--- name: nvidia-cuda-workflows description: Develop and verify bounded CUDA, GPU Python and video-encoding workloads while preserving a desktop GPU's availability. --- # NVIDIA and CUDA Inspect the GPU, active driver, installed toolkit and workload requirements. The CUDA version displayed by `nvidia-smi` is a driver capability, not proof of an installed development toolkit. - Prefer distribution-managed drivers. Do not mix runfiles, foreign-release repositories or driver branches casually. - Check native host compiler compatibility and select it within the project. Do not replace the system GCC or assume a particular GPU architecture. - Use a framework's documented runtime wheels inside its environment. A system CUDA toolkit is not required merely to run many prebuilt GPU wheels. - Verify allocation, actual device execution, synchronization and correct output. A silent CPU fallback is not working CUDA. - Keep tests representative and bounded. Leave VRAM for the desktop and other applications; never kill another job or change clocks/power settings as a shortcut. - Diagnose GPU errors with scoped workload/kernel evidence, not stress tests or certificate/sandbox bypasses. - Use generated frames to test NVENC unless the user requested a recording. Do not capture or upload an unrelated desktop.