Repository navigation
[BUG]: cuda.core tests fail with torch cu126 installed (NVRTC 12.6 options, torch CAI v2) #3333
Description
Is this a duplicate?
- I confirmed there appear to be no duplicate issues for this bug and that I agree to the Code of Conduct
Type of Bug
Something else
Component
cuda.core
Describe the bug
With cuda-core 1.2.2 [cu12] and torch 2.14.1+cu126 in the same environment, cuda_core/tests has 10 failures that come from the tests, not from cuda.core. The same parametrizations are unchanged on main.
-
test_program.py::test_cpp_program_with_various_options:ProgramOptions(no_cache=True),ProgramOptions(frandom_seed="12345")andProgramOptions(ofast_compile="min")fail withNVRTC_ERROR_INVALID_OPTION. torch cu126 pinsnvidia-cuda-nvrtc-cu12==12.6.85, and these options need NVRTC 12.9. Compiling an empty kernel with each option throughcuda.bindings.nvrtc:Option NVRTC 12.6.85 12.8.93 12.9.86 --no-cacheunrecognized option unrecognized option OK --frandom-seed=12345unrecognized option unrecognized option OK --Ofast-compile=mininvalid argument mininvalid argument minOK --pchunrecognized option OK OK The PCH parametrizations in the same list are already skipped below NVRTC 12.8; these three have no version gate.
-
test_utils.py::TestViewCudaArrayInterfaceGPU::test_cuda_array_interface_gpu[torch-*](7 cases) fail withBufferError: only CUDA Array Interface v3 or above is supported. torch still exports__cuda_array_interface__with"version": 2, soStridedMemoryView.from_cuda_array_interfacerejects every torch sample that_gpu_array_samples()adds. This test is listed in EPIC: Tests that never run in CI #2077 because CI has no array library installed for it; with torch installed it does run and these samples fail.
How to Reproduce
pip install "cuda-core[cu12]==1.2.2" torch --index-url ...(torch 2.14.1+cu126), plus the test dependencies.- Run
pytest cuda_core/tests/test_program.py cuda_core/tests/test_utils.pyfrom thecuda-core-v1.2.2sources on a CUDA GPU.
Expected behavior
The three option cases skip on NVRTC older than 12.9, like the PCH cases skip below 12.8. The CAI test either leaves out torch samples or expects the BufferError for CAI v2 producers.
Operating System
Ubuntu 24.04, x86_64
nvidia-smi output
Tesla V100-SXM2-16GB, driver 580 (CUDA 13.0). The failures are not GPU-specific.
Activity
leofang commentedon Nov 2, 2026
2 was fixed in #3033 I believe.
1 just needs updated NVRTC. I don't think we have very extensive per-major-minor checks for everything provided by cuda-core. While we don't want to maintain an effectively lookup table in cuda-core, @Andy-Jost might have some thoughts on how we want to approach this in the long term.
Andy-Jost commentedon Nov 2, 2026
The option tests only check that cuda-core encodes each NVRTC option correctly. For instance, they check that
ProgramOptions(no_cache=True)emits a flag that NVRTC accepts (--no-cache), not one it rejects (such as--no_cache).I agree we should not keep a per-option version table for that. A correct encoding only has to pass once, on a new-enough NVRTC, and CI already runs the newest NVRTC for each major version. I suggest we gate the option tests on the installed cuda-bindings version: they run when the NVRTC major.minor is at least the cuda-bindings major.minor, and they skip on an older NVRTC. The gate follows cuda-bindings releases, so it needs no updates when NVRTC adds options.
test(cuda.core): gate the NVRTC option tests on the cuda-bindings ver…