Installation

Runtime Requirements

CUDA Python is supported on all platforms that CUDA is supported. Specific dependencies are as follows:

  • Driver: Linux (450.80.02 or later) Windows (456.38 or later)

  • CUDA Toolkit 12.0 to 12.6

Note

Only the NVRTC redistributable component is required from the CUDA Toolkit. CUDA Toolkit Documentation Installation Guides can be used for guidance. Note that the NVRTC component in the Toolkit can be obtained via PYPI, Conda or Local Installer.

Installing from PyPI

pip install cuda-python

Installing from Conda

conda install -c nvidia cuda-python

Conda packages are assigned a dependency to CUDA Toolkit:

  • cuda-cudart (Provides CUDA headers to enable writting NVRTC kernels with CUDA types)

  • cuda-nvrtc (Provides NVRTC shared library)

Installing from Source

Build Requirements

  • CUDA Toolkit headers

  • Cython

  • pyclibrary

Remaining build and test dependencies are outlined in requirements.txt

The version of CUDA Toolkit headers must match the major.minor of CUDA Python. Note that minor version compatibility will still be maintained.

During the build process, environment variable CUDA_HOME or CUDA_PATH are used to find the location of CUDA headers. In particular, if your headers are located in path /usr/local/cuda/include, then you should set CUDA_HOME as follows:

export CUDA_HOME=/usr/local/cuda

In-place

To compile the extension in-place, run:

python setup.py build_ext --inplace

To compile for debugging the extension modules with gdb, pass the --debug argument to setup.py.

Develop

You can use

pip install -e .

to install the module as editible in your current Python environment (e.g. for testing of porting other libraries to use the binding).

Build the Docs

conda env create -f docs_src/environment-docs.yml
conda activate cuda-python-docs

Then compile and install cuda-python following the steps above.

cd docs_src
make html
open build/html/index.html

Publish the Docs

git checkout gh-pages
cd docs_src
make html
cp -a build/html/. ../docs/