Skip to content

Commit bb8d9d9

Browse files
committed
CI: broaden mlir deselect list to full NVIDIA#135 union across platforms
The previous per-platform-tight lists were incomplete: NVIDIA/numba-cuda-mlir#135's import-time contamination poisons whichever tests reference cuda.cudadrv.driver AFTER the polluting xfail runs, and collection order varies between runs. Two consecutive Windows CI runs failed on different subsets (3 slicing tests one run, 5 interface tests the next). Deselect the full union of NVIDIA#135-listed tests + test_fortran_contiguous (observed to hit the same contamination) on both Linux and Windows. Same version guard (<= 0.4.0) still applies, so the whole block drops automatically when a newer numba-cuda-mlir release ships with the root-cause fix. Linux keeps the extra cuobjdump deselect (Linux-only environment issue).
1 parent 01ac84e commit bb8d9d9

2 files changed

Lines changed: 22 additions & 6 deletions

File tree

‎.github/workflows/test-wheel-linux.yml‎

Lines changed: 13 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -463,14 +463,23 @@ jobs:
463463
if python -c "from packaging.version import Version; import sys; sys.exit(0 if Version('${NUMBA_CUDA_MLIR_VER}') <= Version('0.4.0') else 1)"; then
464464
# NVIDIA/numba-cuda-mlir#135: serial-pytest contamination of
465465
# numba_cuda_mlir.cuda.cudadrv from an xfailed test in
466-
# test_nrt_comprehensive.py. Upstream CI hides it via
467-
# `-n auto --dist loadscope`.
466+
# test_nrt_comprehensive.py contaminates any later test that
467+
# touches cuda.cudadrv.driver. Upstream CI hides it via
468+
# `-n auto --dist loadscope`. Which specific tests fail depends
469+
# on collection order (we saw different subsets on linux-64 vs
470+
# win-64 across runs), so we deselect the union of all tests
471+
# #135 lists as vulnerable + test_fortran_contiguous (observed
472+
# to hit the same contamination in our runs).
468473
#
469474
# test_nvjitlink_jit_with_linkable_code_lto_dump_assembly_warn:
470475
# subprocess-invokes `cuobjdump`, not on PATH in the base
471-
# ubuntu:24.04 container. (No upstream fix yet — pending a
472-
# skip-guard bug to be filed against NVIDIA/numba-cuda-mlir.)
476+
# ubuntu:24.04 container. (Linux-only; Windows runners ship
477+
# cuobjdump with the local CTK. No upstream fix yet — pending
478+
# a skip-guard bug to be filed against NVIDIA/numba-cuda-mlir.)
473479
DESELECTS+=(
480+
--deselect 'tests/numba_cuda_tests/cudadrv/test_cuda_array_slicing.py::CudaArraySetting::test_no_sync_default_stream'
481+
--deselect 'tests/numba_cuda_tests/cudadrv/test_cuda_array_slicing.py::CudaArraySetting::test_no_sync_supplied_stream'
482+
--deselect 'tests/numba_cuda_tests/cudadrv/test_cuda_array_slicing.py::CudaArraySetting::test_sync'
474483
--deselect 'tests/numba_cuda_tests/cudapy/test_cuda_array_interface.py::TestCudaArrayInterface::test_consume_no_sync'
475484
--deselect 'tests/numba_cuda_tests/cudapy/test_cuda_array_interface.py::TestCudaArrayInterface::test_consume_sync'
476485
--deselect 'tests/numba_cuda_tests/cudapy/test_cuda_array_interface.py::TestCudaArrayInterface::test_launch_no_sync'

‎.github/workflows/test-wheel-windows.yml‎

Lines changed: 9 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -436,14 +436,21 @@ jobs:
436436
pip install --group test
437437
# Version-gated deselects — dropped automatically when newer
438438
# cuSIMT release ships. See linux step for full rationale.
439+
# NVIDIA/numba-cuda-mlir#135 poisons a subset of tests that
440+
# varies across runs based on collection order, so we deselect
441+
# the full union rather than trying to enumerate what happened
442+
# to fail on the most recent nightly.
439443
DESELECTS=()
440444
if python -c "from packaging.version import Version; import sys; sys.exit(0 if Version('${NUMBA_CUDA_MLIR_VER}') <= Version('0.4.0') else 1)"; then
441-
# Subset of NVIDIA/numba-cuda-mlir#135 tests that surface on
442-
# win-64 (different serial-pytest ordering than Linux).
443445
DESELECTS+=(
444446
--deselect 'tests/numba_cuda_tests/cudadrv/test_cuda_array_slicing.py::CudaArraySetting::test_no_sync_default_stream'
445447
--deselect 'tests/numba_cuda_tests/cudadrv/test_cuda_array_slicing.py::CudaArraySetting::test_no_sync_supplied_stream'
446448
--deselect 'tests/numba_cuda_tests/cudadrv/test_cuda_array_slicing.py::CudaArraySetting::test_sync'
449+
--deselect 'tests/numba_cuda_tests/cudapy/test_cuda_array_interface.py::TestCudaArrayInterface::test_consume_no_sync'
450+
--deselect 'tests/numba_cuda_tests/cudapy/test_cuda_array_interface.py::TestCudaArrayInterface::test_consume_sync'
451+
--deselect 'tests/numba_cuda_tests/cudapy/test_cuda_array_interface.py::TestCudaArrayInterface::test_launch_no_sync'
452+
--deselect 'tests/numba_cuda_tests/cudapy/test_cuda_array_interface.py::TestCudaArrayInterface::test_launch_sync'
453+
--deselect 'tests/numba_cuda_tests/cudapy/test_cuda_array_interface.py::TestCudaArrayInterface::test_launch_sync_two_streams'
447454
--deselect 'tests/numba_cuda_tests/cudapy/test_cuda_array_interface.py::TestCudaArrayInterface::test_fortran_contiguous'
448455
)
449456
fi

0 commit comments

Comments
 (0)