cuda-core 1.0 fails to import with cuda-bindings 13.0: cuDevSmResourceSplit not found #2063

Bug description

cuda-core==1.0.0 fails to import when paired with cuda-bindings==13.0.x, despite the documented compatibility claim that cuda-core 1.0 works with cuda-bindings 13.0, 13.1, 13.2, etc.

Error

ImportError while loading conftest '/.../tests/conftest.py'.
    from cuda.core import Device
/.../cuda/core/__init__.py:71: in <module>
    from cuda.core import checkpoint, system, utils
/.../cuda/core/cu13/checkpoint.py:12: in <module>
    from cuda.core.typing import ProcessStateType as _ProcessStateType
/.../cuda/core/cu13/typing.py:13: in <module>
    from cuda.core._context import DeviceResourcesType
cuda/core/_context.pyx:1: in init cuda.core._context
cuda/core/_device_resources.pyx:1: in init cuda.core._device_resources
E   ImportError: cuda.bindings.cydriver does not export expected C function cuDevSmResourceSplit

Observed in: https://github.com/NVIDIA/numba-cuda-mlir/actions/runs/25706811043/job/75479358550?pr=23#step:18:155

Root cause

_device_resources.pyx uses Cython cimport to access cydriver.cuDevSmResourceSplit(...) inside an IF CUDA_CORE_BUILD_MAJOR >= 13: block (line 303). When Cython cimports a module, it validates at C-level module init (PyInit__device_resources) that every referenced cdef function exists in the target module's __pyx_capi__ dict — before any Python-level code can execute.

• cuDevSmResourceSplit was introduced in CUDA 13.1. cuda-bindings 13.0 only has cuDevSmResourceSplitByCount.
• The IF CUDA_CORE_BUILD_MAJOR >= 13 compile-time guard doesn't distinguish 13.0 from 13.1, so the cuDevSmResourceSplit reference is compiled into the cu13/ variant of the wheel.
• cuda-core 1.0.0 was built against cuda-bindings ≥ 13.1 (which exports cuDevSmResourceSplit), so the compiled .so has a hard ABI dependency on that symbol.
• The runtime version check in _can_use_structured_sm_split() (line 208–218) correctly guards against calling the function on older bindings, but it never gets a chance to run because the module fails to load at the Cython __Pyx_ImportFunction level.

This is a different layer from the PyCapsule-based indirection used in _resource_handles.pyx, which does Python-level __pyx_capi__ dict lookups with try/except and therefore handles missing functions gracefully.

Verified

$ pip download cuda-bindings==13.0.3 ...
# cydriver.pxd only declares cuDevSmResourceSplitByCount — no cuDevSmResourceSplit
# CU_DEV_SM_RESOURCE_GROUP_PARAMS and CUdevSmResourceGroup_flags also absent from .pxd
# (but struct/enum types don't cause __pyx_capi__ failures — only cdef functions do)

Suggested fix

Replace the direct cydriver.cuDevSmResourceSplit(...) call in _device_resources.pyx with a PyCapsule-based function pointer lookup, following the pattern established in _resource_handles.pyx:

1. Add p_cuDevSmResourceSplit = _get_optional_driver_fn("cuDevSmResourceSplit") in _resource_handles.pyx
2. Declare the function pointer in resource_handles.hpp / resource_handles.cpp
3. Replace cydriver.cuDevSmResourceSplit(...) in _device_resources.pyx with a call through the function pointer
4. The struct/enum types (CU_DEV_SM_RESOURCE_GROUP_PARAMS, CUdevSmResourceGroup_flags) can remain as direct cimports — they are resolved at compile time and do not trigger __pyx_capi__ lookups

Environment

• cuda-core 1.0.0
• cuda-bindings 13.0.3
• Python 3.11, Linux aarch64