Array API compatibility library

This is a small wrapper around common array libraries that is compatible with the Array API standard. Currently, NumPy, CuPy, and PyTorch are supported. If you want support for other array libraries, or if you encounter any issues, please open an issue.

Note that some of the functionality in this library is backwards incompatible with the corresponding wrapped libraries. The end-goal is to eventually make each array library itself fully compatible with the array API, but this requires making backwards incompatible changes in many cases, so this will take some time.

Currently all libraries here are implemented against the 2022.12 version of the standard.

Install

array-api-compat is available on both PyPI

python -m pip install array-api-compat

and Conda-forge

conda install --channel conda-forge array-api-compat

Usage

The typical usage of this library will be to get the corresponding array API compliant namespace from the input arrays using array_namespace(), like

def your_function(x, y):
    xp = array_api_compat.array_namespace(x, y)
    # Now use xp as the array library namespace
    return xp.mean(x, axis=0) + 2*xp.std(y, axis=0)

If you wish to have library-specific code-paths, you can import the corresponding wrapped namespace for each library, like

import array_api_compat.numpy as np

import array_api_compat.cupy as cp

import array_api_compat.torch as torch

Each will include all the functions from the normal NumPy/CuPy/PyTorch namespace, except that functions that are part of the array API are wrapped so that they have the correct array API behavior. In each case, the array object used will be the same array object from the wrapped library.

Difference between array_api_compat and numpy.array_api

numpy.array_api is a strict minimal implementation of the Array API (see NEP 47). For example, numpy.array_api does not include any functions that are not part of the array API specification, and will explicitly disallow behaviors that are not required by the spec (e.g., cross-kind type promotions). (cupy.array_api is similar to numpy.array_api)

array_api_compat, on the other hand, is just an extension of the corresponding array library namespaces with changes needed to be compliant with the array API. It includes all additional library functions not mentioned in the spec, and allows any library behaviors not explicitly disallowed by it, such as cross-kind casting.

In particular, unlike numpy.array_api, this package does not use a separate Array object, but rather just uses the corresponding array library array objects (numpy.ndarray, cupy.ndarray, torch.Tensor, etc.) directly. This is because those are the objects that are going to be passed as inputs to functions by end users. This does mean that a few behaviors cannot be wrapped (see below), but most of the array API functional, so this does not affect most things.

Array consuming library authors coding against the array API may wish to test against numpy.array_api to ensure they are not using functionality outside of the standard, but prefer this implementation for the default behavior for end-users.

Helper Functions

In addition to the wrapped library namespaces and functions in the array API specification, there are several helper functions included here that aren't part of the specification but which are useful for using the array API:

• is_array_api_obj(x): Return True if x is an array API compatible array object.

• array_namespace(*xs): Get the corresponding array API namespace for the arrays xs. For example, if the arrays are NumPy arrays, the returned namespace will be array_api_compat.numpy. Note that this function will also work for namespaces that aren't supported by this compat library but which do support the array API (i.e., arrays that have the __array_namespace__ attribute).

• device(x): Equivalent to x.device in the array API specification. Included because numpy.ndarray does not include the device attribute and this library does not wrap or extend the array object. Note that for NumPy, device(x) is always "cpu".

• to_device(x, device, /, *, stream=None): Equivalent to x.to_device. Included because neither NumPy's, CuPy's, nor PyTorch's array objects include this method. For NumPy, this function effectively does nothing since the only supported device is the CPU, but for CuPy, this method supports CuPy CUDA Device and Stream objects. For PyTorch, this is the same as x.to(device) (the stream argument is not supported in PyTorch).

• size(x): Equivalent to x.size, i.e., the number of elements in the array. Included because PyTorch's Tensor defines size as a method which returns the shape, and this cannot be wrapped because this compat library doesn't wrap or extend the array objects.

Known Differences from the Array API Specification

There are some known differences between this library and the array API specification:

NumPy and CuPy

• The array methods __array_namespace__, device (for NumPy), to_device, and mT are not defined. This reuses np.ndarray and cp.ndarray and we don't want to monkeypatch or wrap it. The helper functions device() and to_device() are provided to work around these missing methods (see above). x.mT can be replaced with xp.linalg.matrix_transpose(x). array_namespace(x) should be used instead of x.__array_namespace__.

• Value-based casting for scalars will be in effect unless explicitly disabled with the environment variable NPY_PROMOTION_STATE=weak or np._set_promotion_state('weak') (requires NumPy 1.24 or newer, see NEP 50 and https://github.com/numpy/numpy/issues/22341)

• asarray() does not support copy=False.

• Functions which are not wrapped may not have the same type annotations as the spec.

• Functions which are not wrapped may not use positional-only arguments.

The minimum supported NumPy version is 1.21. However, this older version of NumPy has a few issues:

• unique_* will not compare nans as unequal.
• finfo() has no smallest_normal.
• No from_dlpack or __dlpack__.
• argmax() and argmin() do not have keepdims.
• qr() doesn't support matrix stacks.
• asarray() doesn't support copy=True (as noted above, copy=False is not supported even in the latest NumPy).
• Type promotion behavior will be value based for 0-D arrays (and there is no NPY_PROMOTION_STATE=weak to disable this).

If any of these are an issue, it is recommended to bump your minimum NumPy version.

PyTorch

• Like NumPy/CuPy, we do not wrap the torch.Tensor object. It is missing the __array_namespace__ and to_device methods, so the corresponding helper functions array_namespace() and to_device() in this library should be used instead (see above).

• The x.size attribute on torch.Tensor is a function that behaves differently from x.size in the spec. Use the size(x) helper function as a portable workaround (see above).

• PyTorch does not have unsigned integer types other than uint8, and no attempt is made to implement them here.

• PyTorch has type promotion semantics that differ from the array API specification for 0-D tensor objects. The array functions in this wrapper library do work around this, but the operators on the Tensor object do not, as no operators or methods on the Tensor object are modified. If this is a concern, use the functional form instead of the operator form, e.g., add(x, y) instead of x + y.

• unique_all() is not implemented, due to the fact that torch.unique does not support returning the indices array. The other unique_* functions are implemented.

• Slices do not support negative steps.

• std() and var() do not support floating-point correction.

• The stream argument of the to_device() helper (see above) is not supported.

• As with NumPy, type annotations and positional-only arguments may not exactly match the spec for functions that are not wrapped at all.

The minimum supported PyTorch version is 1.13.

Vendoring

This library supports vendoring as an installation method. To vendor the library, simply copy array_api_compat into the appropriate place in the library, like

cp -R array_api_compat/ mylib/vendored/array_api_compat

You may also rename it to something else if you like (nowhere in the code references the name "array_api_compat").

Alternatively, the library may be installed as dependency on PyPI.

Implementation Notes

As noted before, the goal of this library is to reuse the NumPy and CuPy array objects, rather than wrapping or extending them. This means that the functions need to accept and return np.ndarray for NumPy and cp.ndarray for CuPy.

Each namespace (array_api_compat.numpy, array_api_compat.cupy, and array_api_compat.torch) is populated with the normal library namespace (like from numpy import *). Then specific functions are replaced with wrapped variants.

Since NumPy and CuPy are nearly identical in behavior, most wrapping logic can be shared between them. Wrapped functions that have the same logic between NumPy and CuPy are in array_api_compat/common/. These functions are defined like

# In array_api_compat/common/_aliases.py

def acos(x, /, xp):
    return xp.arccos(x)

The xp argument refers to the original array namespace (either numpy or cupy). Then in the specific array_api_compat/numpy/ and array_api_compat/cupy/ namespaces, the @get_xp decorator is applied to these functions, which automatically removes the xp argument from the function signature and replaces it with the corresponding array library, like

# In array_api_compat/numpy/_aliases.py

from ..common import _aliases

import numpy as np

acos = get_xp(np)(_aliases.acos)

This acos now has the signature acos(x, /) and calls numpy.arccos.

Similarly, for CuPy:

# In array_api_compat/cupy/_aliases.py

from ..common import _aliases

import cupy as cp

acos = get_xp(cp)(_aliases.acos)

Since NumPy and CuPy are nearly identical in their behaviors, this allows writing the wrapping logic for both libraries only once.

PyTorch uses a similar layout in array_api_compat/torch/, but it differs enough from NumPy/CuPy that very few common wrappers for those libraries are reused.

See https://numpy.org/doc/stable/reference/array_api.html for a full list of changes from the base NumPy (the differences for CuPy are nearly identical). A corresponding document does not yet exist for PyTorch, but you can examine the various comments in the implementation to see what functions and behaviors have been wrapped.