cuda-python

What it is and what it does

cuda-python is NVIDIA's official Python interface to the CUDA platform, restructured as a metapackage that bundles multiple subcomponents. It provides both high-level, Pythonic abstractions (cuda.core for CUDA Runtime) and low-level C API bindings (cuda.bindings for CUDA Driver, NVRTC, NVML, and related libraries), plus utilities like cuda.pathfinder for locating CUDA installations. The package aims to enable end-to-end GPU development in Python without leaving the language.

The metapackage approach allows developers to install only the components they need—whether that's just the runtime bindings, low-level driver access, or specialized modules like numba.cuda for kernel JIT compilation or nvmath-python for math library access. It targets developers, researchers, and end users on Windows and Linux, supporting Python 3.10 through 3.14.

Use it for:

• Write GPU kernels in Python using numba.cuda or cuda.tile DSLs and compile them to CUDA without leaving Python.
• Access CUDA Driver and Runtime APIs directly from Python for fine-grained GPU memory and kernel control.
• Build GPU-accelerated libraries that expose CUDA functionality through idiomatic Python interfaces.
• Profile and debug CUDA Python applications using Nsight Python or CUPTI Python profiling tools.
• Perform distributed GPU computing with nvshmem4py's PGAS programming model across multiple GPUs.

Worth the install?

AI-flagged interpretation of the facts on this page — verify before relying

cuda-python is a metapackage providing Pythonic access to NVIDIA's CUDA platform, bundling low-level bindings, runtime abstractions, and utilities for GPU-accelerated computing from Python.

Yes, if you are developing GPU-accelerated Python applications and have an NVIDIA GPU with compatible drivers. The package is actively maintained, has low install friction, and provides both high-level and low-level access to CUDA. However, verify that the proprietary NVIDIA license aligns with your project's requirements before committing to it in production or redistributed code.

Install

cuda-python on PyPI

pip install cuda-python

uv add cuda-python

poetry add cuda-python

Installing cuda-python

Before you install

Low install friction; distributed as a pure-Python wheel. Actively maintained with recent commits and 3342 repository stars. Supports current Python versions (3.10–3.14) and targets both Windows and Linux.

License in practice

Licensed under LicenseRef-NVIDIA-SOFTWARE-LICENSE, a proprietary NVIDIA license. The license treatment is unclear; review NVIDIA's terms before use in commercial or redistributed projects.

Quickstart

pip install cuda-python

import cuda.core
import cuda.bindings

# Access CUDA runtime via cuda.core or low-level APIs via cuda.bindings

Requires Python 3.10 or later and an NVIDIA GPU with compatible CUDA drivers installed on the system.

Verify before relying

• Whether the proprietary NVIDIA license permits commercial use and redistribution without additional agreements.
• Whether all bundled subpackages (cuda.coop, cuda.compute, numba-cuda-mlir, etc.) are included in the base install or require separate installation.
• Performance characteristics and typical overhead of the Pythonic abstractions in cuda.core versus direct cuda.bindings calls.

Package facts

Evidence: cuda_python-13.3.1-py3-none-any.whl

Tags

More Libraries packages