CPU-only build

The Python bindings support the toolkit-free CPU-only build for API testing and debugging. Install from the repository root with:

CUDECOMP_BUILD_CPU_ONLY=ON python -m pip install .

Use the same setting when creating or installing CPU source archives. CPU package versions have a +cpu suffix. Use separate environments for CPU and GPU variants, and load only one variant per process.

The CPU binding accepts C-contiguous NumPy arrays with native byte order and allocates NumPy outputs when out is omitted. CuPy is not required. Operations complete synchronously and stream arguments are ignored. Only MPI backends are supported; set both process dimensions explicitly. Autotuning is unsupported.