[BUG]: glInterop samples waived on Windows CI due to missing OpenGL runtime

[lijinf2]

[BUG]: glInterop samples waived on Windows CI due to missing OpenGL runtime

Type of Bug

Runtime Error

Component

cuda.core

Describe the bug

The glInteropMipmapLod, glInteropFluid, glInteropPlasma, and glInteropFluidNumbaCudaMlir samples are currently waived (skipped) on all Windows CI runners because opengl32.dll or its dependencies (e.g. nvoglv64.dll) cannot be loaded.

The root cause depends on the GPU driver mode:

Mode	Reason	OpenGL supported?
TCC	Compute-only driver; no display/graphics stack	❌
MCDM	Microsoft Compute Driver Model; no display	❌
WDDM	Should support OpenGL, but CI runners appear to lack the full NVIDIA display driver	⚠️

On WDDM runners (rtx4090, rtx2080), the test should be able to run once OpenGL is properly available. The current failure:

FileNotFoundError: Could not find module 'opengl32' (or one of its dependencies).

Found in PR #2266 - Migrating cuda-python samples from cuda-samples.
CI run: https://github.com/NVIDIA/cuda-python/actions/runs/33847204352?pr=2266

Current workaround

OSError is caught in create_window() and the sample exits with EXIT_WAIVED. This prevents CI failures but also means the GL interop samples are never exercised on Windows.

Expected behavior

On WDDM Windows CI runners, OpenGL should be available and the glInterop samples should run and pass. Options to investigate:

1. Ensure full NVIDIA display driver (including nvoglv64.dll) is installed on WDDM CI runners
2. Add explicit TCC/MCDM detection to waive only those modes, and let WDDM runners attempt the test

System information

• GPU: Various (rtx4090 WDDM, rtx2080 WDDM, rtxpro6000 TCC/MCDM, a100 TCC/MCDM, h100 MCDM, ...)
• OS: Windows (self-hosted CI runners)
• Affected samples: glInteropMipmapLod, glInteropFluid, glInteropPlasma, glInteropFluidNumbaCudaMlir