NumPy 2.4.0 Release Notes
The NumPy 2.4.0 release continues the work to improve free threaded Python
support, user dtypes implementation, and annotations. There are many expired
deprecations and bug fixes as well.
This release supports Python versions 3.11-3.14
Highlights
Apart from annotations and same_value kwarg, the 2.4 highlights are mostly
of interest to downstream developers. They should help in implementing new user
dtypes.
- Many annotation improvements. In particular, runtime signature introspection.
- New
casting kwarg 'same_value' for casting by value.
- New
PyUFunc_AddLoopsFromSpec function that can be used to add user sort
loops using the ArrayMethod API.
- New
__numpy_dtype__ protocol.
Deprecations
Setting the strides attribute is deprecated
Setting the strides attribute is now deprecated since mutating
an array is unsafe if an array is shared, especially by multiple
threads. As an alternative, you can create a new view (no copy) via:
np.lib.stride_tricks.strided_window_view if applicable,np.lib.stride_tricks.as_strided for the general case,- or the
np.ndarray constructor (buffer is the original array) for a
light-weight version.
(gh-28925)
Positional out argument to np.maximum, np.minimum is deprecated
Passing the output array out positionally to numpy.maximum and
numpy.minimum is deprecated. For example, np.maximum(a, b, c) will emit
a deprecation warning, since c is treated as the output buffer rather than
a third input.
Always pass the output with the keyword form, e.g. np.maximum(a, b, out=c).
This makes intent clear and simplifies type annotations.
(gh-29052)
align= must be passed as boolean to np.dtype()
When creating a new dtype a VisibleDeprecationWarning will be given if
align= is not a boolean. This is mainly to prevent accidentally passing a
subarray align flag where it has no effect, such as np.dtype("f8", 3)
instead of np.dtype(("f8", 3)). We strongly suggest to always pass
align= as a keyword argument.
(gh-29301)
Assertion and warning control utilities are deprecated
np.testing.assert_warns and np.testing.suppress_warnings are
deprecated. Use warnings.catch_warnings, warnings.filterwarnings,
pytest.warns, or pytest.filterwarnings instead.
(gh-29550)
np.fix is pending deprecation
The numpy.fix function will be deprecated in a future release. It is
recommended to use numpy.trunc instead, as it provides the same
functionality of truncating decimal values to their integer parts. Static type
checkers might already report a warning for the use of numpy.fix.
(gh-30168)
in-place modification of ndarray.shape is pending deprecation
Setting the ndarray.shape attribute directly will be deprecated in a future
release. Instead of modifying the shape in place, it is recommended to use the
numpy.reshape function. Static type checkers might already report a
warning for assignments to ndarray.shape.
(gh-30282)
Deprecation of numpy.lib.user_array.container
The numpy.lib.user_array.container class is deprecated and will be removed
in a future version.
(gh-30284)
Expired deprecations
Removed deprecated MachAr runtime discovery mechanism.
(gh-29836)
Raise TypeError on attempt to convert array with ndim > 0 to scalar
Conversion of an array with ndim > 0 to a scalar was deprecated in NumPy
1.25. Now, attempting to do so raises TypeError. Ensure you extract a
single element from your array before performing this operation.
(gh-29841)
Removed numpy.linalg.linalg and numpy.fft.helper
The following were deprecated in NumPy 2.0 and have been moved to private
modules:
numpy.linalg.linalg
Use numpy.linalg instead.numpy.fft.helper
Use numpy.fft instead.
(gh-29909)
Removed interpolation parameter from quantile and percentile functions
The interpolation parameter was deprecated in NumPy 1.22.0 and has been
removed from the following functions:
numpy.percentilenumpy.nanpercentilenumpy.quantilenumpy.nanquantile
Use the method parameter instead.
(gh-29973)
Removed numpy.in1d
numpy.in1d has been deprecated since NumPy 2.0 and is now removed in favor of numpy.isin.
(gh-29978)
Removed numpy.ndindex.ndincr()
The ndindex.ndincr() method has been deprecated since NumPy 1.20 and is now
removed; use next(ndindex) instead.
(gh-29980)
Removed fix_imports parameter from numpy.save
The fix_imports parameter was deprecated in NumPy 2.1.0 and is now removed.
This flag has been ignored since NumPy 1.17 and was only needed to support
loading files in Python 2 that were written in Python 3.
(gh-29984)
Removal of four undocumented ndarray.ctypes methods
Four undocumented methods of the ndarray.ctypes object have been removed:
_ctypes.get_data() (use _ctypes.data instead)_ctypes.get_shape() (use _ctypes.shape instead)_ctypes.get_strides() (use _ctypes.strides instead)_ctypes.get_as_parameter() (use _ctypes._as_parameter_ instead)
These methods have been deprecated since NumPy 1.21.
(gh-29986)
Removed newshape parameter from numpy.reshape
The newshape parameter was deprecated in NumPy 2.1.0 and has been
removed from numpy.reshape. Pass it positionally or use shape=
on newer NumPy versions.
(gh-29994)
Removal of deprecated functions and arguments
The following long-deprecated APIs have been removed:
numpy.trapz --- deprecated since NumPy 2.0 (2023-08-18). Use numpy.trapezoid or
scipy.integrate functions instead.disp function --- deprecated from 2.0 release and no longer functional. Use
your own printing function instead.bias and ddof arguments in numpy.corrcoef --- these had no effect
since NumPy 1.10.
(gh-29997)
Removed delimitor parameter from numpy.ma.mrecords.fromtextfile()
The delimitor parameter was deprecated in NumPy 1.22.0 and has been
removed from numpy.ma.mrecords.fromtextfile(). Use delimiter instead.
(gh-30021)
numpy.array2string and numpy.sum deprecations finalized
The following long-deprecated APIs have been removed or converted to errors:
- The
style parameter has been removed from numpy.array2string.
This argument had no effect since Numpy 1.14.0. Any arguments following
it, such as formatter have now been made keyword-only.
- Calling
np.sum(generator) directly on a generator object now raises a
TypeError. This behavior was deprecated in NumPy 1.15.0. Use
np.sum(np.fromiter(generator)) or the python sum builtin instead.
(gh-30068)
Compatibility notes
-
NumPy's C extension modules have begun to use multi-phase initialisation, as
defined by PEP 489. As part of this, a new explicit check has been added that
each such module is only imported once per Python process. This comes with
the side-effect that deleting numpy from sys.modules and re-importing
it will now fail with an ImportError. This has always been unsafe, with
unexpected side-effects, though did not previously raise an error.
(gh-29030)
-
numpy.round now always returns a copy. Previously, it returned a view
for integer inputs for decimals >= 0 and a copy in all other cases.
This change brings round in line with ceil, floor and trunc.
(gh-29137)
-
Type-checkers will no longer accept calls to numpy.arange with
start as a keyword argument. This was done for compatibility with
the Array API standard. At runtime it is still possible to use
numpy.arange with start as a keyword argument.
(gh-30147)
-
The Macro NPY_ALIGNMENT_REQUIRED has been removed The macro was defined in
the npy_cpu.h file, so might be regarded as semi public. As it turns out,
with modern compilers and hardware it is almost always the case that
alignment is required, so numpy no longer uses the macro. It is unlikely
anyone uses it, but you might want to compile with the -Wundef flag or
equivalent to be sure.
(gh-29094)
C API changes
The NPY_SORTKIND enum has been enhanced with new variables
This is of interest if you are using PyArray_Sort or PyArray_ArgSort.
We have changed the semantics of the old names in the NPY_SORTKIND enum and
added new ones. The changes are backward compatible, and no recompilation is
needed. The new names of interest are:
NPY_SORT_DEFAULT -- default sort (same value as NPY_QUICKSORT)NPY_SORT_STABLE -- the sort must be stable (same value as NPY_MERGESORT)NPY_SORT_DESCENDING -- the sort must be descending
The semantic change is that NPY_HEAPSORT is mapped to NPY_QUICKSORT when used.
Note that NPY_SORT_DESCENDING is not yet implemented.
(gh-29642)
New NPY_DT_get_constant slot for DType constant retrieval
A new slot NPY_DT_get_constant has been added to the DType API, allowing
dtype implementations to provide constant values such as machine limits and
special values. The slot function has the signature:
int get_constant(PyArray_Descr *descr, int constant_id, void *ptr)
It returns 1 on success, 0 if the constant is not available, or -1 on error.
The function is always called with the GIL held and may write to unaligned memory.
Integer constants (marked with the 1 << 16 bit) return npy_intp values,
while floating-point constants return values of the dtype's native type.
Implementing this can be used by user DTypes to provide numpy.finfo values.
(gh-29836)
A new PyUFunc_AddLoopsFromSpecs convenience function has been added to the C API.
This function allows adding multiple ufunc loops from their specs in one call
using a NULL-terminated array of PyUFunc_LoopSlot structs. It allows
registering sorting and argsorting loops using the new ArrayMethod API.
(gh-29900)
New Features
'same_value' for casting by value
The casting kwarg now has a 'same_value' option that checks the actual
values can be round-trip cast without changing value. Currently it is only
implemented in ndarray.astype. This will raise a ValueError if any of the
values in the array would change as a result of the cast, including rounding of
floats or overflowing of ints.
(gh-29129)
StringDType fill_value support in numpy.ma.MaskedArray
Masked arrays now accept and preserve a Python str as their fill_value
when using the variable‑width StringDType (kind 'T'), including through
slicing and views. The default is 'N/A' and may be overridden by any valid
string. This fixes issue gh‑29421
and was implemented in pull request gh‑29423.
(gh-29423)
ndmax option for numpy.array
The ndmax option is now available for numpy.array.
It explicitly limits the maximum number of dimensions created from nested sequences.
This is particularly useful when creating arrays of list-like objects with dtype=object.
By default, NumPy recurses through all nesting levels to create the highest possible
dimensional array, but this behavior may not be desired when the intent is to preserve
nested structures as objects. The ndmax parameter provides explicit control over
this recursion depth.
# Default behavior: Creates a 2D array
>>> a = np.array([[1, 2], [3, 4]], dtype=object)
>>> a
array([[1, 2],
[3, 4]], dtype=object)
>>> a.shape
(2, 2)
# With ndmax=1: Creates a 1D array
>>> b = np.array([[1, 2], [3, 4]], dtype=object, ndmax=1)
>>> b
array([list([1, 2]), list([3, 4])], dtype=object)
>>> b.shape
(2,)
(gh-29569)
Warning emitted when using where without out
Ufuncs called with a where mask and without an out positional or kwarg will
now emit a warning. This usage tends to trip up users who expect some value in
output locations where the mask is False (the ufunc will not touch those
locations). The warning can be suppressed by using out=None.
(gh-29813)
DType sorting and argsorting supports the ArrayMethod API
User-defined dtypes can now implement custom sorting and argsorting using the
ArrayMethod API. This mechanism can be used in place of the
PyArray_ArrFuncs slots which may be deprecated in the future.
The sorting and argsorting methods are registered by passing the arraymethod
specs that implement the operations to the new PyUFunc_AddLoopsFromSpecs
function. See the ArrayMethod API documentation for details.
(gh-29900)
New __numpy_dtype__ protocol
NumPy now has a new __numpy_dtype__ protocol. NumPy will check
for this attribute when converting to a NumPy dtype via np.dtype(obj)
or any dtype= argument.
Downstream projects are encouraged to implement this for all dtype like
objects which may previously have used a .dtype attribute that returned
a NumPy dtype.
We expect to deprecate .dtype in the future to prevent interpreting
array-like objects with a .dtype attribute as a dtype.
If you wish you can implement __numpy_dtype__ to ensure an earlier
warning or error (.dtype is ignored if this is found).
(gh-30179)
Improvements
Fix flatiter indexing edge cases
The flatiter object now shares the same index preparation logic as
ndarray, ensuring consistent behavior and fixing several issues where
invalid indices were previously accepted or misinterpreted.
Key fixes and improvements:
-
Stricter index validation
- Boolean non-array indices like
arr.flat[[True, True]] were
incorrectly treated as arr.flat[np.array([1, 1], dtype=int)].
They now raise an index error. Note that indices that match the
iterator's shape are expected to not raise in the future and be
handled as regular boolean indices. Use np.asarray(<index>) if
you want to match that behavior.
- Float non-array indices were also cast to integer and incorrectly
treated as
arr.flat[np.array([1.0, 1.0], dtype=int)]. This is now
deprecated and will be removed in a future version.
- 0-dimensional boolean indices like
arr.flat[True] are also
deprecated and will be removed in a future version.
-
Consistent error types:
Certain invalid flatiter indices that previously raised ValueError
now correctly raise IndexError, aligning with ndarray behavior.
-
Improved error messages:
The error message for unsupported index operations now provides more
specific details, including explicitly listing the valid index types,
instead of the generic IndexError: unsupported index operation.
(gh-28590)
Improved error handling in np.quantile
[np.quantile]{.title-ref} now raises errors if:
- All weights are zero
- At least one weight is
np.nan
- At least one weight is
np.inf
(gh-28595)
Improved error message for assert_array_compare
The error message generated by assert_array_compare which is used by functions
like assert_allclose, assert_array_less etc. now also includes information
about the indices at which the assertion fails.
(gh-29112)
Show unit information in __repr__ for datetime64("NaT")
When a datetime64 object is "Not a Time" (NaT), its __repr__ method now
includes the time unit of the datetime64 type. This makes it consistent with
the behavior of a timedelta64 object.
(gh-29396)
Performance increase for scalar calculations
The speed of calculations on scalars has been improved by about a factor 6 for
ufuncs that take only one input (like np.sin(scalar)), reducing the speed
difference from their math equivalents from a factor 19 to 3 (the speed
for arrays is left unchanged).
(gh-29819)
numpy.finfo Refactor
The numpy.finfo class has been completely refactored to obtain floating-point
constants directly from C compiler macros rather than deriving them at runtime.
This provides better accuracy, platform compatibility and corrected
several attribute calculations:
- Constants like
eps, min, max, smallest_normal, and
smallest_subnormal now come directly from standard C macros (FLT_EPSILON,
DBL_MIN, etc.), ensuring platform-correct values.
- The deprecated
MachAr runtime discovery mechanism has been removed.
- Derived attributes have been corrected to match standard definitions:
machep and negep now use int(log2(eps)); nexp accounts for
all exponent patterns; nmant excludes the implicit bit; and minexp
follows the C standard definition.
- longdouble constants, Specifically
smallest_normal now follows the
C standard definitions as per respecitive platform.
- Special handling added for PowerPC's IBM double-double format.
- New test suite added in
test_finfo.py to validate all
finfo properties against expected machine arithmetic values for
float16, float32, and float64 types.
(gh-29836)
Multiple axes are now supported in numpy.trim_zeros
The axis argument of numpy.trim_zeros now accepts a sequence; for example
np.trim_zeros(x, axis=(0, 1)) will trim the zeros from a multi-dimensional
array x along axes 0 and 1. This fixes issue
gh‑29945 and was implemented
in pull request gh‑29947.
(gh-29947)
Runtime signature introspection support has been significantly improved
Many NumPy functions, classes, and methods that previously raised
ValueError when passed to inspect.signature() now return meaningful
signatures. This improves support for runtime type checking, IDE autocomplete,
documentation generation, and runtime introspection capabilities across the
NumPy API.
Over three hundred classes and functions have been updated in total, including,
but not limited to, core classes such as ndarray, generic, dtype,
ufunc, broadcast, nditer, etc., most methods of ndarray and
scalar types, array constructor functions (array, empty, arange,
fromiter, etc.), all ufuncs, and many other commonly used functions,
including dot, concat, where, bincount, can_cast, and
numerous others.
(gh-30208)
Performance improvements and changes
Performance improvements to np.unique for string dtypes
The hash-based algorithm for unique extraction provides an order-of-magnitude
speedup on large string arrays. In an internal benchmark with about 1 billion
string elements, the hash-based np.unique completed in roughly 33.5 seconds,
compared to 498 seconds with the sort-based method -- about 15× faster for
unsorted unique operations on strings. This improvement greatly reduces the
time to find unique values in very large string datasets.
(gh-28767)
Rewrite of np.ndindex using itertools.product
The numpy.ndindex function now uses itertools.product internally,
providing significant improvements in performance for large iteration spaces,
while maintaining the original behavior and interface. For example, for an
array of shape (50, 60, 90) the NumPy ndindex benchmark improves
performance by a factor 5.2.
(gh-29165)
Performance improvements to np.unique for complex dtypes
The hash-based algorithm for unique extraction now also supports
complex dtypes, offering noticeable performance gains.
In our benchmarks on complex128 arrays with 200,000 elements,
the hash-based approach was about 1.4--1.5× faster
than the sort-based baseline when there were 20% of unique values,
and about 5× faster when there were 0.2% of unique values.
(gh-29537)
Changes
-
Multiplication between a string and integer now raises OverflowError instead
of MemoryError if the result of the multiplication would create a string that
is too large to be represented. This follows Python's behavior.
(gh-29060)
-
The accuracy of np.quantile and np.percentile for 16- and 32-bit
floating point input data has been improved.
(gh-29105)
unique_values for string dtypes may return unsorted data
np.unique now supports hash‐based duplicate removal for string dtypes.
This enhancement extends the hash-table algorithm to byte strings ('S'),
Unicode strings ('U'), and the experimental string dtype ('T', StringDType).
As a result, calling np.unique() on an array of strings will use
the faster hash-based method to obtain unique values.
Note that this hash-based method does not guarantee that the returned unique values will be sorted.
This also works for StringDType arrays containing None (missing values)
when using equal_nan=True (treating missing values as equal).
(gh-28767)
Modulate dispatched x86 CPU features
IMPORTANT: The default setting for cpu-baseline on x86 has been raised
to x86-64-v2 microarchitecture. This can be changed to none during build
time to support older CPUs, though SIMD optimizations for pre-2009 processors
are no longer maintained.
NumPy has reorganized x86 CPU features into microarchitecture-based groups
instead of individual features, aligning with Linux distribution standards and
Google Highway requirements.
Key changes:
- Replaced individual x86 features with microarchitecture levels:
X86_V2,
X86_V3, and X86_V4
- Raised the baseline to
X86_V2
- Improved
- operator behavior to properly exclude successor features that
imply the excluded feature
- Added meson redirections for removed feature names to maintain backward
compatibility
- Removed compiler compatibility workarounds for partial feature support (e.g.,
AVX512 without mask operations)
- Removed legacy AMD features (XOP, FMA4) and discontinued Intel Xeon Phi
support
New Feature Group Hierarchy:
Name Implies Includes
X86_V2 SSE SSE2 SSE3 SSSE3 SSE4_1 SSE4_2 POPCNT CX16 LAHF
X86_V3 X86_V2 AVX AVX2 FMA3 BMI BMI2 LZCNT F16C MOVBE
X86_V4 X86_V3 AVX512F AVX512CD AVX512VL AVX512BW AVX512DQ
AVX512_ICL X86_V4 AVX512VBMI AVX512VBMI2 AVX512VNNI AVX512BITALG AVX512VPOPCNTDQ AVX512IFMA VAES GFNI VPCLMULQDQ
AVX512_SPR AVX512_ICL AVX512FP16
These groups correspond to CPU generations:
X86_V2: x86-64-v2 microarchitectures (CPUs since 2009)X86_V3: x86-64-v3 microarchitectures (CPUs since 2015)X86_V4: x86-64-v4 microarchitectures (AVX-512 capable CPUs)AVX512_ICL: Intel Ice Lake and similar CPUsAVX512_SPR: Intel Sapphire Rapids and newer CPUs
On 32-bit x86, cx16 is excluded from X86_V2.
Documentation has been updated with details on using these new feature groups
with the current meson build system.
(gh-28896)
Fix bug in matmul for non-contiguous out kwarg parameter
In some cases, if out was non-contiguous, np.matmul would cause memory
corruption or a c-level assert. This was new to v2.3.0 and fixed in v2.3.1.
(gh-29179)
__array_interface__ with NULL pointer changed
The array interface now accepts NULL pointers (NumPy will do its own dummy
allocation, though). Previously, these incorrectly triggered an undocumented
scalar path. In the unlikely event that the scalar path was actually desired,
you can (for now) achieve the previous behavior via the correct scalar path by
not providing a data field at all.
(gh-29338)
unique_values for complex dtypes may return unsorted data
np.unique now supports hash‐based duplicate removal for complex dtypes. This
enhancement extends the hash‐table algorithm to all complex types ('c'), and
their extended precision variants. The hash‐based method provides faster
extraction of unique values but does not guarantee that the result will be
sorted.
(gh-29537)
Sorting kind='heapsort' now maps to kind='quicksort'
It is unlikely that this change will be noticed, but if you do see a change in
execution time or unstable argsort order, that is likely the cause. Please let
us know if there is a performance regression. Congratulate us if it is improved
:)
(gh-29642)
numpy.typing.DTypeLike no longer accepts None
The type alias numpy.typing.DTypeLike no longer accepts None. Instead of
dtype: DTypeLike = None
it should now be
dtype: DTypeLike | None = None
instead.
(gh-29739)
The npymath and npyrandom libraries now have a .lib rather than a
.a file extension on win-arm64, for compatibility for building with MSVC
and setuptools. Please note that using these static libraries is
discouraged and for existing projects using it, it's best to use it with a
matching compiler toolchain, which is clang-cl on Windows on Arm.
(gh-29750)