[BUG]: ProgramOptions(numba_debug=True) always fails on the NVVM backend

[fedonman]

Is this a duplicate?

No.

Type of Bug

Runtime Error

Component

cuda.core

Describe the bug

ProgramOptions(numba_debug=True) never works on the NVVM backend. cuda.core emits the option to libNVVM as --numba-debug, and libNVVM accepts only the single-dashed -numba-debug, so every such compile raises NVVM_ERROR_INVALID_OPTION. libNVVM rejects the double-dashed form of every option, not just this one, so no toolkit version accepts the current spelling. The NVRTC backend tolerates both spellings and is unaffected.

This also hides itself: test_nvvm_program_numba_debug is gated on a probe that asks libNVVM about the same wrong spelling, so the test skips everywhere. It is one of the entries in #2077. The skip reason blames CTK 13.2, but the option is rejected on CTK 13.3 as well, so the toolkit version is not the cause.

Present in released cuda-core 1.1.0 and 1.1.1, and on main. Added in #2158, closing #1287.

How to Reproduce

from cuda.core import Device, Program, ProgramOptions
from cuda.core._program import _get_nvvm_module

nvvm = _get_nvvm_module()
ir = 'target triple = "nvptx64-unknown-cuda"\ntarget datalayout = "e-p:64:64:64-i64:64:64-n16:32:64"\ndefine void @simple() {\nentry:\n  ret void\n}\n!nvvmir.version = !{!1}\n!1 = !{i32 %d, i32 %d, i32 %d, i32 %d}\n' % nvvm.ir_version()

Device(0).set_current()
opts = ProgramOptions(arch="sm_89", debug=True, numba_debug=True)
print(opts.as_bytes("nvvm"))
Program(ir, "nvvm", opts).compile("ptx")

[b'-arch=compute_89', b'-g', b'--numba-debug']
Traceback (most recent call last):
  ...
cuda.bindings.nvvm.nvvmError: ERROR_INVALID_OPTION (7)
NVVM program log: libnvvm : error: --numba-debug is an unsupported option

Asking libNVVM directly which spelling it accepts:

$ python -c "from cuda.bindings.utils import check_nvvm_compiler_options as c; print([(o, c([o])) for o in [\"--numba-debug\", \"-numba-debug\"]])"
[('--numba-debug', False), ('-numba-debug', True)]

Expected behavior

ProgramOptions(numba_debug=True) should compile on the NVVM backend. The option should be emitted as -numba-debug, which is also what numba-cuda emits for the NVVM path.

Operating System

Ubuntu 24.04 on WSL2, CUDA 13.3