CUDA installation
The base Komorebi Math wheel needs no GPU runtime. CUDA execution uses optional PyTorch and CuPy installations that must be compatible with the NVIDIA driver and the selected Python interpreter.
Use the official PyTorch installer
to choose a CUDA-enabled build for the host. Install one CuPy distribution for
the matching CUDA line using the
official CuPy instructions.
For a configured CUDA 12.x environment, the CuPy package is cupy-cuda12x.
Do not install it alongside cupy or a second CUDA-specific CuPy package.
After those runtimes are available, install the local Komorebi wheel through the same interpreter:
python -m pip install .\dist\komorebi_math-1.0.0-py3-none-any.whl
python -m pip check
python -m komorebi_math doctor
python -m komorebi_math verify gpu
If using an existing environment, invoke its full python.exe path rather
than relying on an unrelated python on PATH. Package detection alone does
not prove that a driver, device or native runtime works.
The GPU validation guide describes the actual computations. Maintainers binding a result to exact release files should follow the release workflow.