Installation

Runtime Requirements

cuda.bindings supports the same platforms as CUDA. Runtime dependencies are:

  • Linux (x86-64, arm64) and Windows (x86-64)

  • Python 3.11 - 3.15

  • Driver: Linux (525.60.13 or later), Windows (528.33 or later)

  • Optionally, NVRTC, nvJitLink, NVVM, nvFatbin, and cuFile (Linux only) from CUDA Toolkit 12.x

These driver versions are the CUDA 12 minor-version compatibility baseline in the CUDA 12.9 release notes. Features introduced with newer drivers and PTX compilation can require a newer driver; see the minor-version compatibility limitations.

Note

The optional CUDA Toolkit components can be installed via PyPI, Conda, OS-specific package managers, or local installers (as described in the CUDA Toolkit Windows and Linux Installation Guides).

The CUDA 12 releases of cuda-python are a metapackage that depends directly on cuda-bindings. The examples below use cuda-python for existing users; install cuda-bindings directly if you only need the low-level interfaces.

Both main-tree CUDA 12 packages require Python 3.11 or newer.

Installing from PyPI

$ pip install -U "cuda-python~=12.9.0"

Install all optional dependencies with:

pip install -U "cuda-python[all]~=12.9.0"

Where the optional dependencies are:

  • nvidia-cuda-nvrtc-cu12 (Provides NVRTC shared library)

  • nvidia-nvjitlink-cu12>=12.3 (Provides nvJitLink shared library)

  • nvidia-cuda-nvcc-cu12 (Provides NVVM shared library)

  • nvidia-nvfatbin-cu12 (Provides nvFatbin shared library)

  • nvidia-cufile-cu12 (Provides cuFile shared library, Linux only)

Installing from Conda

$ conda install -c conda-forge cuda-python=12.9

Installing from Source

Start with the repository recommended clone. On native Windows outside WSL, complete its symlink setup before cloning. CUDA 12 sources live in cuda_bindings_12/; CUDA 13 sources have a separate root. Both build the same cuda-bindings distribution, so use a separate Python environment for each major.

Versions are derived from reachable v12.* tags using setuptools-scm. Historical tags from the old maintenance branch may not be ancestors of main. Before a suitable tag is reachable, follow the maintenance preview-build guidance in an independent clone; worktrees share tags with the rest of their repository.

Requirements

  • CUDA Toolkit headers[1]

  • CUDA Runtime static library[2]

Source builds require that the provided CUDA headers are of the same major.minor version as the cuda.bindings you’re trying to build. Despite this requirement, note that the minor version compatibility is still maintained. Use the CUDA_HOME (or CUDA_PATH) environment variable to specify the location of your headers. For example, if your headers are located in /usr/local/cuda/include, then you should set CUDA_HOME with:

$ export CUDA_HOME=/usr/local/cuda

See Environment Variables for a description of other build-time environment variables.

Note

Only cydriver, cyruntime and cynvrtc are impacted by the header requirement.

Editable Install

From the repository root, run:

$ cd cuda_bindings_12
$ python -m pip install -v -e .

This installs the CUDA 12 bindings as editable in your current Python environment for development and testing.