You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
linux-64 conda CI fails on NumPy 2.3: CUDA-sim test_fp16_intrinsics_common #10807
Same workflow: py 3.13 + np 2.2 passed, py 3.14 + np 2.5 passed, all other cells passed.
5 failures on the failing job were all CUDASIM related: numba.cuda.tests.cudapy.test_intrinsics.TestCudaIntrinsic.test_fp16_intrinsics_common (sin, cos, log, log10, exp). The test compares per-element CUDASIM np.sin(x, dtype=np.float16) to a vectorized fn(x, dtype=np.float16).
NumPy 2.3 changed the array path (AVX512-FP16 SVML, ~1–4 ULP; numpy#30821). Reverted in 2.4.5. Default assert_allclose (rtol=1e-7) fails on 1 float16 ulp.
This job is CPU-only. CUDA tests run because the workflow sets NUMBA_ENABLE_CUDASIM=1 and numba.tests discovery always includes numba.cuda.tests.
Activity
swap357 commentedon Aug 28, 2026
This is CPU-specific behavior.
Failures were observed on -
Intel emerald rapids (https://github.com/numba/numba/actions/runs/32559494414/job/96999073327#step:8:719)
previous to this, workflows with py3.13+np2.3 package testing ran with -
AMD zen3 (https://github.com/numba/numba/actions/runs/31871841507/job/94981897785#step:8:720)
AMD zen4 (https://github.com/numba/numba/actions/runs/31246399942/job/93783038692#step:8:720)
esc commentedon Sep 8, 2026
Failing job: https://github.com/numba/numba/actions/runs/33952384424/job/101269798444
Failure also observed on -
Intel granite rapids (https://github.com/numba/numba/actions/runs/33952384424/job/101269798444#step:8:13424)
numba -sreportsCPU Name: graniterapidswithavx512fp16in CPU Features.Granite Rapids is Emerald Rapids' successor and also has AVX512-FP16.
swap357 commentedon Sep 8, 2026
Based on discussion in maintainers meeting, solution here is to patch the these CUDA sim tests to increase tolerance.