The samples are grouped by the primary API they demonstrate:

• cuda_bindings/ contains low-level examples that call cuda.bindings driver, runtime, NVRTC, or NVML APIs directly.
• cuda_core/ contains high-level examples built primarily with cuda.core and cuda.compute.

Some cuda.core samples use individual cuda.bindings calls where the high-level API does not yet expose the required operation. They remain in the cuda_core collection because that is the API driving the sample workflow.

Each sample has its own README.md and requirements.txt. Run its install and launch commands from that sample's directory. From the repository root, for example:

cd samples/cuda_core/vectorAdd
pip install -r requirements.txt
python vectorAdd.py

Many samples require a supported NVIDIA GPU and CUDA Toolkit. See the individual README for hardware, toolkit, and optional dependency requirements.

Utilities are private implementation helpers for their respective sample collections:

• cuda_bindings/Utilities/ supports low-level bindings samples.
• cuda_core/Utilities/ supports high-level core samples.

They are not installed as part of the CUDA Python packages and should not be treated as public APIs.