fix(bindings): repair the two example command-line paths that #2546 revives

[LeSingh1]

Two defects that both sit behind an argv option. check_cmd_line_flag currently always returns False (enumerate unpacked backwards — that is #2546), so neither can fire today; fixing the helper turns both into immediate failures. tests/test_examples.py runs every example with no arguments, so CI catches neither. Grouped because they are the same class and the same trigger.

1. simple_zero_copy.py: cudaGetDeviceCount() is not unwrapped

cuda_bindings/examples/0_Introduction/simple_zero_copy.py:71-75:

    if check_cmd_line_flag("device="):
        device_count = cudart.cudaGetDeviceCount()
        idev = int(get_cmd_line_argument_int("device="))

        if idev >= device_count or idev < 0:

The bindings return (cudaError_t, count) (runtime.pyx:21309), so device_count is a tuple and the comparison raises:

TypeError: '>=' not supported between instances of 'int' and 'tuple'

Every other call site in the examples wraps it — e.g. simple_p2p.py:52, gpu_n = check_cuda_errors(cudart.cudaGetDeviceCount()). Fixed the same way.

2. global_to_shmem_async_copy.py: grid_shared_state_kernel.z is never set

cuda_bindings/examples/3_CUDA_Features/global_to_shmem_async_copy.py:795-801:

    grid_shared_state_kernel = cudart.dim3()
    grid_shared_state_kernel.x = dims_b.x / threads_shared_state_kernel.x
    grid_shared_state_kernel.y = dims_a.y / threads_shared_state_kernel.x

C++ dim3 defaults every component to 1, which is why the C sample can write dim3 gridSharedStateKernel(a, b). cudart.dim3 is a cdef class whose backing struct is zero-initialised (runtime.pyx:7639-7643 — self._pvt_ptr = &self._pvt_val), so .z stays 0. It is then passed as gridDimZ at both cuLaunchKernel sites for Kernels.AsyncCopyMultiStageSharedState (lines 866 and 1003), which the driver rejects with CUDA_ERROR_INVALID_VALUE.

Every other dim3 in the file and in the sibling examples sets .z explicitly — including grid twelve lines above (grid.z = 1), threads_shared_state_kernel.z immediately before it, simple_cubemap_texture.py:162, and simple_p2p.py:151 — so this one omission is clearly unintended. Reachable via kernel=3.

Deliberately not touched: grid_shared_state_kernel.y = dims_a.y / threads_shared_state_kernel.x uses .x where the sibling grid.y uses .y. That matches the upstream C sample and is intentional — the 16x18 block has only 16 consumer rows.

What I ran

Environment: macOS, no CUDA driver and no CUDA toolkit.

• Did not run: either example, or cuda_bindings/tests/test_examples.py — both need a GPU, and Docs #2 additionally needs kernel=3, which is unreachable until Fix command line parsing in the cuda.bindings example helpers #2546 lands.
• Ran: a reduction of the comparison in Renaming #1 with the tuple the bindings actually return:

before -> TypeError: '>=' not supported between instances of 'int' and 'tuple'
after  -> False

• Ran: python -m py_compile, ruff check, ruff format --check on both files — clean, no new findings against a main baseline.
• Checked by inspection (not executed): cudart.dim3().z == 0. I could not import cuda.bindings here, so this rests on runtime.pyx:7639-7643 (the wrapper points at an in-object struct member, which CPython zero-fills on allocation) plus the fact that dim3 exposes no constructor arguments other than _ptr.

Overlap notes: simple_zero_copy.py is one of the files #2266 (samples migration) deletes; global_to_shmem_async_copy.py is not. If #2266 lands first, fix #1 should travel with the migrated copy — flagging rather than guessing.

Refs #2546

[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.