NumPy 2.1.0 Release Notes#

NumPy 2.1.0 provides support for the upcoming Python 3.13 release and drops support for Python 3.9. In addition to the usual bug fixes and updated Python support, it helps get us back into our usual release cycle after the extended development of 2.0. The highlights for this release are:

  • Support for the array-api 2023.12 standard.

  • Support for Python 3.13.

  • Preliminary support for free threaded Python 3.13.

Python versions 3.10-3.13 are supported in this release.

New functions#

New function numpy.unstack#

A new function np.unstack(array, axis=...) was added, which splits an array into a tuple of arrays along an axis. It serves as the inverse of numpy.stack.

(gh-26579)

Deprecations#

  • The fix_imports keyword argument in numpy.save is deprecated. Since NumPy 1.17, numpy.save uses a pickle protocol that no longer supports Python 2, and ignored fix_imports keyword. This keyword is kept only for backward compatibility. It is now deprecated.

    (gh-26452)

  • Passing non-integer inputs as the first argument of bincount is now deprecated, because such inputs are silently cast to integers with no warning about loss of precision.

    (gh-27076)

Expired deprecations#

  • Scalars and 0D arrays are disallowed for numpy.nonzero and numpy.ndarray.nonzero.

    (gh-26268)

  • set_string_function internal function was removed and PyArray_SetStringFunction was stubbed out.

    (gh-26611)

C API changes#

API symbols now hidden but customizable#

NumPy now defaults to hide the API symbols it adds to allow all NumPy API usage. This means that by default you cannot dynamically fetch the NumPy API from another library (this was never possible on windows).

If you are experiencing linking errors related to PyArray_API or PyArray_RUNTIME_VERSION, you can define the NPY_API_SYMBOL_ATTRIBUTE to opt-out of this change.

If you are experiencing problems due to an upstream header including NumPy, the solution is to make sure you #include "numpy/ndarrayobject.h" before their header and import NumPy yourself based on including-the-c-api.

(gh-26103)

Many shims removed from npy_3kcompat.h#

Many of the old shims and helper functions were removed from npy_3kcompat.h. If you find yourself in need of these, vendor the previous version of the file into your codebase.

(gh-26842)

New PyUFuncObject field process_core_dims_func#

The field process_core_dims_func was added to the structure PyUFuncObject. For generalized ufuncs, this field can be set to a function of type PyUFunc_ProcessCoreDimsFunc that will be called when the ufunc is called. It allows the ufunc author to check that core dimensions satisfy additional constraints, and to set output core dimension sizes if they have not been provided.

(gh-26908)

New Features#

Preliminary Support for Free-Threaded CPython 3.13#

CPython 3.13 will be available as an experimental free-threaded build. See https://py-free-threading.github.io, PEP 703 and the CPython 3.13 release notes for more detail about free-threaded Python.

NumPy 2.1 has preliminary support for the free-threaded build of CPython 3.13. This support was enabled by fixing a number of C thread-safety issues in NumPy. Before NumPy 2.1, NumPy used a large number of C global static variables to store runtime caches and other state. We have either refactored to avoid the need for global state, converted the global state to thread-local state, or added locking.

Support for free-threaded Python does not mean that NumPy is thread safe. Read-only shared access to ndarray should be safe. NumPy exposes shared mutable state and we have not added any locking to the array object itself to serialize access to shared state. Care must be taken in user code to avoid races if you would like to mutate the same array in multiple threads. It is certainly possible to crash NumPy by mutating an array simultaneously in multiple threads, for example by calling a ufunc and the resize method simultaneously. For now our guidance is: “don’t do that”. In the future we would like to provide stronger guarantees.

Object arrays in particular need special care, since the GIL previously provided locking for object array access and no longer does. See Issue #27199 for more information about object arrays in the free-threaded build.

If you are interested in free-threaded Python, for example because you have a multiprocessing-based workflow that you are interested in running with Python threads, we encourage testing and experimentation.

If you run into problems that you suspect are because of NumPy, please open an issue, checking first if the bug also occurs in the “regular” non-free-threaded CPython 3.13 build. Many threading bugs can also occur in code that releases the GIL; disabling the GIL only makes it easier to hit threading bugs.

(gh-26157)

  • numpy.reshape and numpy.ndarray.reshape now support shape and copy arguments.

    (gh-26292)

  • NumPy now supports DLPack v1, support for older versions will be deprecated in the future.

    (gh-26501)

  • numpy.asanyarray now supports copy and device arguments, matching numpy.asarray.

    (gh-26580)

  • numpy.printoptions, numpy.get_printoptions, and numpy.set_printoptions now support a new option, override_repr, for defining custom repr(array) behavior.

    (gh-26611)

  • numpy.cumulative_sum and numpy.cumulative_prod were added as Array API compatible alternatives for numpy.cumsum and numpy.cumprod. The new functions can include a fixed initial (zeros for sum and ones for prod) in the result.

    (gh-26724)

  • numpy.clip now supports max and min keyword arguments which are meant to replace a_min and a_max. Also, for np.clip(a) or np.clip(a, None, None) a copy of the input array will be returned instead of raising an error.

    (gh-26724)

  • numpy.astype now supports device argument.

    (gh-26724)

f2py can generate freethreading-compatible C extensions#

Pass --freethreading-compatible to the f2py CLI tool to produce a C extension marked as compatible with the free threading CPython interpreter. Doing so prevents the interpreter from re-enabling the GIL at runtime when it imports the C extension. Note that f2py does not analyze fortran code for thread safety, so you must verify that the wrapped fortran code is thread safe before marking the extension as compatible.

(gh-26981)

Improvements#

histogram auto-binning now returns bin sizes >=1 for integer input data#

For integer input data, bin sizes smaller than 1 result in spurious empty bins. This is now avoided when the number of bins is computed using one of the algorithms provided by histogram_bin_edges.

(gh-12150)

ndarray shape-type parameter is now covariant and bound to tuple[int, ...]#

Static typing for ndarray is a long-term effort that continues with this change. It is a generic type with type parameters for the shape and the data type. Previously, the shape type parameter could be any value. This change restricts it to a tuple of ints, as one would expect from using ndarray.shape. Further, the shape-type parameter has been changed from invariant to covariant. This change also applies to the subtypes of ndarray, e.g. numpy.ma.MaskedArray. See the typing docs for more information.

(gh-26081)

np.quantile with method closest_observation chooses nearest even order statistic#

This changes the definition of nearest for border cases from the nearest odd order statistic to nearest even order statistic. The numpy implementation now matches other reference implementations.

(gh-26656)

lapack_lite is now thread safe#

NumPy provides a minimal low-performance version of LAPACK named lapack_lite that can be used if no BLAS/LAPACK system is detected at build time.

Until now, lapack_lite was not thread safe. Single-threaded use cases did not hit any issues, but running linear algebra operations in multiple threads could lead to errors, incorrect results, or segfaults due to data races.

We have added a global lock, serializing access to lapack_lite in multiple threads.

(gh-26750)

The numpy.printoptions context manager is now thread and async-safe#

In prior versions of NumPy, the printoptions were defined using a combination of Python and C global variables. We have refactored so the state is stored in a python ContextVar, making the context manager thread and async-safe.

(gh-26846)

Type hinting numpy.polynomial#

Starting from the 2.1 release, PEP 484 type annotations have been included for the functions and convenience classes in numpy.polynomial and its sub-packages.

(gh-26897)

Improved numpy.dtypes type hints#

The type annotations for numpy.dtypes are now a better reflection of the runtime: The numpy.dtype type-aliases have been replaced with specialized dtype subtypes, and the previously missing annotations for numpy.dtypes.StringDType have been added.

(gh-27008)

Performance improvements and changes#

  • numpy.save now uses pickle protocol version 4 for saving arrays with object dtype, which allows for pickle objects larger than 4GB and improves saving speed by about 5% for large arrays.

    (gh-26388)

  • OpenBLAS on x86_64 and i686 is built with fewer kernels. Based on benchmarking, there are 5 clusters of performance around these kernels: PRESCOTT NEHALEM SANDYBRIDGE HASWELL SKYLAKEX.

    (gh-27147)

  • OpenBLAS on windows is linked without quadmath, simplifying licensing

    (gh-27147)

  • Due to a regression in OpenBLAS on windows, the performance improvements when using multiple threads for OpenBLAS 0.3.26 were reverted.

    (gh-27147)

ma.cov and ma.corrcoef are now significantly faster#

The private function has been refactored along with ma.cov and ma.corrcoef. They are now significantly faster, particularly on large, masked arrays.

(gh-26285)

Changes#

  • As numpy.vecdot is now a ufunc it has a less precise signature. This is due to the limitations of ufunc’s typing stub.

    (gh-26313)

  • numpy.floor, numpy.ceil, and numpy.trunc now won’t perform casting to a floating dtype for integer and boolean dtype input arrays.

    (gh-26766)

ma.corrcoef may return a slightly different result#

A pairwise observation approach is currently used in ma.corrcoef to calculate the standard deviations for each pair of variables. This has been changed as it is being used to normalise the covariance, estimated using ma.cov, which does not consider the observations for each variable in a pairwise manner, rendering it unnecessary. The normalisation has been replaced by the more appropriate standard deviation for each variable, which significantly reduces the wall time, but will return slightly different estimates of the correlation coefficients in cases where the observations between a pair of variables are not aligned. However, it will return the same estimates in all other cases, including returning the same correlation matrix as corrcoef when using a masked array with no masked values.

(gh-26285)

Cast-safety fixes in copyto and full#

copyto now uses NEP 50 correctly and applies this to its cast safety. Python integer to NumPy integer casts and Python float to NumPy float casts are now considered “safe” even if assignment may fail or precision may be lost. This means the following examples change slightly:

  • np.copyto(int8_arr, 1000) previously performed an unsafe/same-kind cast

    of the Python integer. It will now always raise, to achieve an unsafe cast you must pass an array or NumPy scalar.

  • np.copyto(uint8_arr, 1000, casting="safe") will raise an OverflowError rather than a TypeError due to same-kind casting.

  • np.copyto(float32_arr, 1e300, casting="safe") will overflow to inf (float32 cannot hold 1e300) rather raising a TypeError.

Further, only the dtype is used when assigning NumPy scalars (or 0-d arrays), meaning that the following behaves differently:

  • np.copyto(float32_arr, np.float64(3.0), casting="safe") raises.

  • np.coptyo(int8_arr, np.int64(100), casting="safe") raises. Previously, NumPy checked whether the 100 fits the int8_arr.

This aligns copyto, full, and full_like with the correct NumPy 2 behavior.

(gh-27091)