array_api.rst

Python Array API Support
========================

The `Python array API standard <https://data-apis.org/array-api/2021.12/>`_ aims to provide a coherent set of
APIs for array and tensor libraries developed by the community to build upon. This solves the API fragmentation
issue across the community by offering concrete function signatures, semantics and scopes of coverage, enabling
writing backend-agnostic codes for better portability.

CuPy provides **experimental** support based on NumPy's `NEP-47 <https://numpy.org/neps/nep-0047-array-api-standard.html>`_,
which is in turn based on the v2021 standard. All of the functionalities can be accessed
through the :mod:`cupy.array_api` namespace.

NumPy's `Array API Standard Compatibility <https://numpy.org/devdocs/reference/array_api.html>`_ is an excellent starting
point to understand better the differences between the APIs under the main namespace and the :mod:`~cupy.array_api` namespace.
Keep in mind, however, that the key difference between NumPy and CuPy is that we are a GPU-only library, therefore CuPy users should be aware
of potential `device management <https://data-apis.org/array-api/latest/design_topics/device_support.html>`_ issues.
Same as in regular CuPy codes, the GPU-to-use can be specified via the :class:`~cupy.cuda.Device` objects, see
:ref:`device_management`. GPU-related semantics (e.g. streams, asynchronicity, etc) are also respected.
Finally, remember there are already :doc:`differences between NumPy and CuPy <../user_guide/difference>`,
although some of which are rectified in the standard.

.. toctree::
   :maxdepth: 2

   array_api_functions
   array_api_array