Ctrl+K

NEP 56 — Array API standard support in NumPy’s main namespace#

Author:

Ralf Gommers <ralf.gommers@gmail.com>

Author:

Mateusz Sokół <msokol@quansight.com>

Author:

Nathan Goldbaum <ngoldbaum@quansight.com>

Status:

Final

Replaces:

NEP 30 — Duck typing for NumPy arrays - implementation, NEP 31 — Context-local and global overrides of the NumPy API, NEP 37 — A dispatch protocol for NumPy-like modules, NEP 47 — Adopting the array API standard

Type:

Standards Track

Created:

2023-12-19

Resolution:

https://mail.python.org/archives/list/numpy-discussion@python.org/message/Z6AA5CL47NHBNEPTFWYOTSUVSRDGHYPN/

Abstract#

This NEP proposes adding nearly full support for the 2022.12 version of the array API standard in NumPy’s main namespace for the 2.0 release.

Adoption in the main namespace has a number of advantages; most importantly for libraries that depend on NumPy and want to start supporting other array libraries. SciPy and scikit-learn are two prominent libraries already moving along this path. The need to support the array API standard in the main namespace draws from lessons learned by those libraries and the experimental numpy.array_api implementation with a different array object. There will also be benefits for other array libraries, JIT compilers like Numba, and for end users who may have an easier time switching between different array libraries.

Motivation and scope#

Note

The main changes proposed in this NEP were presented in the NumPy 2.0 Developer Meeting in April 2023 (see here for presentations from that meeting) and given a thumbs up there. The majority of the implementation work for NumPy 2.0 has already been merged. For the rest, PRs are ready - those are mainly the items that are specific to array API support and we’d probably not consider for inclusion in NumPy without that context. This NEP will focus on those APIs and PRs in a bit more detail.

NEP 47 — Adopting the array API standard contains the motivation for adding array API support to NumPy. This NEP expands on and supersedes NEP 47. The main reason NEP 47 aimed for a separate numpy.array_api submodule rather than the main namespace is that casting rules differed too much. With value-based casting being removed (NEP 50 — Promotion rules for Python scalars), that will be resolved in NumPy 2.0. Having NumPy be a superset of the array API standard will be a significant improvement for code portability to other libraries (CuPy, JAX, PyTorch, etc.) and thereby address one of the top user requests from the 2020 NumPy user survey [4] (GPU support). See the numpy.array_api API docs (1.26.x) for an overview of differences between it and the main namespace (note that the “strictness” ones are not applicable).

Experiences with numpy.array_api, which is still marked as experimental, have shown that the separate strict implementation and separate array object are mostly good for testing purposes, but not for regular usage in downstream libraries. Having support in the main namespace resolves this issue. Hence this NEP supersedes NEP 47. The numpy.array_api module will be moved to a standalone package, to facilitate easier updates not tied to a NumPy release cycle.

Some of the key design rules from the array API standard (e.g., output dtypes predictable from input dtypes, no polymorphic APIs with varying number of returns controlled by keywords) will also be applied to NumPy functions that are not part of the array API standard, because those design rules are now understood to be good practice in general. Those two design rules in particular make it easier for Numba and other JIT compilers to support NumPy or NumPy-compatible APIs. We’ll note that making existing arguments positional-only and keyword-only is a good idea for functions added to NumPy in the future, but will not be done for existing functions since each such change is a backwards compatibility break and it’s not necessary for writing code that is portable across libraries supporting the standard. An additional reason to apply those design rules to all functions in the main namespace now is that it then becomes much easier to deal with potential standardization of new functions already present in NumPy - those could otherwise be blocked or forced to use alternative function names due to the need for backwards compatibility.

It is important that new functions added to the main namespace integrate well with the rest of NumPy. So they should for example follow broadcasting and other rules as expected, and work with all NumPy’s dtypes rather than only the ones in the standard. The same goes for backwards-incompatible changes (e.g., linear algebra functions need to all support batching in the same way, and consider the last two axes as matrices). As a result, NumPy should become more rather than less consistent.

Here are what we see as the main expected benefits and costs of the complete set of proposed changes:

Benefits:

  • It will enable array-consuming libraries (the likes of SciPy and scikit-learn, as well as smaller libraries higher up the stack) to implement support for multiple array libraries,

  • It will remove the “having to make a choice between the NumPy API and the array API standard” issue for other array libraries when choosing what API to implement,

  • Easier for CuPy, JAX, PyTorch, Dask, Numba, and other such libraries and compilers to match or support NumPy, through providing a more well-defined and minimal API surface to target, as well as through resolving some differences that were caused by Numpy semantics that were hard to support in JIT compilers,

  • A few new features that have benefits independent of the standard: adding matrix_transpose and ndarray.mT, adding vecdot, introducing matrix_norm/vector_norm (they can be made gufuncs, vecdot already has a PR making it one),

  • Closer correspondence between the APIs of NumPy and other array libraries will lower the learning curve for end users when they switch from one array library to another one,

  • The array API standard tends to have more consistent behavior than NumPy itself has (in cases where there are differences between the two, see for example the linear algebra design principles and data-dependent output shapes page in the standard),

Costs:

  • A number of backwards compatibility breaks (mostly minor, see the Backwards compatibility section further down),

  • Expanding the size of the main namespace with about ~20 aliases (e.g., acos & co. with C99 names aliasing arccos & co.).

Overall we believe that the benefits significantly outweigh the costs - and are permanent, while the costs are largely temporary. In particular, the benefits to array libraries and compilers that want to achieve compatibility with NumPy are significant. And as a result, the long-term benefits for the PyData (or scientific Python) ecosystem as a whole - because of downstream libraries being able to support multiple array libraries much more easily - are significant too. The number of breaking changes needed is fairly limited, and the impact of those changes seems modest. Not painless, but we believe the impact is smaller than the impact of other breaking changes in NumPy 2.0, and a price worth paying.

In scope for this NEP are:

  • Changes to NumPy’s Python API needed to support the 2022.12 version of the array API standard, in the main namespace as well as numpy.linalg and numpy.fft,

  • Changes in the behavior of existing NumPy functions not (or not yet) present in the array API standard, to align with key design principles of the standard.

Out of scope for this NEP are:

  • Other changes to NumPy’s Python API unrelated to the array API standard,

  • Changes to NumPy’s C API.

This NEP will supersede the following NEPs:

Usage and impact#

We have several different types of users in mind: end users writing numerical code, downstream packages that depend on NumPy who want to start supporting multiple array libraries, and other array libraries and tools which aim to implement NumPy-like or NumPy-compatible APIs.

The most prominent users who will benefit from array API support are probably downstream libraries that want to start supporting CuPy, PyTorch, JAX, Dask, or other such libraries. SciPy and scikit-learn are already fairly far along the way of doing just that, and successfully support CuPy arrays and PyTorch tensors in a small part of their own APIs (that support is still marked as experimental).

The main principle they use is that they replace the regular import numpy as np with a utility function to retrieve the array library namespace from the input array. They call it xp, which is effectively an alias to np if the input is a NumPy array, cupy for a CuPy array, torch for a PyTorch tensor. This xp then allows writing code that works for all these libraries - because the array API standard is the common denominator. As a concrete example, this code is taken from scipy.cluster:

def vq_py(obs, code_book, check_finite=True):
    """Python version of vq algorithm"""
    xp = array_namespace(obs, code_book)
    obs = as_xparray(obs, xp=xp, check_finite=check_finite)
    code_book = as_xparray(code_book, xp=xp, check_finite=check_finite)

    if obs.ndim != code_book.ndim:
        raise ValueError("Observation and code_book should have the same rank")

    if obs.ndim == 1:
        obs = obs[:, xp.newaxis]
        code_book = code_book[:, xp.newaxis]

    # Once `cdist` has array API support, this `xp.asarray` call can be removed
    dist = xp.asarray(cdist(obs, code_book))
    code = xp.argmin(dist, axis=1)
    min_dist = xp.min(dist, axis=1)
    return code, min_dist

It mostly looks like normal NumPy code, but will run with for example PyTorch tensors as input and then return PyTorch tensors. There is a lot more to this story of course then this basic example. These blog posts on scikit-learn [1] and SciPy’s [2] experiences and impact (large performance gains in some cases - LinearDiscriminantAnalysis.fit showed ~28x gain with PyTorch on GPU vs. NumPy) paint a more complete picture.

For end users who are using NumPy directly, little changes aside from there being fewer differences between NumPy and other libraries they may want to use as well. This shortens their learning curve and makes it easier to switch between NumPy and PyTorch/JAX/CuPy. In addition, they should benefit from array-consuming libraries starting to support multiple array libraries, making their experience of using a stack of Python packages for scientific computing or data science more seamless.

Finally, for authors of other array libraries as well as tools like Numba, API improvements which align NumPy with the array API standard will also save them time. The design rules ([3]), and in some cases new APIs like the unique_* ones, are easier to implement on GPU and for JIT compilers as a result of more predictable behavior.

Backward compatibility#

The changes that have a backwards compatibility impact fall into these categories:

  1. Raising errors for consistency/strictness in some places where NumPy now allows more flexible behavior,

  2. Dtypes of returned arrays for some element-wise functions and reductions,

  3. Numerical behavior for a few tolerance keywords,

  4. Functions moved to numpy.linalg and supporting stacking/batching,

  5. The semantics of the copy keyword in asarray and array,

  6. Changes to numpy.fft functionality.

Raising errors for consistency/strictness includes:

  1. Making .T error for >2 dimensions,

  2. Making cross error on size-2 vectors (only size-3 vectors are supported),

  3. Making solve error on ambiguous input (only accept x2 as vector if x2.ndim == 1),

  4. outer raises rather than flattens on >1-D inputs,

We expect the impact of this category of changes to be small.

Dtypes of returned arrays for some element-wise functions and reductions includes functions where dtypes need to be preserved: ceil, floor, and trunc will start returning arrays with the same integer dtypes if the input has an integer dtype.

We expect the impact of this category of changes to be small.

Changes in numerical behavior include:

  • The rtol default value for pinv changes from 1e-15 to a dtype-dependent default value of None, interpreted as max(M, N) * finfo(result_dtype).eps,

  • The tol keyword to matrix_rank changes to rtol with a different interpretation. In addition, matrix_rank will no longer support 1-D array input,

Raising a FutureWarning for these tolerance changes doesn’t seem reasonable; they’d be spurious warnings for the vast majority of users, and it would force users to hardcode a tolerance value to avoid the warning. Changes in numerical results are in principle undesirable, so while we expect the impact to be small it would be good to do this in a major release.

We expect the impact of this category of changes to be medium. It is the only category of changes that does not result in clear exceptions or warnings, and hence if it does matter (e.g., downstream tests start failing or users notice a change in behavior) it may require more work from users to track down the problem. This should happen infrequently - one month after the PR implementing this change was merged (see gh-25437), the impact reported so far is a single test failure in AstroPy.

Functions moved to numpy.linalg and supporting stacking/batching are the diagonal and trace functions. They part of the linalg submodule in the standard, rather than the main namespace. Hence they will be introduced in numpy.linalg. They will operate on the last two rather than first two axes. This is done for consistency, since this is now other NumPy functions work, and to support “stacking” (or “batching” in more commonly used terminology in other libraries). Hence the linalg and main namespace functions of the same names will differ. This is technically not breaking, but potentially confusing because of the different behavior for functions with the same name. We may deprecate np.trace and np.diagonal to resolve it, but preferably not immediately to avoid users having to write if-2.0-else conditional code.

We expect the impact of this category of changes to be small.

The semantics of the copy keyword in asarray and array for copy=False will change from “copy if needed” to “never copy”. there are now three types of behavior rather than two - copy=None means “copy if needed”.

We expect the impact of this category of changes to be medium. In case users get an exception because they use copy=False explicitly in their copy but a copy was previously made anyway, they have to inspect their code and determine whether the intent of the code was the old or the new semantics (both seem rougly equally likely), and adapt the code as appropriate. We expect most cases to be np.array(..., copy=False), because until a few years ago that had lower overhead than np.asarray(...). This was solved though, and np.asarray(...) is idiomatic NumPy usage.

Changes to numpy.fft: all functions in the numpy.fft submodule need to preserve precision for 32-bit input dtypes rather than upcast to float64/complex128. This is a desirable change, consistent with the design of NumPy as a whole - but it’s possible that the lower precision or the dtype of the returned arrays from calls to functions in this module may affect users. This change was made by via a new gufunc-based implementation and vendoring of the C++ version of PocketFFT in (gh-25711).

A smaller backwards-incompatible change to numpy.fft is to make the behavior of the s and axes arguments in n-D transforms easier to understand by disallowing None values in s and requiring that if s is used, axes must be specified as well (see gh-25495).

We expect the impact of this category of changes to be small.

Adapting to the changes & tooling support#

Some parts of the array API have already been implemented as part of the general Python API cleanup for NumPy 2.0 (see NEP 52), such as:

  • establishing one and way for naming inf and nan that is array API compatible.

  • removing cryptic dtype names and establishing (array API compatible) canonical names for each dtype.

All instructions for migrating to a NEP 52 compatible codebase are available in the NumPy 2.0 Migration Guide .

Additionally, a new ruff rule was implemented for an automatic migration of Python API changes. It’s worth pointing out that the new rule NP201 is only to adhere to the NEP 52 changes, and does not cover using new functions that are part of the array API standard nor APIs with some types of backwards incompatible changes discussed above.

For an automated migration to an array API compatible codebase, a new rule is being implemented (see issue ruff#8615 and PR ruff#8910).

With both rules in place a downstream user should be able to update their project, to the extent that is possible with automation, to a library agnostic codebase that can benefit from different array libraries and devices.

Backwards incompatible changes that cannot be handled automatically (e.g., a change in rtol defaults for a linear algebra function) will be handled the in same way as any other backwards incompatible change in NumPy 2.0 - through documentation, release notes, API migrations and deprecations over several releases.

Detailed description#

In this section we’ll focus on specific API additions and functionality that we would not consider introducing into NumPy if the standard did not exist and we didn’t have to think/worry about its main goal: writing code that is portable across multiple array libraries and their supported features like GPUs and other hardware accelerators or JIT compilers.

device support#

Device support is perhaps the most obvious example. NumPy is and will remain a CPU-only library, so why bother introducing a ndarray.device attribute or device= keywords in several functions? This one feature is purely meant to make it easier to write code that is portable across libraries. The .device attribute will return an object representing CPU, and that object will be accepted as an input to device= keywords. For example:

# Should work when `xp` is `np` and `x1` a numpy array
x2 = xp.asarray([0, 1, 2, 3], dtype=xp.float64, device=x1.device)

This will work as expected for NumPy, creating a 1-D numpy array from the input list. It will also work for CuPy & co, where it may create a new array on a GPU or other supported device.

isdtype#

The array API standard introduced a new function isdtype for introspection of dtypes, because there was no suitable alternative in NumPy. The closest one is np.issubdtype, however that assumes a complex class hierarchy which other array libraries don’t have, isn’t the most ergonomic API, and required a larger API surface (np.floating and friends). isdtype will be the new and canonical way to introspect dtypes. All it requires from a dtype is that __eq__ is implemented and has the expected behavior when compared with other dtypes from the same library.

Note that as part of the effort on NEP 52, some dtype aliases were removed and canonical Python and C names documented. See also gh-17325 covering issues with NumPy’s lack of a good API for this.

copy keyword semantics#

The copy keyword in asarray and array will now support True/False/None with new meanings:

  • True - Always make a copy.

  • False - Never make a copy. If a copy is required, a ValueError is raised.

  • None - A copy will only be made if it is necessary (previously False).

The copy keyword in astype will stick to its current meaning, because “never copy” when asking for a cast to a different dtype doesn’t quite make sense.

There is still one hiccup for the change in semantics: if for user code np.array(obj, copy=False), NumPy may end up calling obj.__array__ and in that case turning the result into a NumPy array is the responsibility of the implementer of obj.__array__. Therefore, we need to add a copy=None keyword to __array__ as well, and pass the copy keyword value along - taking care to not break backwards compatibility when the implementer of __array__ does not yet have the new keyword (a DeprecationWarning will be emitted in that case, to allow for a gradual transition).

New function name aliases#

In the Python API cleanup for NumPy 2.0 (see NEP 52 — Python API cleanup for NumPy 2.0) we spent a lot of effort removing aliases. So introducing new aliases has to have a good rationale. In this case, it is needed in order to match other libraries. The main set of aliases added is for trigonometric functions, where the array API standard chose to follow C99 and other libraries in using acos, asin etc. rather than arccos, arcsin, etc. NumPy usually also follows C99; it is not entirely clear why this naming choice was made many years ago.

In total 13 aliases are added to the main namespace and 2 aliases to numpy.linalg:

  • trigonometry functions: acos, acosh, asin, asinh, atan, atanh, atan2

  • bit-wise functions: bitwise_left_shift, bitwise_invert, bitwise_right_shift

  • other functions: concat, permute_dims, pow

  • in numpy.linalg: tensordot, matmul

In the future NumPy can choose to hide the original names from its __dir__ to nudge users to the preferred spelling for each function.

New keywords with overlapping semantics#

Similarly to function name aliases, there are a couple of new keywords which have overlap with existing ones:

  • correction keyword for std and var (overlaps with ddof)

  • stable keyword for sort and argsort (overlaps with kind)

The correction name is for clarity (“delta degrees of freedom” is not easy to understand). stable is complementary to kind, which already has 'stable' as an option (a separate keyword may be more discoverable though and hence nice to have anyway), allowing a library to reserve the right to change/improve the stable and unstable sorting algorithms.

New unique_* functions#

The unique function, with return_index, return_inverse, and return_counts arguments that influence the cardinality of the returned tuple, is replaced in the array API by four respective functions: unique_all, unique_counts, unique_inverse, and unique_values. These new functions avoid polymorphism, which tends to be a problem for JIT compilers and static typing. Use of these functions therefore helps tools like Numba as well as users of static type checkers like Mypy.

np.bool addition#

One of the aliases that used to live in NumPy but was removed is np.bool. To comply with the array API it was reintroduced with a different meaning, as now it points to NumPy’s bool instead of a Python builtin. This change is a good idea and we were planning to make it anyway, because bool is a nicer name than bool_. However, we may not have scheduled that reintroduction of the name for 2.0 if it had not been part of the array API standard.

Parts of the standard that are not adopted#

There are a couple of things that the standard prescribes which we propose not to follow (at least at this time). These are:

  1. The requirement for sum and prod to always upcast lower-precision floating-point dtypes to float64 when dtype=None.

    Rationale: this is potentially disruptive (e.g., float32_arr - float32_arr.mean() would yield a float64 array, and double memory use). While this upcasting is already done for inputs with lower-precision integer dtypes and seems useful there to prevent overflows, it seems less reasonable to require this for floating-point dtypes.

    array-api#731 was opened to reconsider this design choice in the standard, and that was accepted for the next standard version.

  2. Making function signatures positional-only and keyword-only in many places.

    Rationale: the 2022.12 version of the standard said “must”, but this has already been softened to “should” in the about-to-be-released 2023.12 version, to recognize that it’s okay to not do this - it’s still possible for users of the array library to write their code using the recommended style after all. For NumPy these changes would be useful, and it seems likely that we may introduce many or all of them over time (and in fact ufuncs are already compliant), however there is no need to rush this change - doing so for 2.0 would be unnecessarily disruptive.

  3. The requirement “An in-place operation must have the same behavior (including special cases) as its respective binary (i.e., two operand, non-assignment) operation” (excluding the effect on views).

    Rationale: the requirement is very reasonable and probably expected behavior for most NumPy users. However, deprecating unsafe casts for in-place operators is a change for which the impact is hard to predict. Hence this needs to be investigated first, and then if the impact is low enough it may be possible to deprecate the current behavior according to NumPy’s normal backwards compatibility guidelines.

    This topic is tracked in gh-25621.

Note

We note that one NumPy-specific behavior that remains is returning array scalars rather than 0-D arrays in most cases where the standard, and other array libraries, return 0-D arrays (e.g., indexing and reductions). Array scalars basically duck type 0-D arrays, which is allowed by the standard (it doesn’t mandate that there is only one array type, nor contains isinstance checks or other semantics that won’t work with array scalars). There have been multiple discussions over the past year about the feasibility of removing array scalars from NumPy, or at least no longer returning them by default. However, this would be a large effort with some uncertainty about technical risks and impact of the change, and no one has taken it on. Given that array scalars implement a largely array-compatible interface, this doesn’t seem like the highest-prio item regarding array API standard compatibility (or in general).

Implementation#

The tracking issue for Array API standard support (gh-25076) records progress of implementing full support and links to related discussions. It lists all relevant PRs (merged and pending) that verify or provide array API support.

As NEP 52 blends to some degree with this NEP, we can find some relevant implementations and discussion also on its tracking issue (gh-23999).

The PR that was merged as one of the first contained a new CI job that adds the array-api-tests test suite. This way we had a better control over which batch of functions/aliases were being added each time, and could be sure that the implementations conformed to the array API standard (see gh-25167).

Then, we continued to merge one batch at the time, adding a specific API section. Below we list some of the more substantial ones, including some that we discussed in the previous sections of this NEP:

Alternatives#

The alternatives to implementing support for the array API standard in NumPy’s main namespace include:

  • one or more of the superseded NEPs, or

  • making ndarray.__array_namespace__() return a hidden namespace (or even another new public namespace) with compatible functions,

  • not implementing support for the array API standard at all.

The superseded NEPs all have some drawbacks compared to the array API standard, and by now a lot of work has gone into the standard - as well as adoption by other key libraries. So those alternatives are not appealing. Given the amount of interest in this topic, doing nothing also is not appealing. The “hidden namespace” option would be a smaller change to this proposal. We prefer not to do that since it leads to duplicate implementations staying around, a more complex implementation (e.g., potential issues with static typing), and still having two flavors of essentially the same API.

An alternative to removing numpy.array_api from NumPy is to keep it in its current place, since it is still useful - it is the best way to test if downstream code is actually portable between array libraries. This is a very reasonable alternative, however there is a slight preference for taking that module and turning it into a standalone package.

Discussion#

References and footnotes#

On this page