NumPy 2.0.0 Release Notes#

Note

The release of 2.0 is in progress and the current release overview and highlights are still in a draft state. However, the highlights should already list the most significant changes detailed in the full notes below, and those full notes should be complete (if not copy-edited well enough yet).

NumPy 2.0.0 is the first major release since 2006. It is the result of X months of development since the last feature release by Y contributors, and contains a large amount of exciting new features as well as a large amount of changes to both the Python and C APIs.

This major release includes breaking changes that could not happen in a regular minor (feature) release - including an ABI break, changes to type promotion rules, and API changes which may not have been emitting deprecation warnings in 1.26.x. Key documents related to how to adapt to changes in NumPy 2.0, in addition to these release notes, include:

Highlights#

Highlights of this release include:

  • New features:

    • A new variable-length string dtype, StringDType and a new numpy.strings namespace with performant ufuncs for string operations,

    • Support for float32 and longdouble in all numpy.fft functions,

    • Support for the array API standard in the main numpy namespace.

  • Performance improvements:

    • Sorting functions (sort, argsort, partition, argpartition) have been accelerated through the use of the Intel x86-simd-sort and Google Highway libraries, and may see large (hardware-specific) speedups,

    • macOS Accelerate support and binary wheels for macOS >=14, with significant performance improvements for linear algebra operations on macOS, and wheels that are about 3 times smaller,

    • numpy.char fixed-length string operations have been accelerated by implementing ufuncs that also support StringDType in addition to the the fixed-length string dtypes,

    • A new tracing and introspection API, opt_func_info, to determine which hardware-specific kernels are available and will be dispatched to.

  • Python API improvements:

    • A clear split between public and private API, with a new module structure, and each public function now available in a single place,

    • Many removals of non-recommended functions and aliases. This should make it easier to learn and use NumPy. The number of objects in the main namespace decreased by ~10% and in numpy.lib by ~80%,

    • Canonical dtype names and a new isdtype introspection function,

  • C API improvements:

  • Improved behavior:

    • Improvements to type promotion behavior was changed by adopting NEP 50. This fixes many user surprises about promotions which previously often depended on data values of input arrays rather than only their dtypes. Please see the NEP and the NumPy 2.0 migration guide for details as this change can lead to changes in output dtypes and lower precision results for mixed-dtype operations.

    • The default integer type on Windows is now int64 rather than int32, matching the behavior on other platforms,

    • The maximum number of array dimensions is changed from 32 to 64

  • Documentation:

    • The reference guide navigation was signficantly improved, and there is now documentation on NumPy’s module structure,

    • The building from source documentation was completely rewritten,

Furthermore there are many changes to NumPy internals, including continuing to migrate code from C to C++, that will make it easier to improve and maintain NumPy in the future.

The “no free lunch” theorem dictates that there is a price to pay for all these API and behavior improvements and better future extensibility. This price is:

  1. Backwards compatibility. There are a significant number of breaking changes to both the Python and C APIs. In the majority of cases, there are clear error messages that will inform the user how to adapt their code. However, there are also changes in behavior for which it was not possible to give such an error message - these cases are all covered in the Deprecation and Compatibility sections below, and in the NumPy 2.0 migration guide.

    Note that there is a ruff mode to auto-fix many things in Python code.

  2. Breaking changes to the NumPy ABI. As a result, binaries of packages that use the NumPy C API and were built against a NumPy 1.xx release will not work with NumPy 2.0. On import, such packages will see an ImportError with a message about binary incompatibiliy.

    It is possible to build binaries against NumPy 2.0 that will work at runtime with both NumPy 2.0 and 1.x. See NumPy 2.0-specific advice for more details.

    All downstream packages that depend on the NumPy ABI are advised to do a new release built against NumPy 2.0 and verify that that release works with both 2.0 and 1.26 - ideally in the period between 2.0.0rc1 (which will be ABI-stable) and the final 2.0.0 release to avoid problems for their users.

The Python versions supported by this release are 3.9-3.12.

NumPy 2.0 Python API removals#

  • np.geterrobj, np.seterrobj and the related ufunc keyword argument extobj= have been removed. The preferred replacement for all of these is using the context manager with np.errstate():.

    (gh-23922)

  • np.cast has been removed. The literal replacement for np.cast[dtype](arg) is np.asarray(arg, dtype=dtype).

  • np.source has been removed. The preferred replacement is inspect.getsource.

  • np.lookfor has been removed.

    (gh-24144)

  • numpy.who has been removed. As an alternative for the removed functionality, one can use a variable explorer that is available in IDEs such as Spyder or Jupyter Notebook.

    (gh-24321)

  • Warnings and exceptions present in numpy.exceptions (e.g, ComplexWarning, VisibleDeprecationWarning) are no longer exposed in the main namespace.

  • Multiple niche enums, expired members and functions have been removed from the main namespace, such as: ERR_*, SHIFT_*, np.fastCopyAndTranspose, np.kernel_version, np.numarray, np.oldnumeric and np.set_numeric_ops.

    (gh-24316)

  • Replaced from ... import * in the numpy/__init__.py with explicit imports. As a result, these main namespace members got removed: np.FLOATING_POINT_SUPPORT, np.FPE_*, np.NINF, np.PINF, np.NZERO, np.PZERO, np.CLIP, np.WRAP, np.WRAP, np.RAISE, np.BUFSIZE, np.UFUNC_BUFSIZE_DEFAULT, np.UFUNC_PYVALS_NAME, np.ALLOW_THREADS, np.MAXDIMS, np.MAY_SHARE_EXACT, np.MAY_SHARE_BOUNDS, add_newdoc, np.add_docstring and np.add_newdoc_ufunc.

    (gh-24357)

  • Alias np.float_ has been removed. Use np.float64 instead.

  • Alias np.complex_ has been removed. Use np.complex128 instead.

  • Alias np.longfloat has been removed. Use np.longdouble instead.

  • Alias np.singlecomplex has been removed. Use np.complex64 instead.

  • Alias np.cfloat has been removed. Use np.complex128 instead.

  • Alias np.longcomplex has been removed. Use np.clongdouble instead.

  • Alias np.clongfloat has been removed. Use np.clongdouble instead.

  • Alias np.string_ has been removed. Use np.bytes_ instead.

  • Alias np.unicode_ has been removed. Use np.str_ instead.

  • Alias np.Inf has been removed. Use np.inf instead.

  • Alias np.Infinity has been removed. Use np.inf instead.

  • Alias np.NaN has been removed. Use np.nan instead.

  • Alias np.infty has been removed. Use np.inf instead.

  • Alias np.mat has been removed. Use np.asmatrix instead.

  • np.issubclass_ has been removed. Use the issubclass builtin instead.

  • np.asfarray has been removed. Use np.asarray with a proper dtype instead.

  • np.set_string_function has been removed. Use np.set_printoptions instead with a formatter for custom printing of NumPy objects.

  • np.tracemalloc_domain is now only available from np.lib.

  • np.recfromcsv and recfromtxt are now only available from np.lib.npyio.

  • np.issctype, np.maximum_sctype, np.obj2sctype, np.sctype2char, np.sctypes, np.issubsctype were all removed from the main namespace without replacement, as they where niche members.

  • Deprecated np.deprecate and np.deprecate_with_doc has been removed from the main namespace. Use DeprecationWarning instead.

  • Deprecated np.safe_eval has been removed from the main namespace. Use ast.literal_eval instead.

    (gh-24376)

  • np.find_common_type has been removed. Use numpy.promote_types or numpy.result_type instead. To achieve semantics for the scalar_types argument, use numpy.result_type and pass 0, 0.0, or 0j as a Python scalar instead.

  • np.round_ has been removed. Use np.round instead.

  • np.nbytes has been removed. Use np.dtype(<dtype>).itemsize instead.

    (gh-24477)

  • np.compare_chararrays has been removed from the main namespace. Use np.char.compare_chararrays instead.

  • The charrarray in the main namespace has been deprecated. It can be imported without a deprecation warning from np.char.chararray for now, but we are planning to fully deprecate and remove chararray in the future.

  • np.format_parser has been removed from the main namespace. Use np.rec.format_parser instead.

    (gh-24587)

  • Support for seven data type string aliases has been removed from np.dtype: int0, uint0, void0, object0, str0, bytes0 and bool8.

    (gh-24807)

  • The experimental numpy.array_api submodule has been removed. Use the main numpy namespace for regular usage instead, or the separate array-api-strict package for the compliance testing use case for which numpy.array_api was mostly used.

    (gh-25911)

__array_prepare__ is removed#

UFuncs called __array_prepare__ before running computations for normal ufunc calls (not generalized ufuncs, reductions, etc.). The function was also called instead of __array_wrap__ on the results of some linear algebra functions.

It is now removed. If you use it, migrate to __array_ufunc__ or rely on __array_wrap__ which is called with a context in all cases, although only after the result array is filled. In those code paths, __array_wrap__ will now be passed a base class, rather than a subclass array.

(gh-25105)

Deprecations#

  • np.compat has been deprecated, as Python 2 is no longer supported.

  • np.safe_eval has been deprecated. ast.literal_eval should be used instead.

    (gh-23830)

  • np.recfromcsv, np.recfromtxt, np.disp, np.get_array_wrap, np.maximum_sctype, np.deprecate and np.deprecate_with_doc have been deprecated.

    (gh-24154)

  • np.trapz has been deprecated. Use np.trapezoid or a scipy.integrate function instead.

  • np.in1d has been deprecated. Use np.isin instead.

  • Alias np.row_stack has been deprecated. Use np.vstack directly.

    (gh-24445)

  • __array_wrap__ is now passed arr, context, return_scalar and support for implementations not accepting all three are deprecated. Its signature should be __array_wrap__(self, arr, context=None, return_scalar=False)

    (gh-25408)

  • Arrays of 2-dimensional vectors for np.cross have been deprecated. Use arrays of 3-dimensional vectors instead.

    (gh-24818)

  • np.dtype("a") alias for np.dtype(np.bytes_) was deprecated. Use np.dtype("S") alias instead.

    (gh-24854)

  • Use of keyword arguments x and y with functions assert_array_equal and assert_array_almost_equal has been deprecated. Pass the first two arguments as positional arguments instead.

    (gh-24978)

numpy.fft deprecations for n-D transforms with None values in arguments#

Using fftn, ifftn, rfftn, irfftn, fft2, ifft2, rfft2 or irfft2 with the s parameter set to a value that is not None and the axes parameter set to None has been deprecated, in line with the array API standard. To retain current behaviour, pass a sequence [0, …, k-1] to axes for an array of dimension k.

Furthermore, passing an array to s which contains None values is deprecated as the parameter is documented to accept a sequence of integers in both the NumPy docs and the array API specification. To use the default behaviour of the corresponding 1-D transform, pass the value matching the default for its n parameter. To use the default behaviour for every axis, the s argument can be omitted.

(gh-25495)

np.linalg.lstsq now defaults to a new rcond value#

lstsq now uses the new rcond value of the machine precision times max(M, N). Previously, the machine precision was used but a FutureWarning was given to notify that this change will happen eventually. That old behavior can still be achieved by passing rcond=-1.

(gh-25721)

Expired deprecations#

  • The np.core.umath_tests submodule has been removed from the public API. (Deprecated in NumPy 1.15)

    (gh-23809)

  • The PyDataMem_SetEventHook deprecation has expired and it is removed. Use tracemalloc and the np.lib.tracemalloc_domain domain. (Deprecated in NumPy 1.23)

    (gh-23921)

  • The deprecation of set_numeric_ops and the C functions PyArray_SetNumericOps and PyArray_GetNumericOps has been expired and the functions removed. (Deprecated in NumPy 1.16)

    (gh-23998)

  • The fasttake, fastclip, and fastputmask ArrFuncs deprecation is now finalized.

  • The deprecated function fastCopyAndTranspose and its C counterpart are now removed.

  • The deprecation of PyArray_ScalarFromObject is now finalized.

    (gh-24312)

  • np.msort has been removed. For a replacement, np.sort(a, axis=0) should be used instead.

    (gh-24494)

  • np.dtype(("f8", 1) will now return a shape 1 subarray dtype rather than a non-subarray one.

    (gh-25761)

  • Assigning to the .data attribute of an ndarray is disallowed and will raise.

  • np.binary_repr(a, width) will raise if width is too small.

  • Using NPY_CHAR in PyArray_DescrFromType() will raise, use NPY_STRING NPY_UNICODE, or NPY_VSTRING instead.

    (gh-25794)

Compatibility notes#

loadtxt and genfromtxt default encoding changed#

loadtxt and genfromtxt now both default to encoding=None which may mainly modify how converters work. These will now be passed str rather than bytes. Pass the encoding explicitly to always get the new or old behavior. For genfromtxt the change also means that returned values will now be unicode strings rather than bytes.

(gh-25158)

f2py compatibility notes#

f2py will no longer accept ambiguous -m and .pyf CLI combinations. When more than one .pyf file is passed, an error is raised. When both -m and a .pyf is passed, a warning is emitted and the -m provided name is ignored.

(gh-25181)

The f2py.compile() helper has been removed because it leaked memory, has been marked as experimental for several years now, and was implemented as a thin subprocess.run wrapper. It is also one of the test bottlenecks. See gh-25122 for the full rationale. It also used several np.distutils features which are too fragile to be ported to work with meson.

Users are urged to replace calls to f2py.compile with calls to subprocess.run("python", "-m", "numpy.f2py",... instead, and to use environment variables to interact with meson. Native files are also an option.

(gh-25193)

arange’s start argument is positional-only#

The first argument of arange is now positional only. This way, specifying a start argument as a keyword, e.g. arange(start=0, stop=4), raises a TypeError. Other behaviors, are unchanged so arange(stop=4), arange(2, stop=4) and so on, are still valid and have the same meaning as before.

(gh-25336)

Minor changes in behavior of sorting functions#

Due to algorithmic changes and use of SIMD code, sorting functions with methods that aren’t stable may return slightly different results in 2.0.0 compared to 1.26.x. This includes the default method of argsort and argpartition.

Removed ambiguity when broadcasting in np.solve#

The broadcasting rules for np.solve(a, b) were ambiguous when b had 1 fewer dimensions than a. This has been resolved in a backward-incompatible way and is now compliant with the Array API. The old behaviour can be reconstructed by using np.solve(a, b[..., None])[..., 0].

(gh-25914)

Modified representation for Polynomial#

The representation method for Polynomial was updated to include the domain in the representation. The plain text and latex representations are now consistent. For example the output of str(np.polynomial.Polynomial([1, 1], domain=[.1, .2])) used to be 1.0 + 1.0 x, but now is 1.0 + 1.0 (-3.0000000000000004 + 20.0 x).

(gh-21760)

C API changes#

  • The PyArray_CGT, PyArray_CLT, PyArray_CGE, PyArray_CLE, PyArray_CEQ, PyArray_CNE macros have been removed.

  • PyArray_MIN and PyArray_MAX have been moved from ndarraytypes.h to npy_math.h.

    (gh-24258)

  • A C API for working with numpy.dtypes.StringDType arrays has been exposed. This includes functions for acquiring and releasing mutexes which lock access to the string data, as well as packing and unpacking UTF-8 bytestreams from array entries.

  • NPY_NTYPES has been renamed to NPY_NTYPES_LEGACY as it does not include new NumPy built-in DTypes. In particular the new string DType will likely not work correctly with code that handles legacy DTypes.

    (gh-25347)

  • The C-API now only exports the static inline function versions of the array accessors (previously this depended on using “deprecated API”). While we discourage it, the struct fields can still be used directly.

    (gh-25789)

  • NumPy now defines PyArray_Pack to set an individual memory address. Unlike PyArray_SETITEM this function is equivalent to setting an individual array item and does not require a NumPy array input.

    (gh-25954)

  • The ->f slot has been removed from PyArray_Descr. If you use this slot, replace accessing it with PyDataType_GetArrFuncs (see its documentation and the NumPy 2.0 migration guide). In some cases using other functions like PyArray_GETITEM may be an alternatives.

  • PyArray_GETITEM and PyArray_SETITEM now require the import of the NumPy API table to be used and are no longer defined in ndarraytypes.h.

    (gh-25812)

  • Due to runtime dependencies, the definition for functionality accessing the dtype flags was moved from numpy/ndarraytypes.h and is only available after including numpy/ndarrayobject.h as it requires import_array(). This includes PyDataType_FLAGCHK, PyDataType_REFCHK and NPY_BEGIN_THREADS_DESCR.

  • The dtype flags on PyArray_Descr must now be accessed through the PyDataType_FLAGS inline function to be compatible with both 1.x and 2.x. This function is defined in npy_2_compat.h to allow backporting. Most or all users should use PyDataType_FLAGCHK which is available on 1.x and does not require backporting. Cython users should use Cython 3. Otherwise access will go through Python unless they use PyDataType_FLAGCHK instead.

    (gh-25816)

Datetime functionality exposed in the C API and Cython bindings#

The functions NpyDatetime_ConvertDatetime64ToDatetimeStruct, NpyDatetime_ConvertDatetimeStructToDatetime64, NpyDatetime_ConvertPyDateTimeToDatetimeStruct, NpyDatetime_GetDatetimeISO8601StrLen, NpyDatetime_MakeISO8601Datetime, and NpyDatetime_ParseISO8601Datetime have been added to the C API to facilitate converting between strings, Python datetimes, and NumPy datetimes in external libraries.

(gh-21199)

Const correctness for the generalized ufunc C API#

The NumPy C API’s functions for constructing generalized ufuncs (PyUFunc_FromFuncAndData, PyUFunc_FromFuncAndDataAndSignature, PyUFunc_FromFuncAndDataAndSignatureAndIdentity) take types and data arguments that are not modified by NumPy’s internals. Like the name and doc arguments, third-party Python extension modules are likely to supply these arguments from static constants. The types and data arguments are now const-correct: they are declared as const char *types and void *const *data, respectively. C code should not be affected, but C++ code may be.

(gh-23847)

Larger NPY_MAXDIMS and NPY_MAXARGS, NPY_RAVEL_AXIS introduced#

NPY_MAXDIMS is now 64, you may want to review its use. This is usually used in a stack allocation, where the increase should be safe. However, we do encourage generally to remove any use of NPY_MAXDIMS and NPY_MAXARGS to eventually allow removing the constraint completely. For the conversion helper and C-API functions mirroring Python ones such as take, NPY_MAXDIMS was used to mean axis=None. Such usage must be replaced with NPY_RAVEL_AXIS. See also Increased maximum number of dimensions.

(gh-25149)

NPY_MAXARGS not constant and PyArrayMultiIterObject size change#

Since NPY_MAXARGS was increased, it is now a runtime constant and not compile-time constant anymore. We expect almost no users to notice this. But if used for stack allocations it now must be replaced with a custom constant using NPY_MAXARGS as an additional runtime check.

The sizeof(PyArrayMultiIterObject) no longer includes the full size of the object. We expect nobody to notice this change. It was necessary to avoid issues with Cython.

(gh-25271)

Required changes for custom legacy user dtypes#

In order to improve our DTypes it is unfortunately necessary to break the ABI, which requires some changes for dtypes registered with PyArray_RegisterDataType. Please see the documentation of PyArray_RegisterDataType for how to adapt your code and achieve compatibility with both 1.x and 2.x.

(gh-25792)

New Public DType API#

The C implementation of the NEP 42 DType API is now public. While the DType API has shipped in NumPy for a few versions, it was only usable in sessions with a special environment variable set. It is now possible to write custom DTypes outside of NumPy using the new DType API and the normal import_array() mechanism for importing the numpy C API.

See Custom Data Types for more details about the API. As always with a new feature, please report any bugs you run into implementing or using a new DType. It is likely that downstream C code that works with dtypes will need to be updated to work correctly with new DTypes.

(gh-25754)

New C-API import functions#

We have now added PyArray_ImportNumPyAPI and PyUFunc_ImportUFuncAPI as static inline functions to import the NumPy C-API tables. The new functions have two advantages over import_array and import_ufunc:

  • They check whether the import was already performed and are light-weight if not, allowing to add them judiciously (although this is not preferable in most cases).

  • The old mechanisms were macros rather than functions which included a return statement.

The PyArray_ImportNumPyAPI() function is included in npy_2_compat.h for simpler backporting.

(gh-25866)

Structured dtype information access through functions#

The dtype structures fields c_metadata, names, fields, and subarray must now be accessed through new functions following the same names, such as PyDataType_NAMES. Direct access of the fields is not valid as they do not exist for all PyArray_Descr instances. The metadata field is kept, but the macro version should also be preferred.

(gh-25802)

Descriptor elsize and alignment access#

Unless compiling only with NumPy 2 support, the elsize and aligment fields must now be accessed via PyDataType_ELSIZE, PyDataType_SET_ELSIZE, and PyDataType_ALIGNMENT. In cases where the descriptor is attached to an array, we advise using PyArray_ITEMSIZE as it exists on all NumPy versions. Please see The PyArray_Descr struct has been changed for more information.

(gh-25943)

NumPy 2.0 C API removals#

  • npy_interrupt.h and the corresponding macros like NPY_SIGINT_ON have been removed. We recommend querying PyErr_CheckSignals() or PyOS_InterruptOccurred() periodically (these do currently require holding the GIL though).

  • The noprefix.h header has been removed. Replace missing symbols with their prefixed counterparts (usually an added NPY_ or npy_).

    (gh-23919)

  • PyUFunc_GetPyVals, PyUFunc_handlefperr, and PyUFunc_checkfperr have been removed. If needed, a new backwards compatible function to raise floating point errors could be restored. Reason for removal: there are no known users and the functions would have made with np.errstate() fixes much more difficult).

    (gh-23922)

  • The numpy/old_defines.h which was part of the API deprecated since NumPy 1.7 has been removed. This removes macros of the form PyArray_CONSTANT. The replace_old_macros.sed script may be useful to convert them to the NPY_CONSTANT version.

    (gh-24011)

  • The legacy_inner_loop_selector member of the ufunc struct is removed to simplify improvements to the dispatching system. There are no known users overriding or directly accessing this member.

    (gh-24271)

  • NPY_INTPLTR has been removed to avoid confusion (see intp redefinition).

    (gh-24888)

  • The advanced indexing MapIter and related API has been removed. The (truly) public part of it was not well tested and had only one known user (Theano). Making it private will simplify improvements to speed up ufunc.at, make advanced indexing more maintainable, and was important for increasing the maximum number of dimensions of arrays to 64. Please let us know if this API is important to you so we can find a solution together.

    (gh-25138)

  • The NPY_MAX_ELSIZE macro has been removed, as it only ever reflected builtin numeric types and served no internal purpose.

    (gh-25149)

  • PyArray_REFCNT and NPY_REFCOUNT are removed. Use Py_REFCNT instead.

    (gh-25156)

  • PyArrayFlags_Type and PyArray_NewFlagsObject as well as PyArrayFlagsObject are private now. There is no known use-case; use the Python API if needed.

  • PyArray_MoveInto, PyArray_CastTo, PyArray_CastAnyTo are removed use PyArray_CopyInto and if absolutely needed PyArray_CopyAnyInto (the latter does a flat copy).

  • PyArray_FillObjectArray is removed, its only true use is for implementing np.empty. Create a new empty array or use PyArray_FillWithScalar() (decrefs existing objects).

  • PyArray_CompareUCS4 and PyArray_CompareString are removed. Use the standard C string comparison functions.

  • PyArray_ISPYTHON is removed as it is misleading, has no known use-cases, and is easy to replace.

  • PyArray_FieldNames is removed, as it is unclear what it would be useful for. It also has incorrect semantics in some possible use-cases.

  • PyArray_TypestrConvert is removed, since it seems a misnomer and unlikely to be used by anyone. If you know the size or are limited to few types, just use it explicitly, otherwise go via Python strings.

    (gh-25292)

  • PyDataType_GetDatetimeMetaData has been removed, it did not actually do anything since at least NumPy 1.7.

    (gh-25802)

PyArray_GetCastFunc was removed#

Note that custom legacy user dtypes can still provide a castfunc as their implementation, but any access to them is now removed. The reason for this is that NumPy never used these internally for many years. If you use simple numeric types, please just use C casts directly. In case you require an alternative, please let us know so we can create new API such as PyArray_CastBuffer() which could use old or new cast functions depending on the NumPy version.

(gh-25161)

New Features#

  • np.add was extended to work with unicode and bytes dtypes.

    (gh-24858)

A new bitwise_count function#

This new function counts the number of 1-bits in a number. bitwise_count works on all the numpy integer types and integer-like objects.

>>> a = np.array([2**i - 1 for i in range(16)])
>>> np.bitwise_count(a)
array([ 0,  1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12, 13, 14, 15],
      dtype=uint8)

(gh-19355)

macOS Accelerate support, including the ILP64#

Support for the updated Accelerate BLAS/LAPACK library, including ILP64 (64-bit integer) support, in macOS 13.3 has been added. This brings arm64 support, and significant performance improvements of up to 10x for commonly used linear algebra operations. When Accelerate is selected at build time, or if no explicit BLAS library selection is done, the 13.3+ version will automatically be used if available.

(gh-24053)

Binary wheels are also available. On macOS >=14.0, users who install NumPy from PyPI will get wheels built against Accelerate rather than OpenBLAS.

(gh-25255)

Option to use weights for quantile and percentile functions#

A weights keyword is now available for quantile, percentile, nanquantile and nanpercentile. Only method="inverted_cdf" supports weights.

(gh-24254)

Improved CPU optimization tracking#

A new tracer mechanism is available which enables tracking of the enabled targets for each optimized function (i.e., that uses hardware-specific SIMD instructions) in the NumPy library. With this enhancement, it becomes possible to precisely monitor the enabled CPU dispatch targets for the dispatched functions.

A new function named opt_func_info has been added to the new namespace numpy.lib.introspect, offering this tracing capability. This function allows you to retrieve information about the enabled targets based on function names and data type signatures.

(gh-24420)

A new Meson backend for f2py#

f2py in compile mode (i.e. f2py -c) now accepts the --backend meson option. This is the default option for Python >=3.12. For older Python versions, f2py will still default to --backend distutils.

To support this in realistic use-cases, in compile mode f2py takes a --dep flag one or many times which maps to dependency() calls in the meson backend, and does nothing in the distutils backend.

There are no changes for users of f2py only as a code generator, i.e. without -c.

(gh-24532)

bind(c) support for f2py#

Both functions and subroutines can be annotated with bind(c). f2py will handle both the correct type mapping, and preserve the unique label for other C interfaces.

Note: bind(c, name = 'routine_name_other_than_fortran_routine') is not honored by the f2py bindings by design, since bind(c) with the name is meant to guarantee only the same name in C and Fortran, not in Python and Fortran.

(gh-24555)

A new strict option for several testing functions#

The strict keyword is now available for assert_allclose, assert_equal, and assert_array_less. Setting strict=True will disable the broadcasting behaviour for scalars and ensure that input arrays have the same data type.

(gh-24680, gh-24770, gh-24775)

Add np.core.umath.find and np.core.umath.rfind UFuncs#

Add two find and rfind UFuncs that operate on unicode or byte strings and are used in np.char. They operate similar to str.find and str.rfind.

(gh-24868)

diagonal and trace for numpy.linalg#

numpy.linalg.diagonal and numpy.linalg.trace have been added, which are array API standard-compatible variants of numpy.diagonal and numpy.trace. They differ in the default axis selection which define 2-D sub-arrays.

(gh-24887)

New long and ulong dtypes#

numpy.long and numpy.ulong have been added as NumPy integers mapping to C’s long and unsigned long. Prior to NumPy 1.24, numpy.long was an alias to Python’s int.

(gh-24922)

svdvals for numpy.linalg#

numpy.linalg.svdvals has been added. It computes singular values for (a stack of) matrices. Executing np.svdvals(x) is the same as calling np.svd(x, compute_uv=False, hermitian=False). This function is compatible with the array API standard.

(gh-24940)

A new isdtype function#

numpy.isdtype was added to provide a canonical way to classify NumPy’s dtypes in compliance with the array API standard.

(gh-25054)

A new astype function#

numpy.astype was added to provide an array API standard-compatible alternative to the numpy.ndarray.astype method.

(gh-25079)

Array API compatible functions’ aliases#

13 aliases for existing functions were added to improve compatibility with the array API standard:

  • Trigonometry: acos, acosh, asin, asinh, atan, atanh, atan2.

  • Bitwise: bitwise_left_shift, bitwise_invert, bitwise_right_shift.

  • Misc: concat, permute_dims, pow.

  • In numpy.linalg: tensordot, matmul.

(gh-25086)

New unique_* functions#

The unique_all, unique_counts, unique_inverse, and unique_values functions have been added. They provide functionality of unique with different sets of flags. They are array API standard-compatible, and because the number of arrays they return does not depend on the values of input arguments, they are easier to target for JIT compilation.

(gh-25088)

Matrix transpose support for ndarrays#

NumPy now offers support for calculating the matrix transpose of an array (or stack of arrays). The matrix transpose is equivalent to swapping the last two axes of an array. Both np.ndarray and np.ma.MaskedArray now expose a .mT attribute, and there is a matching new numpy.matrix_transpose function.

(gh-23762)

Array API compatible functions for numpy.linalg#

Six new functions and two aliases were added to improve compatibility with the Array API standard for numpy.linalg:

A correction argument for var and std#

A correction argument was added to var and std, which is an array API standard compatible alternative to ddof. As both arguments serve a similar purpose, only one of them can be provided at the same time.

(gh-25169)

ndarray.device and ndarray.to_device#

An ndarray.device attribute and ndarray.to_device method were added to numpy.ndarray for array API standard compatibility.

Additionally, device keyword-only arguments were added to: asarray, arange, empty, empty_like, eye, full, full_like, linspace, ones, ones_like, zeros, and zeros_like.

For all these new arguments, only device="cpu" is supported.

(gh-25233)

StringDType has been added to NumPy#

We have added a new variable-width UTF-8 encoded string data type, implementing a “NumPy array of Python strings”, including support for a user-provided missing data sentinel. It is intended as a drop-in replacement for arrays of Python strings and missing data sentinels using the object dtype. See NEP 55 and the documentation for more details.

(gh-25347)

New keywords for cholesky and pinv#

The upper and rtol keywords were added to numpy.linalg.cholesky and numpy.linalg.pinv, respectively, to improve array API standard compatibility.

For pinv, if neither rcond nor rtol is specified, the rcond’s default is used. We plan to deprecate and remove rcond in the future.

(gh-25388)

New keywords for sort, argsort and linalg.matrix_rank#

New keyword parameters were added to improve array API standard compatibility:

(gh-25437)

New numpy.strings namespace for string ufuncs#

NumPy now implements some string operations as ufuncs. The old np.char namespace is still available, and where possible the string manipulation functions in that namespace have been updated to use the new ufuncs, substantially improving their performance.

Where possible, we suggest updating code to use functions in np.strings instead of np.char. In the future we may deprecate np.char in favor of np.strings.

(gh-25463)

numpy.fft support for different precisions and in-place calculations#

The various FFT routines in numpy.fft now do their calculations natively in float, double, or long double precision, depending on the input precision, instead of always calculating in double precision. Hence, the calculation will now be less precise for single and more precise for long double precision. The data type of the output array will now be adjusted accordingly.

Furthermore, all FFT routines have gained an out argument that can be used for in-place calculations.

(gh-25536)

configtool and pkg-config support#

A new numpy-config CLI script is available that can be queried for the NumPy version and for compile flags needed to use the NumPy C API. This will allow build systems to better support the use of NumPy as a dependency. Also, a numpy.pc pkg-config file is now included with Numpy. In order to find its location for use with PKG_CONFIG_PATH, use numpy-config --pkgconfigdir.

(gh-25730)

Array API standard support in the main namespace#

The main numpy namespace now supports the array API standard. See Array API standard compatibility for details.

(gh-25911)

Improvements#

  • Strings are now supported by any, all, and the logical ufuncs.

    (gh-25651)

Integer sequences as the shape argument for memmap#

numpy.memmap can now be created with any integer sequence as the shape argument, such as a list or numpy array of integers. Previously, only the types of tuple and int could be used without raising an error.

(gh-23729)

errstate is now faster and context safe#

The numpy.errstate context manager/decorator is now faster and safer. Previously, it was not context safe and had (rare) issues with thread-safety.

(gh-23936)

AArch64 quicksort speed improved by using Highway’s VQSort#

The first introduction of the Google Highway library, using VQSort on AArch64. Execution time is improved by up to 16x in some cases, see the PR for benchmark results. Extensions to other platforms will be done in the future.

(gh-24018)

Complex types - underlying C type changes#

  • The underlying C types for all of NumPy’s complex types have been changed to use C99 complex types.

  • While this change does not affect the memory layout of complex types, it changes the API to be used to directly retrieve or write the real or complex part of the complex number, since direct field access (as in c.real or c.imag) is no longer an option. You can now use utilities provided in numpy/npy_math.h to do these operations, like this:

    npy_cdouble c;
    npy_csetreal(&c, 1.0);
    npy_csetimag(&c, 0.0);
    printf("%d + %di\n", npy_creal(c), npy_cimag(c));
    
  • To ease cross-version compatibility, equivalent macros and a compatibility layer have been added which can be used by downstream packages to continue to support both NumPy 1.x and 2.x. See Support for complex numbers for more info.

  • numpy/npy_common.h now includes complex.h, which means that complex is now a reserved keyword.

(gh-24085)

iso_c_binding support and improved common blocks for f2py#

Previously, users would have to define their own custom f2cmap file to use type mappings defined by the Fortran2003 iso_c_binding intrinsic module. These type maps are now natively supported by f2py

(gh-24555)

f2py now handles common blocks which have kind specifications from modules. This further expands the usability of intrinsics like iso_fortran_env and iso_c_binding.

(gh-25186)

Call str automatically on third argument to functions like assert_equal#

The third argument to functions like assert_equal now has str called on it automatically. This way it mimics the built-in assert statement, where assert_equal(a, b, obj) works like assert a == b, obj.

(gh-24877)

Support for array-like atol/rtol in isclose, allclose#

The keywords atol and rtol in isclose and allclose now accept both scalars and arrays. An array, if given, must broadcast to the shapes of the first two array arguments.

(gh-24878)

Consistent failure messages in test functions#

Previously, some numpy.testing assertions printed messages that referred to the actual and desired results as x and y. Now, these values are consistently referred to as ACTUAL and DESIRED.

(gh-24931)

n-D FFT transforms allow s[i] == -1#

The fftn, ifftn, rfftn, irfftn, fft2, ifft2, rfft2 and irfft2 functions now use the whole input array along the axis i if s[i] == -1, in line with the array API standard.

(gh-25495)

Guard PyArrayScalar_VAL and PyUnicodeScalarObject for the limited API#

PyUnicodeScalarObject holds a PyUnicodeObject, which is not available when using Py_LIMITED_API. Add guards to hide it and consequently also make the PyArrayScalar_VAL macro hidden.

(gh-25531)

Changes#

  • np.gradient() now returns a tuple rather than a list making the return value immutable.

    (gh-23861)

  • Being fully context and thread-safe, np.errstate can only be entered once now.

  • np.setbufsize is now tied to np.errstate(): leaving an np.errstate context will also reset the bufsize.

    (gh-23936)

  • A new public np.lib.array_utils submodule has been introduced and it currently contains three functions: byte_bounds (moved from np.lib.utils), normalize_axis_tuple and normalize_axis_index.

    (gh-24540)

  • Introduce numpy.bool as the new canonical name for NumPy’s boolean dtype, and make numpy.bool_ an alias to it. Note that until NumPy 1.24, np.bool was an alias to Python’s builtin bool. The new name helps with array API standard compatibility and is a more intuitive name.

    (gh-25080)

  • The dtype.flags value was previously stored as a signed integer. This means that the aligned dtype struct flag lead to negative flags being set (-128 rather than 128). This flag is now stored unsigned (positive). Code which checks flags manually may need to adapt. This may include code compiled with Cython 0.29.x.

    (gh-25816)

Representation of NumPy scalars changed#

As per NEP 51, the scalar representation has been updated to include the type information to avoid confusion with Python scalars.

Scalars are now printed as np.float64(3.0) rather than just 3.0. This may disrupt workflows that store representations of numbers (e.g., to files) making it harder to read them. They should be stored as explicit strings, for example by using str() or f"{scalar!s}". For the time being, affected users can use np.set_printoptions(legacy="1.25") to get the old behavior (with possibly a few exceptions). Documentation of downstream projects may require larger updates, if code snippets are tested. We are working on tooling for doctest-plus to facilitate updates.

(gh-22449)

Truthiness of NumPy strings changed#

NumPy strings previously were inconsistent about how they defined if the string is True or False and the definition did not match the one used by Python. Strings are now considered True when they are non-empty and False when they are empty. This changes the following distinct cases:

  • Casts from string to boolean were previously roughly equivalent to string_array.astype(np.int64).astype(bool), meaning that only valid integers could be cast. Now a string of "0" will be considered True since it is not empty. If you need the old behavior, you may use the above step (casting to integer first) or string_array == "0" (if the input is only ever 0 or 1). To get the new result on old NumPy versions use string_array != "".

  • np.nonzero(string_array) previously ignored whitespace so that a string only containing whitespace was considered False. Whitespace is now considered True.

This change does not affect np.loadtxt, np.fromstring, or np.genfromtxt. The first two still use the integer definition, while genfromtxt continues to match for "true" (ignoring case). However, if np.bool_ is used as a converter the result will change.

The change does affect np.fromregex as it uses direct assignments.

(gh-23871)

A mean keyword was added to var and std function#

Often when the standard deviation is needed the mean is also needed. The same holds for the variance and the mean. Until now the mean is then calculated twice, the change introduced here for the var and std functions allows for passing in a precalculated mean as an keyword argument. See the docstrings for details and an example illustrating the speed-up.

(gh-24126)

Remove datetime64 deprecation warning when constructing with timezone#

The numpy.datetime64 method now issues a UserWarning rather than a DeprecationWarning whenever a timezone is included in the datetime string that is provided.

(gh-24193)

Default integer dtype is now 64-bit on 64-bit Windows#

The default NumPy integer is now 64-bit on all 64-bit systems as the historic 32-bit default on Windows was a common source of issues. Most users should not notice this. The main issues may occur with code interfacing with libraries written in a compiled language like C. For more information see Windows default integer.

(gh-24224)

Renamed numpy.core to numpy._core#

Accessing numpy.core now emits a DeprecationWarning. In practice we have found that most downstream usage of numpy.core was to access functionality that is available in the main numpy namespace. If for some reason you are using functionality in numpy.core that is not available in the main numpy namespace, this means you are likely using private NumPy internals. You can still access these internals via numpy._core without a deprecation warning but we do not provide any backward compatibility guarantees for NumPy internals. Please open an issue if you think a mistake was made and something needs to be made public.

(gh-24634)

The “relaxed strides” debug build option, which was previously enabled through the NPY_RELAXED_STRIDES_DEBUG environment variable or the -Drelaxed-strides-debug config-settings flag has been removed.

(gh-24717)

Redefinition of np.intp/np.uintp (almost never a change)#

Due to the actual use of these types almost always matching the use of size_t/Py_ssize_t this is now the definition in C. Previously, it matched intptr_t and uintptr_t which would often have been subtly incorrect. This has no effect on the vast majority of machines since the size of these types only differ on extremely niche platforms.

However, it means that:

  • Pointers may not necessarily fit into an intp typed array anymore. The p and P character codes can still be used, however.

  • Creating intptr_t or uintptr_t typed arrays in C remains possible in a cross-platform way via PyArray_DescrFromType('p').

  • The new character codes nN were introduced.

  • It is now correct to use the Python C-API functions when parsing to npy_intp typed arguments.

(gh-24888)

numpy.fft.helper made private#

numpy.fft.helper was renamed to numpy.fft._helper to indicate that it is a private submodule. All public functions exported by it should be accessed from numpy.fft.

(gh-24945)

numpy.linalg.linalg made private#

numpy.linalg.linalg was renamed to numpy.linalg._linalg to indicate that it is a private submodule. All public functions exported by it should be accessed from numpy.linalg.

(gh-24946)

Out-of-bound axis not the same as axis=None#

In some cases axis=32 or for concatenate any large value was the same as axis=None. Except for concatenate this was deprecate. Any out of bound axis value will now error, make sure to use axis=None.

(gh-25149)

New copy keyword meaning for array and asarray constructors#

Now numpy.array and numpy.asarray support three values for copy parameter:

  • None - A copy will only be made if it is necessary.

  • True - Always make a copy.

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

The meaning of False changed as it now raises an exception if a copy is needed.

(gh-25168)

The __array__ special method now takes a copy keyword argument.#

NumPy will pass copy to the __array__ special method in situations where it would be set to a non-default value (e.g. in a call to np.asarray(some_object, copy=False)). Currently, if an unexpected keyword argument error is raised after this, NumPy will print a warning and re-try without the copy keyword argument. Implementations of objects implementing the __array__ protocol should accept a copy keyword argument with the same meaning as when passed to numpy.array or numpy.asarray.

(gh-25168)

Cleanup of initialization of numpy.dtype with strings with commas#

The interpretation of strings with commas is changed slightly, in that a trailing comma will now always create a structured dtype. E.g., where previously np.dtype("i") and np.dtype("i,") were treated as identical, now np.dtype("i,") will create a structured dtype, with a single field. This is analogous to np.dtype("i,i") creating a structured dtype with two fields, and makes the behaviour consistent with that expected of tuples.

At the same time, the use of single number surrounded by parenthesis to indicate a sub-array shape, like in np.dtype("(2)i,"), is deprecated. Instead; one should use np.dtype("(2,)i") or np.dtype("2i"). Eventually, using a number in parentheses will raise an exception, like is the case for initializations without a comma, like np.dtype("(2)i").

(gh-25434)

Change in how complex sign is calculated#

Following the array API standard, the complex sign is now calculated as z / |z| (instead of the rather less logical case where the sign of the real part was taken, unless the real part was zero, in which case the sign of the imaginary part was returned). Like for real numbers, zero is returned if z==0.

(gh-25441)

Return types of functions that returned a list of arrays#

Functions that returned a list of ndarrays have been changed to return a tuple of ndarrays instead. Returning tuples consistently whenever a sequence of arrays is returned makes it easier for JIT compilers like Numba, as well as for static type checkers in some cases, to support these functions. Changed functions are: atleast_1d, atleast_2d, atleast_3d, broadcast_arrays, meshgrid, ogrid, histogramdd.

np.unique return_inverse shape for multi-dimensional inputs#

When multi-dimensional inputs are passed to np.unique with return_inverse=True, the unique_inverse output is now shaped such that the input can be reconstructed directly using np.take(unique, unique_inverse) when axis=None, and np.take_along_axis(unique, unique_inverse, axis=axis) otherwise.

(gh-25553, gh-25570)

any and all return booleans for object arrays#

The any and all functions and methods now return booleans also for object arrays. Previously, they did a reduction which behaved like the Python or and and operators which evaluates to one of the arguments. You can use np.logical_or.reduce and np.logical_and.reduce to achieve the previous behavior.

(gh-25712)

Content from release note snippets in doc/release/upcoming_changes:

Expired deprecations#