Hi all,
Here is what is probably the second-to-last NEP for NumPy 2.0 (the last one
being the informational summary NEP of all major changes):
https://github.com/numpy/numpy/pull/25542. Full text below.
A lot of the work has been under discussion since the 2.0 developer meeting
in April and has been merged. A few of PRs that didn't make sense as
standalone changes without this NEP are still open (see the "NumPy 2.0 API
Changes" label), and there's a couple ones that still need to be opened.
For editorial comments on the text, please comment on GitHub. For
significant conceptual/design comments, please post them on this thread.
Cheers,
Ralf
=============================================================
NEP 56 — Array API standard support in NumPy's main namespace
=============================================================
:Author: Ralf Gommers <ralf.gomm...@gmail.com>
:Author: Mateusz Sokół <mso...@quansight.com>
:Author: Nathan Goldbaum <ngoldb...@quansight.com>
:Status: Draft
:Type: Standards Track
:Created: 2023-12-19
:Resolution: TODO mailing list link
Abstract
--------
This NEP proposes adding full support for the 2022.12 version of the array
API
standard in NumPy's main namespace for the 2.0 release.
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
<
https://github.com/numpy/archive/blob/main/2.0_developer_meeting/NumPy_2.0_devmeeting_array_API_adoption.pdf
>`__
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.
:ref:`NEP47` 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
(:ref:`NEP50`), 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) <
https://numpy.org/doc/1.26/reference/array_api.html#table-of-differences-between-numpy-array-api-and-numpy
>`__
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.
We'll note that one remaining incompatibility will be that NumPy 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).
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 mostly array-compatible interface, this
doesn't
seem like the highest-prio item regarding array API standard compatibility.
Here are what we see as the main expected benefits and costs of the complete
set of proposed changes:
Benefits:
- It will remove the "having to make a choice between the NumPy API and the
Array API" issue for other libraries,
- 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 <
https://data-apis.org/array-api/2022.12/extensions/linear_algebra_functions.html#design-principles
>`__
and `data-dependent output shapes page <
https://data-apis.org/array-api/2022.12/design_topics/data_dependent_output_shapes.html
>`__
in the standard),
- Easier for CuPy, JAX, PyTorch, Dask, Numba, and other such libraries and
compilers to fully match or support NumPy,
- 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),
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 gains - 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 as easily as possible - are
significant too. The number of breaking changes needed is fairly limited,
and
the impact of those changes seems modest. Not painless, but 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,
- 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:
- :ref:`NEP30` (never implemented)
- :ref:`NEP31` (never implemented)
- :ref:`NEP37` (never implemented; the ``__array_module__`` idea is
basically
the same as ``__array_namespace__``)
- :ref:`NEP47` (implemented with an experimental label in
``numpy.array_api``,
will be removed)
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``:
.. code:: python
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,
array API standard support should 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 few tolerance keywords,
4. Functions moved to ``numpy.linalg`` and supporting stacking/batching.
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,
5. In-place operators are disallowed when the left-hand side would be
promoted.
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. It also includes dtype changes: ``sum`` and ``prod``
always upcast lower-precision floating-point dtypes to ``float64`` when
``dtype=None`` (this upcasting is already done for inputs with
lower-precision
integer dtypes).
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,
- ``argsort`` and ``sort`` will gain a ``stable`` keyword argument in
addition
to ``kind``, and the default will become ``stable=True``.
- The ``ddof`` keyword in ``std`` and ``var`` changes its name to
``correction``.
The ``diagonal`` and ``trace`` functions are 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 (for consistency, and to support stacking). 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.
A related note on terminology: "stacking" is a confusing and fairly
NumPy-specific term, this is called "batching" in deep learning frameworks
and
elsewhere - we plan to change the terminology in NumPy to "batching".
There may be other minor changes that don't quite fall in one of the
categories
above. For example, ``numpy.fft`` functions need to preserve precision for
32-bit input dtypes rather than upcast to ``float64``/``complex128``. And
there's an issue with the ``s``/``axes`` argument in n-D transforms that
needs
solving (see `gh-25495 <https://github.com/numpy/numpy/pull/25495>`__).
Adapting to the changes
^^^^^^^^^^^^^^^^^^^^^^^
Some part of the Array API has 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 canonical names for each
dtype.
All instructions for migrating to a NEP 52 compatible codebase are
available in
the `NumPy 2.0 Migration Guide
<https://numpy.org/devdocs/numpy_2_0_migration_guide.html>`__ .
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 Array API compatible codebase, a new rule is
being implemented (see issue `ruff#8615 <
https://github.com/astral-sh/ruff/issues/8615>`__
and PR `ruff#8910 <https://github.com/astral-sh/ruff/pull/8910>`__).
With both rules in place a downstream user should be able to update their
project, to the extent that 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), this will be
handled the same way as any other backwards incompatible change in NumPy
2.0 -
through documentation, release notes, and 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 would not exist
and
we wouldn't have to think/worry about its main goal of 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:
.. code::
# 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
<https://github.com/numpy/numpy/issues/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.
New function name aliases
^^^^^^^^^^^^^^^^^^^^^^^^^
In the Python API cleanup for NumPy 2.0 (see :ref:`NEP52`) 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.
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) and ``stable`` is complementary to ``kind``, 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``.
This new API helps to better anticipate the return object and support
clearer
typing.
``np.bool`` addition
^^^^^^^^^^^^^^^^^^^^
One of the aliases that used to live in NumPy but got removed was
``np.bool``.
To comply with the Array API it got 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 perhaps if it had not been part of the array API standard.
Related work
------------
The array API standard (`html docs <https://data-apis.org/array-api/2022.12/
>`__,
`repository <https://github.com/data-apis/array-api/>`__) is the first
related
work; a lot of design discussion in its issue tracker may be relevant in
case
reasons for particular decisions need to be found.
Downstream adoption from array-consuming libraries is actively happening at
the moment,
see for example:
- scikit-learn `docs on array API support <
https://scikit-learn.org/dev/modules/array_api.html>`__ and
`PRs <
https://github.com/scikit-learn/scikit-learn/pulls?q=is%3Aopen+is%3Apr+label%3A%22Array+API%22>`__
and
`issues <
https://github.com/scikit-learn/scikit-learn/issues?q=is%3Aopen+is%3Aissue+label%3A%22Array+API%22
>`__
labeled with *Array API*.
- SciPy `docs on array API support <
http://scipy.github.io/devdocs/dev/api-dev/array_api.html>`__
and `PRs <
https://github.com/scipy/scipy/pulls?q=is%3Aopen+is%3Apr+label%3A%22array+types%22
>`__
and `issues <
https://github.com/scipy/scipy/issues?q=is%3Aopen+is%3Aissue+label%3A%22array+types%22>`__
labeled with *array types*.
- Einops `docs on supported frameworks <
https://einops.rocks/#supported-frameworks>`__
and `PR to implement array API standard support <
https://github.com/arogozhnikov/einops/pull/261>`__.
Other array libraries either already have support or are implementing
support
for the array API standard (in sync with the changes for NumPy 2.0, since
they
usually try to be as compatible to NumPy as possible). For example:
- CuPy's `docs on array API support <
https://docs.cupy.dev/en/stable/reference/array_api.html>`__
and `PRs labelled with array-api <
https://github.com/cupy/cupy/pulls?q=is%3Aopen+is%3Apr+label%3Aarray-api
>`__.
- JAX: enhancement proposal `Scope of JAX NumPy & SciPy Wrappers <
https://jax.readthedocs.io/en/latest/jep/18137-numpy-scipy-scope.html#axis-2-array-api-alignment
>`__
and `PR with initial implementation <
https://github.com/google/jax/pull/16099>`__.
Implementation
--------------
The tracking issue for Array API standard support
(`gh-25076 <https://github.com/numpy/numpy/issues/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 <
https://github.com/numpy/numpy/issues/23999>`__).
The PR that was merged as one of the first contained a new CI job that adds
the
`array-api-tests <https://github.com/data-apis/array-api-tests>`__ test
suite.
This way we had a better control over which batch of functions/aliases is
being
added each time, and be sure that the implementation conforms with the Array
API standard (see `gh-25167 <https://github.com/numpy/numpy/pull/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:
- `gh-25167: MAINT: Add array-api-tests CI stage, add
ndarray.__array_namespace__ <https://github.com/numpy/numpy/pull/25167>`__.
- `gh-25088: API: Add Array API setops [Array API] <
https://github.com/numpy/numpy/pull/25088>`__
- `gh-25155: API: Add matrix_norm, vector_norm, vecdot and matrix_transpose
[Array API] <https://github.com/numpy/numpy/pull/25155>`__
- `gh-25080: API: Add and redefine numpy.bool [Array API] <
https://github.com/numpy/numpy/pull/25080>`__
- `gh-25054: API: Introduce np.isdtype function [Array API] <
https://github.com/numpy/numpy/pull/25054>`__
- `gh-25168: API: Introduce copy argument for np.asarray [Array API] <
https://github.com/numpy/numpy/pull/25168>`__
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 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
------------------------
.. [1] https://labs.quansight.org/blog/array-api-support-scikit-learn
.. [2] https://labs.quansight.org/blog/scipy-array-api
.. [3] A. Meurer et al., "Python Array API Standard: Toward Array
Interoperability in the Scientific Python Ecosystem." (2023),
https://conference.scipy.org/proceedings/scipy2023/pdfs/aaron_meurer.pdf
.. [4] https://numpy.org/user-survey-2020/, 2020 NumPy User Survey results
Copyright
---------
This document has been placed in the public domain.
_______________________________________________
NumPy-Discussion mailing list -- numpy-discussion@python.org
To unsubscribe send an email to numpy-discussion-le...@python.org
https://mail.python.org/mailman3/lists/numpy-discussion.python.org/
Member address: arch...@mail-archive.com
