cuPyNumeric

[!IMPORTANT] cuPyNumeric has reached end of life and is no longer maintained or supported. The final release is v26.06.01. No further releases, fixes, or support are planned. Existing packages and documentation remain available for historical reference, but we do not recommend starting new projects with cuPyNumeric.

Consider one of the following alternatives for multi-GPU/multi-node array computing:

• nvmath-python, in particular the nvmath.distributed module
• CuPy combined with mpi4py or Dask Array & Dask-CUDA
• JAX, in particular the multi-controller module
• PyTorch, in particular the torch.distributed module

cuPyNumeric is a high-performance array computing library that implements the NumPy API on top of the Legate framework. It enables you to run existing NumPy workflows on GPUs and distributed systems with little to no code changes.

Whether your work involves large-scale data analysis, complex simulations, or machine learning, cuPyNumeric allows you to seamlessly scale from a single CPU, to a single GPU, and up to thousands of GPUs across multiple nodes.

Installation

Existing pre-built cuPyNumeric packages remain available from conda on the legate channel and from PyPI. See https://docs.nvidia.com/cupynumeric/26.06/installation.html for details about different install configurations, or building cuPyNumeric from source.

📌 Note

Linux packages support Python versions 3.11 to 3.14. Windows is only supported through WSL.

Documentation

The cuPyNumeric documentation can be found here.

Note

The cuPyNumeric project is independent of the CuPy project. CuPy is a trademark of Preferred Networks, Inc, and the name 'cuPyNumeric' is used with their permission.