NVIDIA’s cuPyNumeric 25.03 Goes Fully Open Source with Enhanced PIP and HDF5 Support

Peter Zhang
Apr 23, 2025 13:20

NVIDIA has open-sourced cuPyNumeric 25.03, enhancing its accessibility with PIP installation and native HDF5 support, fostering transparency and collaboration in multi-GPU computing.

NVIDIA has announced that its cuPyNumeric 25.03 library is now completely open source, marking a significant milestone in its development. The update introduces powerful new capabilities, including support for PIP installation and native HDF5 IO, according to NVIDIA.

Full Open Source Transition

With this latest release, NVIDIA has open-sourced the entire stack of cuPyNumeric, including the Legate framework and runtime layer that powers it, under the Apache 2 license. This transition underscores NVIDIA’s commitment to fostering transparency, reproducibility, and collaboration in the development community. The open-source nature allows contributors to explore, audit, and enhance the system without obstacles.

PIP Installation Support

Previously installable only via conda, cuPyNumeric can now be installed using PIP, simplifying the setup process significantly. This enhancement facilitates easier integration into workflows, virtual environments, and CI pipelines. The cuPyNumeric package on PyPI is multinode and multirank capable, supporting both single-node multiple GPUs and multi-GPU multinode clusters.

Native HDF5 IO Support

Another notable feature of cuPyNumeric 25.03 is its native support for HDF5 through GPU Direct Storage, which optimizes the handling of large datasets. This feature ensures efficient IO operations, critical for high-performance computing and data-intensive applications. Users can now manage complex data structures with improved performance and portability.

Installation and Usage

The installation process has been streamlined to include a simple PIP command: pip install nvidia-cupynumeric. This update bundles all major dependencies except MPI, which are otherwise easily resolvable through PyPI. NVIDIA provides detailed guidance on setting up and running cuPyNumeric on SLURM clusters, emphasizing the ease of use in multinode and multirank environments.

For further details and to explore the comprehensive capabilities of cuPyNumeric 25.03, NVIDIA encourages users to review the official release notes and contribute to the ongoing development via the GitHub repository.

Image source: Shutterstock