Skip, not error, when nvJitLink is not installed

[LeSingh1]

def check_nvjitlink_usable():
    from cuda.bindings._internal import nvjitlink as inner_nvjitlink

    return inner_nvjitlink._inspect_function_pointer("__nvJitLinkVersion") != 0


pytestmark = pytest.mark.skipif(
    not check_nvjitlink_usable(), reason="nvJitLink not usable, maybe not installed or too old (<12.3)"
)

The reason string covers "maybe not installed", but the check cannot detect that case. A
zero pointer only means "library loaded, symbol not exported".
_inspect_function_pointer() loads nvJitLink lazily:

_inspect_function_pointer → _inspect_function_pointers → _check_or_init_nvjitlink →
load_library() → load_nvidia_dynamic_lib("nvJitLink")

which raises DynamicLibNotFoundError when the library is absent. Nothing catches it — and
because this runs at module import time (the pytestmark line), the whole module is a
collection ERROR on a machine without nvJitLink, rather than the skip it asks for.

The repo already handles exactly this call shape correctly in two other places:

• tests/test_cudart.py::test_getLocalRuntimeVersion → except pathfinder.DynamicLibNotFoundError: pytest.skip(...)
• tests/test_utils.py::_is_libnvvm_available → except DynamicLibNotFoundError: return False

Fix

Fold both probes into one _nvjitlink_exports(symbol_name) helper that catches
DynamicLibNotFoundError and returns False. That covers
check_nvjitlink_get_linked_ltoir_usable() too, which has the identical shape and gates
the skipif on line 182.

Test

test_check_nvjitlink_usable_without_the_library stubs
cuda.bindings._internal.nvjitlink with an _inspect_function_pointer that raises
DynamicLibNotFoundError — reproducing the not-installed condition on a machine that
does have nvJitLink, so it actually runs in CI. It fails on main (the exception escapes)
and passes with this change. Both probes are asserted.

I verified the stubbing mechanism locally against both the old and new helper shapes: the
old one propagates DynamicLibNotFoundError, the new one returns False.

ruff check and ruff format --check are clean.

This is the same defect class as #2541, which fixes the production-code copy in
cuda.bindings.utils.check_nvvm_compiler_options().

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.