Support for the array API standard#

Note

Array API standard support is still experimental and hidden behind an environment variable. Only a small part of the public API is covered right now.

This guide describes how to use and add support for the Python array API standard. This standard allows users to use any array API compatible array library with SciPy out of the box.

The RFC defines how SciPy implements support for the standard, with the main principle being “array type in equals array type out”. In addition, the implementation does more strict validation of allowed array-like inputs, e.g. rejecting numpy matrix and masked array instances, and arrays with object dtype.

In the following, an array API compatible namespace is noted as xp.

Using array API standard support#

To enable the array API standard support, an environment variable must be set before importing SciPy:

export SCIPY_ARRAY_API=1

This both enables array API standard support and the more strict input validation for array-like arguments. Note that this environment variable is meant to be temporary, as a way to make incremental changes and merge them into ``main`` without affecting backwards compatibility immediately. We do not intend to keep this environment variable around long-term.

This clustering example shows usage with PyTorch tensors as inputs and return values:

>>> import torch
>>> from scipy.cluster.vq import vq
>>> code_book = torch.tensor([[1., 1., 1.],
...                           [2., 2., 2.]])
>>> features  = torch.tensor([[1.9, 2.3, 1.7],
...                           [1.5, 2.5, 2.2],
...                           [0.8, 0.6, 1.7]])
>>> code, dist = vq(features, code_book)
>>> code
tensor([1, 1, 0], dtype=torch.int32)
>>> dist
tensor([0.4359, 0.7348, 0.8307])

Note that the above example works for PyTorch CPU tensors. For GPU tensors or CuPy arrays, the expected result for vq is a TypeError, because vq is not a pure Python function and hence won’t work on GPU.

More strict array input validation will reject np.matrix and np.ma.MaskedArray instances, as well as arrays with object dtype:

>>> import numpy as np
>>> from scipy.cluster.vq import vq
>>> code_book = np.array([[1., 1., 1.],
...                       [2., 2., 2.]])
>>> features  = np.array([[1.9, 2.3, 1.7],
...                       [1.5, 2.5, 2.2],
...                       [0.8, 0.6, 1.7]])
>>> vq(features, code_book)
(array([1, 1, 0], dtype=int32), array([0.43588989, 0.73484692, 0.83066239]))

>>> # The above uses numpy arrays; trying to use np.matrix instances or object
>>> # arrays instead will yield an exception with `SCIPY_ARRAY_API=1`:
>>> vq(np.asmatrix(features), code_book)
...
TypeError: 'numpy.matrix' are not supported

>>> vq(np.ma.asarray(features), code_book)
...
TypeError: 'numpy.ma.MaskedArray' are not supported

>>> vq(features.astype(np.object_), code_book)
...
TypeError: object arrays are not supported

Currently supported functionality#

The following modules provide array API standard support when the environment variable is set:

Support is provided in scipy.special for the following functions: scipy.special.log_ndtr, scipy.special.ndtr, scipy.special.ndtri, scipy.special.erf, scipy.special.erfc, scipy.special.i0, scipy.special.i0e, scipy.special.i1, scipy.special.i1e, scipy.special.gammaln, scipy.special.gammainc, scipy.special.gammaincc, scipy.special.logit, and scipy.special.expit.

Implementation notes#

A key part of the support for the array API standard and specific compatibility functions for Numpy, CuPy and PyTorch is provided through array-api-compat. This package is included in the SciPy code base via a git submodule (under scipy/_lib), so no new dependencies are introduced.

array-api-compat provides generic utility functions and adds aliases such as xp.concat (which, for numpy, maps to np.concatenate). This allows using a uniform API across NumPy, PyTorch and CuPy (as of right now; support for other libraries like JAX is expected to be added in the future).

When the environment variable isn’t set and hence array API standard support in SciPy is disabled, we still use the “augmented” version of the NumPy namespace, which is array_api_compat.numpy. That should not change behavior of SciPy functions, it’s effectively the existing numpy namespace with a number of aliases added and a handful of functions amended/added for array API standard support. When support is enabled, depending on the type of arrays, xp will return the standard-compatible namespace matching the input array type to a function (e.g., if the input to cluster.vq.kmeans is a PyTorch array, then xp is array_api_compat.torch).

Adding array API standard support to a SciPy function#

As much as possible, new code added to SciPy should try to follow as closely as possible the array API standard (these functions typically are best-practice idioms for NumPy usage as well). By following the standard, effectively adding support for the array API standard is typically straightforward, and we ideally don’t need to maintain any customization.

Three helper functions are available:

  • array_namespace: return the namespace based on input arrays and do some input validation (like refusing to work with masked arrays, please see the RFC.)

  • _asarray: a drop-in replacement for asarray with the additional parameters check_finite and order. As stated above, try to limit the use of non-standard features. In the end we would want to upstream our needs to the compatibility library. Passing xp=xp avoids duplicate calls of array_namespace internally.

  • copy: an alias for _asarray(x, copy=True). The copy parameter was only introduced to np.asarray in NumPy 2.0, so use of the helper is needed to support <2.0. Passing xp=xp avoids duplicate calls of array_namespace internally.

To add support to a SciPy function which is defined in a .py file, what you have to change is:

  1. Input array validation,

  2. Using xp rather np functions,

  3. When calling into compiled code, convert the array to a NumPy array before and convert it back to the input array type after.

Input array validation uses the following pattern:

xp = array_namespace(arr) # where arr is the input array
# alternatively, if there are multiple array inputs, include them all:
xp = array_namespace(arr1, arr2)

# uses of non-standard parameters of np.asarray can be replaced with _asarray
arr = _asarray(arr, order='C', dtype=xp.float64, xp=xp)

Note that if one input is a non-numpy array type, all array-like inputs have to be of that type; trying to mix non-numpy arrays with lists, Python scalars or other arbitrary Python objects will raise an exception. For NumPy arrays, those types will continue to be accepted for backwards compatibility reasons.

If a function calls into a compiled code just once, use the following pattern:

x = np.asarray(x)  # convert to numpy right before compiled call(s)
y = _call_compiled_code(x)
y = xp.asarray(y)  # convert back to original array type

If there are multiple calls to compiled code, ensure doing the conversion just once to avoid too much overhead.

Here is an example for a hypothetical public SciPy function toto:

def toto(a, b):
    a = np.asarray(a)
    b = np.asarray(b, copy=True)

    c = np.sum(a) - np.prod(b)

    # this is some C or Cython call
    d = cdist(c)

    return d

You would convert this like so:

def toto(a, b):
    xp = array_namespace(a, b)
    a = xp.asarray(a)
    b = copy(b, xp=xp)  # our custom helper is needed for copy

    c = xp.sum(a) - xp.prod(b)

    # this is some C or Cython call
    c = np.asarray(c)
    d = cdist(c)
    d = xp.asarray(d)

    return d

Going through compiled code requires going back to a NumPy array, because SciPy’s extension modules only work with NumPy arrays (or memoryviews in the case of Cython), but not with other array types. For arrays on CPU, the conversions should be zero-copy, while on GPU and other devices the attempt at conversion will raise an exception. The reason for that is that silent data transfer between devices is considered bad practice, as it is likely to be a large and hard-to-detect performance bottleneck.

Adding tests#

The following pytest markers are available:

  • array_api_compatible -> xp: use a parametrisation to run a test on multiple array backends.

  • skip_if_array_api(*backends, reasons=None, np_only=False, cpu_only=False): skip certain backends and/or devices. np_only skips tests for all backends other than the default NumPy backend. @pytest.mark.usefixtures("skip_if_array_api") must be used alongside this marker for the skipping to apply.

The following is an example using the markers:

from scipy.conftest import array_api_compatible
...
@pytest.mark.skip_if_array_api(np_only=True,
                               reasons=['skip reason'])
@pytest.mark.usefixtures("skip_if_array_api")
@array_api_compatible
def test_toto1(self, xp):
    a = xp.asarray([1, 2, 3])
    b = xp.asarray([0, 2, 5])
    toto(a, b)
...
@pytest.mark.skip_if_array_api('array_api_strict', 'cupy',
                               reasons=['skip reason 1',
                                        'skip reason 2',])
@pytest.mark.usefixtures("skip_if_array_api")
@array_api_compatible
def test_toto2(self, xp):
    a = xp.asarray([1, 2, 3])
    b = xp.asarray([0, 2, 5])
    toto(a, b)

Passing a custom reason to reasons when cpu_only=True is unsupported since cpu_only=True can be used alongside passing backends. Also, the reason for using cpu_only is likely just that compiled code is used in the function(s) being tested.

When every test function in a file has been updated for array API compatibility, one can reduce verbosity by telling pytest to apply the markers to every test function using pytestmark:

from scipy.conftest import array_api_compatible

pytestmark = [array_api_compatible, pytest.mark.usefixtures("skip_if_array_api")]
skip_if_array_api = pytest.mark.skip_if_array_api
...
@skip_if_array_api(np_only=True, reasons=['skip reason'])
def test_toto1(self, xp):
    a = xp.asarray([1, 2, 3])
    b = xp.asarray([0, 2, 5])
    toto(a, b)

After applying these markers, dev.py test can be used with the new option -b or --array-api-backend:

python dev.py test -b numpy -b pytorch -s cluster

This automatically sets SCIPY_ARRAY_API appropriately. To test a library that has multiple devices with a non-default device, a second environment variable (SCIPY_DEVICE, only used in the test suite) can be set. Valid values depend on the array library under test, e.g. for PyTorch (currently the only library with multi-device support that is known to work) valid values are "cpu", "cuda", "mps". So to run the test suite with the PyTorch MPS backend, use: SCIPY_DEVICE=mps python dev.py test -b pytorch.

Note that there is a GitHub Actions workflow which runs pytorch-cpu.

Additional information#

Here are some additional resources which motivated some design decisions and helped during the development phase:

  • Initial PR with some discussions

  • Quick started from this PR and some inspiration taken from scikit-learn.

  • PR adding Array API surpport to scikit-learn

  • Some other relevant scikit-learn PRs: #22554 and #25956