numpy - PyPI Package Compare versions

+91

.github/workflows/linux-ibm.yml

		name: Linux IBM tests

		on:
		pull_request:
		branches:
		- main
		- maintenance/**
		paths-ignore:
		- '**.pyi'
		- '**.md'
		- '**.rst'
		workflow_dispatch:

		permissions:
		contents: read

		concurrency:
		group: ${{ github.workflow }}-${{ github.head_ref \|\| github.run_id }}
		cancel-in-progress: true

		jobs:
		native_ibm:
		# These jobs runs only in the main NumPy repository.
		# It requires a native ppc64le and s390x GHA runners, which are not available on forks.
		# For more details, see: https://github.com/numpy/numpy/issues/29125
		if: github.repository == 'numpy/numpy'
		runs-on: ${{ matrix.config.runner }}

		strategy:
		fail-fast: false
		matrix:
		config:
		- name: "ppc64le/gcc - baseline(default)"
		args: "-Dallow-noblas=false"
		runner: ubuntu-24.04-ppc64le-p10
		compiler: "gcc"
		- name: "ppc64le/clang - baseline(default)"
		args: "-Dallow-noblas=false"
		runner: ubuntu-24.04-ppc64le-p10
		compiler: "clang"
		- name: "s390x/gcc - baseline(default)"
		args: "-Dallow-noblas=false"
		runner: ubuntu-24.04-s390x
		compiler: "gcc"
		- name: "s390x/clang - baseline(default)"
		args: "-Dallow-noblas=false"
		runner: ubuntu-24.04-s390x
		compiler: "clang"
		- name: "s390x/gcc - baseline(Z15/VXE2)"
		args: "-Dallow-noblas=false -Dcpu-baseline=vxe2"
		runner: ubuntu-24.04-s390x
		compiler: "gcc"
		- name: "s390x/clang - baseline(Z15/VXE2)"
		args: "-Dallow-noblas=false -Dcpu-baseline=vxe2"
		runner: ubuntu-24.04-s390x
		compiler: "clang"

		name: "${{ matrix.config.name }}"
		steps:
		- uses: actions/checkout@1af3b93b6815bc44a9784bd300feb67ff0d1eeb3 # v6.0.0
		with:
		submodules: recursive
		fetch-tags: true

		- name: Install dependencies
		run: \|
		sudo apt update
		sudo apt install -y python3.12 python3-pip python3-dev ninja-build gfortran \
		build-essential libopenblas-dev liblapack-dev pkg-config
		pip install --upgrade pip
		pip install -r requirements/build_requirements.txt -r requirements/test_requirements.txt
		echo "/home/runner/.local/bin" >> $GITHUB_PATH

		- name: Install clang
		if: matrix.config.compiler == 'clang'
		run: \|
		sudo apt install -y clang-20
		echo CC=clang-20 >> $GITHUB_ENV
		echo CXX=clang++-20 >> $GITHUB_ENV

		- name: Meson Build
		run: \|
		spin build -- ${{ matrix.config.args }}

		- name: Meson Log
		if: always()
		run: cat build/meson-logs/meson-log.txt

		- name: Run Tests
		run: \|
		spin test -- --timeout=60 --durations=10

+59

doc/changelog/2.4.5-changelog.rst


		Contributors
		============

		A total of 17 people contributed to this release. People with a "+" by their
		names contributed a patch for the first time.

		* Aleksei Nikiforov
		* Ankit Ahlawat
		* Breno Favaretto +
		* Charles Harris
		* Igor Krivenko +
		* Ijtihed Kilani +
		* Joren Hammudoglu
		* Maarten Baert +
		* Matti Picus
		* Nathan Goldbaum
		* Praneeth Kodumagulla +
		* Ralf Gommers
		* RoomWithOutRoof +
		* Sebastian Berg
		* Warren Weckesser
		* Anarion Zuo +
		* div +

		Pull requests merged
		====================

		A total of 28 pull requests were merged for this release.

		* `#31093 <https://github.com/numpy/numpy/pull/31093>`__: MAINT: Prepare 2.4.x for further development
		* `#31182 <https://github.com/numpy/numpy/pull/31182>`__: TYP: fix ``np.shape`` assignability issue for python lists (#31171)
		* `#31197 <https://github.com/numpy/numpy/pull/31197>`__: ENH: Return rank 0 for empty matrices in matrix_rank (#30422)
		* `#31198 <https://github.com/numpy/numpy/pull/31198>`__: CI/BUG: add native jobs for s390x, fix bug in ``pack_inner``...
		* `#31199 <https://github.com/numpy/numpy/pull/31199>`__: BUG: f2py map complex_long_double to NPY_CLONGDOUBLE
		* `#31205 <https://github.com/numpy/numpy/pull/31205>`__: MAINT: f2py: Stop setting re._MAXCACHE to 50.
		* `#31206 <https://github.com/numpy/numpy/pull/31206>`__: BUG: fix heap buffer overflow in timedelta to string casts
		* `#31207 <https://github.com/numpy/numpy/pull/31207>`__: MAINT: Rename ppc64le and s390x workflow (#31121)
		* `#31208 <https://github.com/numpy/numpy/pull/31208>`__: BUG: Fix matvec/vecmat in-place aliasing (out=input produces...
		* `#31209 <https://github.com/numpy/numpy/pull/31209>`__: TYP: ``tile``\ : accept numpy scalars and arrays as second argument...
		* `#31211 <https://github.com/numpy/numpy/pull/31211>`__: DEP: Undo deprecation for np.dtype() signature used by old pickles...
		* `#31212 <https://github.com/numpy/numpy/pull/31212>`__: REV: Manual revert of float16 svml use (#31178)
		* `#31222 <https://github.com/numpy/numpy/pull/31222>`__: TYP: ``ix_`` fix for boolean and non-1d input (#31218)
		* `#31329 <https://github.com/numpy/numpy/pull/31329>`__: BUG: incorrect temp elision for new-style (NEP 43) user-defined...
		* `#31330 <https://github.com/numpy/numpy/pull/31330>`__: TYP: fix sliding_window_view axis parameter typing
		* `#31335 <https://github.com/numpy/numpy/pull/31335>`__: BUG: Prevent deadlock due to downstream importing NumPy in dlopen...
		* `#31336 <https://github.com/numpy/numpy/pull/31336>`__: BUG: Fix segfault in nditer.multi_index when __getitem__ raises...
		* `#31338 <https://github.com/numpy/numpy/pull/31338>`__: TYP: Fix ruff lint error
		* `#31357 <https://github.com/numpy/numpy/pull/31357>`__: BUG: fix memory leak in np.zeros when fill-zero loop raises (#31320)
		* `#31358 <https://github.com/numpy/numpy/pull/31358>`__: BUG: np.einsum() fails with a 0-dimensional out argument and...
		* `#31379 <https://github.com/numpy/numpy/pull/31379>`__: BUG: Fix signed overflow issue in npy_gcd for INT_MIN on s390x...
		* `#31383 <https://github.com/numpy/numpy/pull/31383>`__: CI: remove Cirrus CI FreeBSD job (#31380)
		* `#31390 <https://github.com/numpy/numpy/pull/31390>`__: BUILD: newer MKL uses so.3
		* `#31391 <https://github.com/numpy/numpy/pull/31391>`__: BLD/MAINT: improve support for Intel LLVM compilers
		* `#31401 <https://github.com/numpy/numpy/pull/31401>`__: BUG: Avoid UB in safe_[add,sub,mul] helpers (#31396)
		* `#31402 <https://github.com/numpy/numpy/pull/31402>`__: BUG: exclude __pycache__ directories from wheels (#31397)
		* `#31404 <https://github.com/numpy/numpy/pull/31404>`__: TYP: ``_NestedSequence`` type parameter default to work around...
		* `#31426 <https://github.com/numpy/numpy/pull/31426>`__: TYP: Fix ``DTypeLike`` runtime type-checker support (#31425)

+71

doc/source/release/2.4.5-notes.rst

		.. currentmodule:: numpy

		=========================
		NumPy 2.4.5 Release Notes
		=========================

		The NumPy 2.4.5 is a patch release that fixes bugs discovered after the 2.4.4
		release, has some typing improvements, and maintains infrastructure.

		This release supports Python versions 3.11-3.14


		Contributors
		============

		A total of 17 people contributed to this release. People with a "+" by their
		names contributed a patch for the first time.

		* Aleksei Nikiforov
		* Anarion Zuo +
		* Ankit Ahlawat
		* Breno Favaretto +
		* Charles Harris
		* Igor Krivenko +
		* Ijtihed Kilani +
		* Joren Hammudoglu
		* Maarten Baert +
		* Matti Picus
		* Nathan Goldbaum
		* Praneeth Kodumagulla +
		* Ralf Gommers
		* RoomWithOutRoof +
		* Sebastian Berg
		* Warren Weckesser
		* div +


		Pull requests merged
		====================

		A total of 28 pull requests were merged for this release.

		* `#31093 <https://github.com/numpy/numpy/pull/31093>`__: MAINT: Prepare 2.4.x for further development
		* `#31182 <https://github.com/numpy/numpy/pull/31182>`__: TYP: fix ``np.shape`` assignability issue for python lists (#31171)
		* `#31197 <https://github.com/numpy/numpy/pull/31197>`__: ENH: Return rank 0 for empty matrices in matrix_rank (#30422)
		* `#31198 <https://github.com/numpy/numpy/pull/31198>`__: CI/BUG: add native jobs for s390x, fix bug in ``pack_inner``...
		* `#31199 <https://github.com/numpy/numpy/pull/31199>`__: BUG: f2py map complex_long_double to NPY_CLONGDOUBLE
		* `#31205 <https://github.com/numpy/numpy/pull/31205>`__: MAINT: f2py: Stop setting re._MAXCACHE to 50.
		* `#31206 <https://github.com/numpy/numpy/pull/31206>`__: BUG: fix heap buffer overflow in timedelta to string casts
		* `#31207 <https://github.com/numpy/numpy/pull/31207>`__: MAINT: Rename ppc64le and s390x workflow (#31121)
		* `#31208 <https://github.com/numpy/numpy/pull/31208>`__: BUG: Fix matvec/vecmat in-place aliasing (out=input produces...
		* `#31209 <https://github.com/numpy/numpy/pull/31209>`__: TYP: ``tile``\ : accept numpy scalars and arrays as second argument...
		* `#31211 <https://github.com/numpy/numpy/pull/31211>`__: DEP: Undo deprecation for np.dtype() signature used by old pickles...
		* `#31212 <https://github.com/numpy/numpy/pull/31212>`__: REV: Manual revert of float16 svml use (#31178)
		* `#31222 <https://github.com/numpy/numpy/pull/31222>`__: TYP: ``ix_`` fix for boolean and non-1d input (#31218)
		* `#31329 <https://github.com/numpy/numpy/pull/31329>`__: BUG: incorrect temp elision for new-style (NEP 43) user-defined...
		* `#31330 <https://github.com/numpy/numpy/pull/31330>`__: TYP: fix sliding_window_view axis parameter typing
		* `#31335 <https://github.com/numpy/numpy/pull/31335>`__: BUG: Prevent deadlock due to downstream importing NumPy in dlopen...
		* `#31336 <https://github.com/numpy/numpy/pull/31336>`__: BUG: Fix segfault in nditer.multi_index when __getitem__ raises...
		* `#31338 <https://github.com/numpy/numpy/pull/31338>`__: TYP: Fix ruff lint error
		* `#31357 <https://github.com/numpy/numpy/pull/31357>`__: BUG: fix memory leak in np.zeros when fill-zero loop raises (#31320)
		* `#31358 <https://github.com/numpy/numpy/pull/31358>`__: BUG: np.einsum() fails with a 0-dimensional out argument and...
		* `#31379 <https://github.com/numpy/numpy/pull/31379>`__: BUG: Fix signed overflow issue in npy_gcd for INT_MIN on s390x...
		* `#31383 <https://github.com/numpy/numpy/pull/31383>`__: CI: remove Cirrus CI FreeBSD job (#31380)
		* `#31390 <https://github.com/numpy/numpy/pull/31390>`__: BUILD: newer MKL uses so.3
		* `#31391 <https://github.com/numpy/numpy/pull/31391>`__: BLD/MAINT: improve support for Intel LLVM compilers
		* `#31401 <https://github.com/numpy/numpy/pull/31401>`__: BUG: Avoid UB in safe_[add,sub,mul] helpers (#31396)
		* `#31402 <https://github.com/numpy/numpy/pull/31402>`__: BUG: exclude __pycache__ directories from wheels (#31397)
		* `#31404 <https://github.com/numpy/numpy/pull/31404>`__: TYP: ``_NestedSequence`` type parameter default to work around...
		* `#31426 <https://github.com/numpy/numpy/pull/31426>`__: TYP: Fix ``DTypeLike`` runtime type-checker support (#31425)

+10

numpy/f2py/tests/test_capi_maps.py

		from numpy.f2py import capi_maps


		def test_complex_long_double_capi_map():
		assert capi_maps.c2capi_map["complex_long_double"] == "NPY_CLONGDOUBLE"


		def test_complex_long_double_is_distinct():
		assert capi_maps.c2pycode_map["complex_long_double"] != capi_maps.c2pycode_map["complex_double"]
		assert capi_maps.c2capi_map["complex_long_double"] != capi_maps.c2capi_map["complex_double"]

+1

-0

.github/workflows/linux_blas.yml

		@@ -315,2 +315,3 @@ name: BLAS tests (Linux)
		for i in $( ls libmkl.so.2 ); do ln -s $i ${i%.}; done
		for i in $( ls libmkl.so.3 ); do ln -s $i ${i%.}; done

		@@ -317,0 +318,0 @@ - name: Build with defaults (LP64)

+0

-35

.github/workflows/linux_qemu.yml

		@@ -46,37 +46,2 @@ # Meson's Python module doesn't support crosscompiling,
		- [
		"ppc64le",
		"powerpc64le-linux-gnu",
		"ppc64le/ubuntu:22.04",
		"-Dallow-noblas=true",
		"test_kind or test_multiarray or test_simd or test_umath or test_ufunc",
		"ppc64le"
		]
		- [
		"ppc64le - baseline(Power9)",
		"powerpc64le-linux-gnu",
		"ppc64le/ubuntu:22.04",
		"-Dallow-noblas=true -Dcpu-baseline=vsx3",
		"test_kind or test_multiarray or test_simd or test_umath or test_ufunc",
		"ppc64le"
		]
		- [
		"s390x",
		"s390x-linux-gnu",
		"s390x/ubuntu:22.04",
		"-Dallow-noblas=true",
		# Skipping TestRationalFunctions.test_gcd_overflow test
		# because of a possible qemu bug that appears to be related to int64 overflow in absolute operation.
		# TODO(@seiko2plus): Confirm the bug and provide a minimal reproducer, then report it to upstream.
		"(test_kind or test_multiarray or test_simd or test_umath or test_ufunc) and not test_gcd_overflow",
		"s390x"
		]
		- [
		"s390x - baseline(Z13)",
		"s390x-linux-gnu",
		"s390x/ubuntu:22.04",
		"-Dallow-noblas=true -Dcpu-baseline=vx",
		"(test_kind or test_multiarray or test_simd or test_umath or test_ufunc) and not test_gcd_overflow",
		"s390x"
		]
		- [
		"riscv64",
		@@ -83,0 +48,0 @@ "riscv64-linux-gnu",

+1

-0

.mailmap

		@@ -83,2 +83,3 @@ # Prevent git from showing duplicate names with commands like "git shortlog"
		Akhil Kannan <akhilkannan10a@gmail.com> <143798318+Alverok@users.noreply.github.com>
		Anarion Zuo <anarionzuo@outlook.com>
		Arun Kota <arunkumarkota@gmail.com>
		@@ -85,0 +86,0 @@ Arun Kota <arunkumarkota@gmail.com> Arun Kota <arunkota@Aruns-iMac.local>

+1

-0

doc/source/release.rst

		@@ -8,2 +8,3 @@ *************

		2.4.5 <release/2.4.5-notes>
		2.4.4 <release/2.4.4-notes>
		@@ -10,0 +11,0 @@ 2.4.3 <release/2.4.3-notes>

+1

-1

doc/source/release/2.4.4-notes.rst

		@@ -7,3 +7,3 @@ .. currentmodule:: numpy

		The NumPy 2.4.3 is a patch release that fixes bugs discovered after the 2.4.2
		The NumPy 2.4.4 is a patch release that fixes bugs discovered after the 2.4.3
		release. It should finally close issue #30816, the OpenBLAS threading problem
		@@ -10,0 +10,0 @@ on ARM.

+8

-1

numpy/__init__.py

		@@ -874,3 +874,10 @@ """
		def blas_fpe_check():
		# Check if BLAS adds spurious FPEs, mostly seen on M4 arms with Accelerate.
		if sys.platform != "darwin":
		# We currently assume this is limited to MacOS as downstream NumPy
		# import during dlopen caused a deadlock regression: gh-31284
		return

		# Check if BLAS adds spurious FPEs, seen on M4 arms with Accelerate.
		# In this case we disable FPE reporting since the use of SME poisons
		# it and Accelerate doesn't sanitize them.
		with errstate(all='raise'):
		@@ -877,0 +884,0 @@ x = ones((20, 20))

+2

-0

numpy/_core/config.h.in

		@@ -35,2 +35,4 @@ #mesondefine SIZEOF_PY_INTPTR_T
		#mesondefine HAVE___BUILTIN_MUL_OVERFLOW
		#mesondefine HAVE___BUILTIN_ADD_OVERFLOW
		#mesondefine HAVE___BUILTIN_SUB_OVERFLOW
		#mesondefine HAVE___BUILTIN_PREFETCH
		@@ -37,0 +39,0 @@

+1

-1

numpy/_core/einsumfunc.py

		@@ -1228,3 +1228,3 @@ """
		# to reshape / transpose ``ab`` after the matmul
		out[:] = ab
		out[...] = ab
		ab = out
		@@ -1231,0 +1231,0 @@ elif output_order is not None:

+9

-8

numpy/_core/fromnumeric.pyi

		@@ -136,2 +136,3 @@ # ruff: noqa: ANN401
		_PyScalar: TypeAlias = complex \| bytes \| str
		_PyScalarT = TypeVar("_PyScalarT", bound=_PyScalar)

		@@ -575,4 +576,5 @@ # TODO: Fix overlapping overloads: https://github.com/numpy/numpy/issues/27032

		# this prevents `Any` from being returned with Pyright
		@overload
		# `collections.abc.Sequence` can't be used here because `bytes` and `str` are
		# subtypes of it, which would make the return types incompatible.
		@overload # this prevents `Any` from being returned with Pyright
		def shape(a: _SupportsShape[Never]) -> _AnyShape: ...
		@@ -583,10 +585,9 @@ @overload
		def shape(a: _PyScalar) -> tuple[()]: ...
		# `collections.abc.Sequence` can't be used hesre, since `bytes` and `str` are
		# subtypes of it, which would make the return types incompatible.
		@overload # an unbound type variable is used because `list` is invariant
		def shape(a: _PyArray[_PyScalarT]) -> tuple[int]: ...
		@overload
		def shape(a: _PyArray[_PyScalar]) -> tuple[int]: ...
		def shape(a: Sequence[_PyArray[_PyScalarT]]) -> tuple[int, int]: ...
		@overload
		def shape(a: _PyArray[_PyArray[_PyScalar]]) -> tuple[int, int]: ...
		# this overload will be skipped by typecheckers that don't support PEP 688
		@overload
		def shape(a: Sequence[Sequence[_PyArray[_PyScalarT]]]) -> tuple[int, int, int]: ...
		@overload # this will be skipped by typecheckers that don't support PEP 688
		def shape(a: memoryview \| bytearray) -> tuple[int]: ...
		@@ -593,0 +594,0 @@ @overload

+2

-2

numpy/_core/include/numpy/npy_common.h

		@@ -395,7 +395,7 @@ #ifndef NUMPY_CORE_INCLUDE_NUMPY_NPY_COMMON_H_

		#if defined(_MSC_VER) && !defined(__INTEL_COMPILER)
		#if defined(_MSC_VER) && !defined(__INTEL_COMPILER) && !defined(__INTEL_LLVM_COMPILER)
		typedef _Dcomplex npy_cdouble;
		typedef _Fcomplex npy_cfloat;
		typedef _Lcomplex npy_clongdouble;
		#else /* !defined(_MSC_VER) \|\| defined(__INTEL_COMPILER) */
		#else /* !defined(_MSC_VER) \|\| defined(__INTEL_COMPILER) && !defined(__INTEL_LLVM_COMPILER) */
		typedef double _Complex npy_cdouble;
		@@ -402,0 +402,0 @@ typedef float _Complex npy_cfloat;

+48

-3

numpy/_core/src/common/npy_extint128.h

		@@ -11,2 +11,16 @@ #ifndef NUMPY_CORE_SRC_COMMON_NPY_EXTINT128_H_

		/*
		* Integer add/sub/mul with overflow checking.
		*
		* On overflow, *overflow_flag is set to 1 and the return value is
		* unspecified (callers must not use it). The arithmetic itself is
		* never performed on values that would overflow signed `npy_int64`,
		* so these helpers are free of the signed-overflow undefined behavior
		* that an unguarded `a + b` / `a - b` / `a * b` would have.
		*
		* `__builtin_{add,sub,mul}_overflow` are probed independently by the
		* meson build (see `numpy/_core/meson.build`); fall back to a
		* branch-and-skip implementation otherwise.
		*/

		/* Integer addition with overflow checking */
		@@ -16,9 +30,19 @@ static inline npy_int64
		{
		#ifdef HAVE___BUILTIN_ADD_OVERFLOW
		npy_int64 result;
		if (__builtin_add_overflow(a, b, &result)) {
		*overflow_flag = 1;
		}
		return result;
		#else
		if (a > 0 && b > NPY_MAX_INT64 - a) {
		*overflow_flag = 1;
		return 0;
		}
		else if (a < 0 && b < NPY_MIN_INT64 - a) {
		if (a < 0 && b < NPY_MIN_INT64 - a) {
		*overflow_flag = 1;
		return 0;
		}
		return a + b;
		#endif
		}
		@@ -31,9 +55,19 @@
		{
		#ifdef HAVE___BUILTIN_SUB_OVERFLOW
		npy_int64 result;
		if (__builtin_sub_overflow(a, b, &result)) {
		*overflow_flag = 1;
		}
		return result;
		#else
		if (a >= 0 && b < a - NPY_MAX_INT64) {
		*overflow_flag = 1;
		return 0;
		}
		else if (a < 0 && b > a - NPY_MIN_INT64) {
		if (a < 0 && b > a - NPY_MIN_INT64) {
		*overflow_flag = 1;
		return 0;
		}
		return a - b;
		#endif
		}
		@@ -46,5 +80,13 @@
		{
		#ifdef HAVE___BUILTIN_MUL_OVERFLOW
		npy_int64 result;
		if (__builtin_mul_overflow(a, b, &result)) {
		*overflow_flag = 1;
		}
		return result;
		#else
		if (a > 0) {
		if (b > NPY_MAX_INT64 / a \|\| b < NPY_MIN_INT64 / a) {
		*overflow_flag = 1;
		return 0;
		}
		@@ -55,8 +97,11 @@ }
		*overflow_flag = 1;
		return 0;
		}
		else if (b < 0 && a < NPY_MAX_INT64 / b) {
		if (b < 0 && a < NPY_MAX_INT64 / b) {
		*overflow_flag = 1;
		return 0;
		}
		}
		return a * b;
		#endif
		}
		@@ -63,0 +108,0 @@

+1

-1

numpy/_core/src/common/simd/vec/arithmetic.h

		@@ -289,3 +289,3 @@ #ifndef NPY_SIMD
		{
		npyv_b64 overflow = npyv_and_b64(vec_cmpeq(a, npyv_setall_s64(-1LL << 63)), (npyv_b64)divisor.val[1]);
		npyv_b64 overflow = npyv_and_b64(vec_cmpeq(a, npyv_setall_s64(0x8000000000000000LL)), (npyv_b64)divisor.val[1]);
		npyv_s64 d = vec_sel(divisor.val[0], npyv_setall_s64(1), overflow);
		@@ -292,0 +292,0 @@ return vec_div(a, d);

+1

-1

numpy/_core/src/multiarray/arrayobject.c

		@@ -445,5 +445,5 @@ /*
		PyDataMem_UserFREE(fa->data, nbytes, fa->mem_handler);
		Py_DECREF(fa->mem_handler);
		}
		}
		Py_CLEAR(fa->mem_handler);

		@@ -450,0 +450,0 @@ /* must match allocation in PyArray_NewFromDescr */

+1

-1

numpy/_core/src/multiarray/calculation.c

		@@ -832,3 +832,3 @@ #define NPY_NO_DEPRECATED_API NPY_API_VERSION
		if (PyArray_ISCOMPLEX(self) \|\| PyArray_ISOBJECT(self) \|\|
		PyArray_ISUSERDEF(self)) {
		PyArray_ISUSERDEF(self) \|\| !NPY_DT_is_legacy(PyArray_DESCR(self))) {
		if (out == NULL) {
		@@ -835,0 +835,0 @@ return PyArray_GenericUnaryFunction(self,

+25

-0

numpy/_core/src/multiarray/compiled_base.c

		@@ -1631,5 +1631,25 @@ #define NPY_NO_DEPRECATED_API NPY_API_VERSION
		#endif

		#if NPY_BYTE_ORDER == NPY_BIG_ENDIAN
		#if NPY_SIMD_WIDTH == 16
		arr[0] = npy_bswap8(arr[0]);
		#elif NPY_SIMD_WIDTH == 32
		arr[0] = npy_bswap8(arr[0]);
		arr[1] = npy_bswap8(arr[1]);
		#else
		arr[0] = npy_bswap8(arr[0]);
		arr[1] = npy_bswap8(arr[1]);
		arr[2] = npy_bswap8(arr[2]);
		arr[3] = npy_bswap8(arr[3]);
		#endif
		#endif
		memcpy(outptr, arr, sizeof(arr));
		outptr += vstepx4;
		} else {
		#if NPY_BYTE_ORDER == NPY_BIG_ENDIAN
		bb[0] = npy_bswap8(bb[0]);
		bb[1] = npy_bswap8(bb[1]);
		bb[2] = npy_bswap8(bb[2]);
		bb[3] = npy_bswap8(bb[3]);
		#endif
		for(int i = 0; i < 4; i++) {
		@@ -1649,2 +1669,7 @@ for (int j = 0; j < vstep; j++) {
		npy_uint64 bb = npyv_tobits_b8(npyv_cmpneq_u8(va, v_zero));

		#if NPY_BYTE_ORDER == NPY_BIG_ENDIAN
		bb = npy_bswap8(bb);
		#endif

		for (int i = 0; i < vstep; ++i) {
		@@ -1651,0 +1676,0 @@ memcpy(outptr, (char*)&bb + i, 1);

+1

-1

numpy/_core/tests/test_cpu_features.py

		@@ -426,3 +426,3 @@ import os
		def load_flags(self):
		self.load_flags_auxv()
		self.load_flags_cpuinfo("features")

		@@ -429,0 +429,0 @@

+7

-0

numpy/_core/tests/test_deprecations.py

		@@ -389,2 +389,7 @@ """
		self.assert_deprecated(lambda: np.dtype("f8", align=3))
		self.assert_deprecated(lambda: np.dtype("f8", align=0, copy=10**100))
		self.assert_deprecated(lambda: np.dtype("f8", align=10**100, copy=0))
		# Subclasses of ints don't hit the below pickle code path:
		self.assert_deprecated(
		lambda: np.dtype("f8", align=np.long(0), copy=np.long(1)))

		@@ -395,2 +400,4 @@ @pytest.mark.parametrize("align", [True, False, np.True_, np.False_])
		self.assert_not_deprecated(lambda: np.dtype("f8", align=align))
		# The following specific case is used by old pickles:
		self.assert_not_deprecated(lambda: np.dtype("f8", align=0, copy=1))

		@@ -397,0 +404,0 @@

+8

-0

numpy/_core/tests/test_einsum.py

		@@ -765,2 +765,10 @@ import itertools

		def test_einsum_0d_out(self):
		# Issue gh-31350, a zero-dimensional out must not cause an error
		# with optimize='optimal'
		a = np.ones(7)
		out = np.array(0)
		np.einsum('i,i->', a, a, out=out, optimize='optimal')
		assert_equal(out, 7)

		def test_object_loop(self):
		@@ -767,0 +775,0 @@

+0

-40

numpy/_core/tests/test_umath_accuracy.py

		@@ -24,4 +24,2 @@ import os

		IS_AVX512FP16 = __cpu_features__.get('AVX512FP16', False)

		# only run on linux with AVX, also avoid old glibc (numpy/numpy#20448).
		@@ -88,4 +86,2 @@ runtest = (sys.platform.startswith('linux')

		@pytest.mark.skipif(IS_AVX512FP16,
		reason="SVML FP16 have slightly higher ULP errors")
		@pytest.mark.parametrize("ufunc", UNARY_OBJECT_UFUNCS)
		@@ -99,37 +95,1 @@ def test_validate_fp16_transcendentals(self, ufunc):
		maxulp=1, dtype=np.float16)

		@pytest.mark.skipif(not IS_AVX512FP16,
		reason="lower ULP only apply for SVML FP16")
		def test_validate_svml_fp16(self):
		max_ulp_err = {
		"arccos": 2.54,
		"arccosh": 2.09,
		"arcsin": 3.06,
		"arcsinh": 1.51,
		"arctan": 2.61,
		"arctanh": 1.88,
		"cbrt": 1.57,
		"cos": 1.43,
		"cosh": 1.33,
		"exp2": 1.33,
		"exp": 1.27,
		"expm1": 0.53,
		"log": 1.80,
		"log10": 1.27,
		"log1p": 1.88,
		"log2": 1.80,
		"sin": 1.88,
		"sinh": 2.05,
		"tan": 2.26,
		"tanh": 3.00,
		}

		with np.errstate(all='ignore'):
		arr = np.arange(65536, dtype=np.int16)
		datafp16 = np.frombuffer(arr.tobytes(), dtype=np.float16)
		datafp32 = datafp16.astype(np.float32)
		for func in max_ulp_err:
		ufunc = getattr(np, func)
		ulp = np.ceil(max_ulp_err[func])
		assert_array_max_ulp(ufunc(datafp16), ufunc(datafp32),
		maxulp=ulp, dtype=np.float16)

+4

-2

numpy/_typing/_dtype_like.py

		@@ -1,3 +0,3 @@
		from collections.abc import Sequence # noqa: F811
		from typing import Any, Protocol, TypeAlias, TypedDict, TypeVar
		from collections.abc import Sequence
		from typing import Any, Protocol, TypeAlias, TypedDict, TypeVar, runtime_checkable

		@@ -44,2 +44,3 @@ import numpy as np

		@runtime_checkable
		class _HasDType(Protocol[_DTypeT_co]):
		@@ -50,2 +51,3 @@ @property

		@runtime_checkable
		class _HasNumPyDType(Protocol[_DTypeT_co]):
		@@ -52,0 +54,0 @@ @property

+9

-3

numpy/_typing/_nested_sequence.py

		"""A module containing the `_NestedSequence` protocol."""

		from typing import TYPE_CHECKING, Any, Protocol, TypeVar, runtime_checkable
		from typing import TYPE_CHECKING, Any, Protocol, runtime_checkable

		if TYPE_CHECKING:
		from collections.abc import Iterator
		from typing_extensions import TypeVar

		_T_co = TypeVar("_T_co", covariant=True, default=Any)
		else:
		from typing import TypeVar

		_T_co = TypeVar("_T_co", covariant=True)


		__all__ = ["_NestedSequence"]

		_T_co = TypeVar("_T_co", covariant=True)


		@runtime_checkable
		@@ -14,0 +20,0 @@ class _NestedSequence(Protocol[_T_co]):

+1

-1

numpy/f2py/capi_maps.py

		@@ -73,3 +73,3 @@ """
		'complex_double': 'NPY_CDOUBLE',
		'complex_long_double': 'NPY_CDOUBLE',
		'complex_long_double': 'NPY_CLONGDOUBLE',
		'string': 'NPY_STRING',
		@@ -76,0 +76,0 @@ 'character': 'NPY_STRING'}

+7

-7

numpy/lib/_index_tricks_impl.pyi

		@@ -238,15 +238,15 @@ from _typeshed import Incomplete, SupportsLenAndGetItem
		@overload
		def ix_(*args: _FiniteNestedSequence[_HasDType[_DTypeT]]) -> tuple[np.ndarray[_AnyShape, _DTypeT], ...]: ...
		def ix_() -> tuple[()]: ...
		@overload
		def ix_(*args: str \| _NestedSequence[str]) -> tuple[NDArray[np.str_], ...]: ...
		def ix_(*args: Sequence[_HasDType[_DTypeT]] \| _HasDType[_DTypeT]) -> tuple[np.ndarray[_AnyShape, _DTypeT], ...]: ...
		@overload
		def ix_(*args: bytes \| _NestedSequence[bytes]) -> tuple[NDArray[np.bytes_], ...]: ...
		def ix_(*args: Sequence[str]) -> tuple[NDArray[np.str_], ...]: ...
		@overload
		def ix_(*args: bool \| _NestedSequence[bool]) -> tuple[NDArray[np.bool], ...]: ...
		def ix_(*args: Sequence[bytes]) -> tuple[NDArray[np.bytes_], ...]: ...
		@overload
		def ix_(*args: int \| _NestedSequence[int]) -> tuple[NDArray[np.intp], ...]: ...
		def ix_(*args: Sequence[int]) -> tuple[NDArray[np.intp], ...]: ...
		@overload
		def ix_(*args: float \| _NestedSequence[float]) -> tuple[NDArray[np.float64], ...]: ...
		def ix_(*args: Sequence[float]) -> tuple[NDArray[np.float64], ...]: ...
		@overload
		def ix_(*args: complex \| _NestedSequence[complex]) -> tuple[NDArray[np.complex128], ...]: ...
		def ix_(*args: Sequence[complex]) -> tuple[NDArray[np.complex128], ...]: ...

		@@ -253,0 +253,0 @@ #

+3

-8

numpy/lib/_shape_base_impl.pyi

		@@ -34,2 +34,3 @@ from collections.abc import Callable, Sequence
		_ArrayLikeFloat_co,
		_ArrayLikeInt,
		_ArrayLikeInt_co,
		@@ -229,10 +230,4 @@ _ArrayLikeObject_co,
		@overload
		def tile(
		A: _ArrayLike[_ScalarT],
		reps: int \| Sequence[int],
		) -> NDArray[_ScalarT]: ...
		def tile(A: _ArrayLike[_ScalarT], reps: _ArrayLikeInt) -> NDArray[_ScalarT]: ...
		@overload
		def tile(
		A: ArrayLike,
		reps: int \| Sequence[int],
		) -> NDArray[Any]: ...
		def tile(A: ArrayLike, reps: _ArrayLikeInt) -> NDArray[Any]: ...

+3

-3

numpy/lib/_stride_tricks_impl.pyi

		from collections.abc import Iterable
		from typing import Any, SupportsIndex, TypeVar, overload
		from typing import Any, TypeVar, overload

		@@ -41,3 +41,3 @@ from numpy import generic
		window_shape: int \| Iterable[int],
		axis: SupportsIndex \| None = None,
		axis: int \| tuple[int, ...] \| None = None,
		*,
		@@ -51,3 +51,3 @@ subok: bool = False,
		window_shape: int \| Iterable[int],
		axis: SupportsIndex \| None = None,
		axis: int \| tuple[int, ...] \| None = None,
		*,
		@@ -54,0 +54,0 @@ subok: bool = False,

+8

-1

numpy/typing/tests/data/reveal/fromnumeric.pyi

		@@ -27,2 +27,6 @@ """Tests for :mod:`_core.fromnumeric`."""

		_py_list_1d: list[int]
		_py_list_2d: list[list[int]]
		_py_list_3d: list[list[list[int]]]

		# integer‑dtype subclass for argmin/argmax
		@@ -144,3 +148,6 @@ class NDArrayIntSubclass(npt.NDArray[np.intp]): ...
		assert_type(np.shape([[2]]), tuple[int, int])
		assert_type(np.shape([[[3]]]), tuple[Any, ...])
		assert_type(np.shape([[[3]]]), tuple[int, int, int])
		assert_type(np.shape(_py_list_1d), tuple[int])
		assert_type(np.shape(_py_list_2d), tuple[int, int])
		assert_type(np.shape(_py_list_3d), tuple[int, int, int])
		assert_type(np.shape(AR_b), tuple[Any, ...])
		@@ -147,0 +154,0 @@ assert_type(np.shape(AR_nd), tuple[Any, ...])

+1

-1

numpy/typing/tests/data/reveal/index_tricks.pyi

		@@ -61,3 +61,3 @@ from types import EllipsisType

		assert_type(np.ix_(AR_LIKE_b), tuple[npt.NDArray[np.bool], ...])
		assert_type(np.ix_(AR_LIKE_b), tuple[npt.NDArray[np.intp], ...])
		assert_type(np.ix_(AR_LIKE_i, AR_LIKE_f), tuple[npt.NDArray[np.float64], ...])
		@@ -64,0 +64,0 @@ assert_type(np.ix_(AR_i8), tuple[npt.NDArray[np.int64], ...])

+2

-2

numpy/version.py

		@@ -5,8 +5,8 @@
		"""
		version = "2.4.4"
		version = "2.4.5"
		__version__ = version
		full_version = version

		git_revision = "be93fe2960dbf49b4647f5783c66d967fb2c65b5"
		git_revision = "26e81854b03fea36ac2b21e3158756fdd40f19bb"
		release = 'dev' not in version and '+' not in version
		short_version = version.split("+")[0]

+1

-1

PKG-INFO

		Metadata-Version: 2.4
		Name: numpy
		Version: 2.4.4
		Version: 2.4.5
		Summary: Fundamental package for array computing in Python
		@@ -5,0 +5,0 @@ Author: Travis E. Oliphant et al.

+1

-1

pyproject.toml

		@@ -10,3 +10,3 @@ [build-system]
		name = "numpy"
		version = "2.4.4"
		version = "2.4.5"
		description = "Fundamental package for array computing in Python"
		@@ -13,0 +13,0 @@ authors = [{name = "Travis E. Oliphant et al."}]

.cirrus.star

Sorry, the diff of this file is not supported yet

-71

.github/workflows/linux-ppc64le.yml

		name: Native ppc64le Linux Test

		on:
		pull_request:
		branches:
		- main
		- maintenance/**
		paths-ignore:
		- '**.pyi'
		- '**.md'
		- '**.rst'
		workflow_dispatch:

		permissions:
		contents: read

		concurrency:
		group: ${{ github.workflow }}-${{ github.head_ref \|\| github.run_id }}
		cancel-in-progress: true

		jobs:
		native_ppc64le:
		# This job runs only in the main NumPy repository.
		# It requires a native ppc64le GHA runner, which is not available on forks.
		# For more details, see: https://github.com/numpy/numpy/issues/29125
		if: github.repository == 'numpy/numpy'
		runs-on: ubuntu-24.04-ppc64le-p10

		strategy:
		fail-fast: false
		matrix:
		config:
		- name: "GCC"
		args: "-Dallow-noblas=false"
		- name: "clang"
		args: "-Dallow-noblas=false"

		name: "${{ matrix.config.name }}"
		steps:
		- uses: actions/checkout@1af3b93b6815bc44a9784bd300feb67ff0d1eeb3 # v6.0.0
		with:
		submodules: recursive
		fetch-tags: true

		- name: Install dependencies
		run: \|
		sudo apt update
		sudo apt install -y python3.12 python3-pip python3-dev ninja-build gfortran \
		build-essential libopenblas-dev liblapack-dev pkg-config
		pip install --upgrade pip
		pip install -r requirements/build_requirements.txt -r requirements/test_requirements.txt
		echo "/home/runner/.local/bin" >> $GITHUB_PATH

		- name: Install clang
		if: matrix.config.name == 'clang'
		run: \|
		sudo apt install -y clang
		export CC=clang
		export CXX=clang++

		- name: Meson Build
		run: \|
		spin build -- ${{ matrix.config.args }}

		- name: Meson Log
		if: always()
		run: cat build/meson-logs/meson-log.txt

		- name: Run Tests
		run: \|
		spin test -- --timeout=60 --durations=10

-68

tools/ci/cirrus_arm.yml

		modified_clone: &MODIFIED_CLONE
		# makes sure that for a PR the CI runs against a merged main
		clone_script: \|
		if [ -z "$CIRRUS_PR" ]; then
		# if you're not in a PR then clone against the branch name that was pushed to.
		git clone --recursive --branch=$CIRRUS_BRANCH https://x-access-token:${CIRRUS_REPO_CLONE_TOKEN}@github.com/${CIRRUS_REPO_FULL_NAME}.git $CIRRUS_WORKING_DIR
		git reset --hard $CIRRUS_CHANGE_IN_REPO
		else
		# it's a PR so clone the main branch then merge the changes from the PR
		git clone https://x-access-token:${CIRRUS_REPO_CLONE_TOKEN}@github.com/${CIRRUS_REPO_FULL_NAME}.git $CIRRUS_WORKING_DIR
		git fetch origin pull/$CIRRUS_PR/head:pull/$CIRRUS_PR

		# CIRRUS_BASE_BRANCH will probably be `main` for the majority of the time
		# However, if you do a PR against a maintenance branch we will want to
		# merge the PR into the maintenance branch, not main
		git checkout $CIRRUS_BASE_BRANCH

		# alpine git package needs default user.name and user.email to be set before a merge
		git -c user.email="you@example.com" merge --no-commit pull/$CIRRUS_PR
		git submodule update --init --recursive
		fi


		freebsd_test_task:
		use_compute_credits: $CIRRUS_USER_COLLABORATOR == 'true'
		compute_engine_instance:
		image_project: freebsd-org-cloud-dev
		image: family/freebsd-14-3
		platform: freebsd
		cpu: 1
		memory: 4G

		install_devtools_script: \|
		pkg install -y git bash ninja ccache blas cblas lapack pkgconf
		pkg install -y python311

		<<: *MODIFIED_CLONE

		ccache_cache:
		folder: .ccache
		populate_script:
		- mkdir -p .ccache
		fingerprint_key: ccache-freebsd

		prepare_env_script: \|
		# Create a venv (the `source` command needs bash, not the default sh shell)
		chsh -s /usr/local/bin/bash
		python3.11 -m venv .venv
		source .venv/bin/activate
		# Minimal build and test requirements
		python3.11 -m pip install -U pip
		python3.11 -m pip install meson-python Cython pytest hypothesis

		build_script: \|
		chsh -s /usr/local/bin/bash
		source .venv/bin/activate
		python3.11 -m pip install . --no-build-isolation -v -Csetup-args="-Dallow-noblas=false"

		test_script: \|
		chsh -s /usr/local/bin/bash
		source .venv/bin/activate
		cd tools
		python3.11 -m pytest --pyargs numpy -m "not slow"
		ccache -s

		on_failure:
		debug_script: \|
		cat build/meson-logs/meson-log.txt

numpy/_core/include/numpy/ndarraytypes.h