretty good sign IMO. :)
The last aspect that made me think "threading" is the name of the test. If the list is indeed huge, then there's a bigger window (relatively speaking) to hit some race conditions, and even more so if error paths are involved.
Note that this test did not fail on any of the other python implementations from 3.10-3.14, including 3.13t. It also doesn't seem to affect osx/win (though not all runs have finished yet).
CC @ngoldbaum