numpy-typing-compat

Overview

This package provides version-specific boolean constants that allow library authors to write NumPy-version-dependent static type annotations. Similar to how you might use if sys.version_info >= (3, 12): for Python version checks, numpy-typing-compat enables static type-checkers to understand which NumPy version is being used and apply appropriate type annotations.

[!TIP] numpy-typing-compat is intended as a helper package for the optype library, which provides high-level typing utilities for annotating NumPy operations. In most situations, it is recommended to use optype instead of numpy-typing-compat. See the optype.numpy documentation for more details.

Installation and Versioning

The numpy-typing-compat packages are available on PyPI and conda-forge.

Modern package managers such as uv and pixi will automatically install the appropriate version of numpy-typing-compat that matches your installed NumPy version, in order to satisfy the numpy dependency restrictions of numpy-typing-compat.

[!WARNING] Legacy package managers such as pip don't always respect dependency restrictions. Running pip install --upgrade numpy will not automatically upgrade numpy-typing-compat to the correct version. If for some reason you need to use pip, then be sure to manually install the correct version of numpy-typing-compat that matches your installed NumPy version. Running pip check will tell you whether the installed numpy-typing-compat and numpy versions are compatible.

To illustrate, the numpy-typing-compat==20250818.2.1 distribution of release v20250818 specifies numpy>=2.1,<2.2 as a required dependency. Modern package managers will ensure that these dependency restrictions are satisfied. That way, if you upgrade numpy from 2.1 to 2.3 (e.g. by running uv sync --upgrade) then uv will also automatically look for a version of numpy-typing-compat that satisfies the new numpy version, which for this example would be numpy-typing-compat==20250818.2.3.

numpy-typing-compat	NumPy	Python
20250818.1.22	>=1.22,<1.23	>=3.8
20250818.1.23	>=1.23,<1.25	>=3.8
20250818.1.25	>=1.25,<2.0	>=3.9
20250818.2.0	>=2.0,<2.1	>=3.9
20250818.2.1	>=2.1,<2.2	>=3.10
20250818.2.2	>=2.2,<2.3	>=3.10
20250818.2.3	>=2.3,<2.4	>=3.11

Reference

array_api

Additionally, the package provides a numpy_typing_compat.array_api namespace that's a re-export of the numpy.array_api module on numpy < 2.0, or the main numpy module on numpy >= 2.0. Note that the numpy.array_api module was introduced in numpy >= 1.23, so it isn't available in e.g. numpy-typing-compat==20250818.1.22.

long and ulong

NumPy 2.0 introduced the new long and ulong scalar types, which are not available in numpy < 2.0, and instead went by the names int_ and uint (which in numpy >= 2.0 are aliases for intp and uintp). If you need to support both NumPy versions, you can use the long and ulong types from numpy_typing_compat, which on numpy < 2.0 are aliases for np.int_ and np.uint, and on numpy >= 2.0 are re-exports of np.long and np.ulong.

StringDType

In NumPy 2.0, the numpy.dtypes.StringDType was introduced, but it wasn't until NumPy 2.1 that it was also available in the numpy stubs. The numpy_typing_compat.StringDType is a re-export of numpy.dtypes.StringDType on numpy >= 2.1, and an alias of np.dtype[Never] on numpy < 2.1. This allows type checkers to also accept StringDType as a valid type on numpy == 2.0.*.

ABCPolyBase

In NumPy 2.1, the numpy.polynomial._polybase.ABCPolyBase was made into a generic type, and in NumPy 2.2 the type parameter was made optional. This can be problematic on numpy==2.1.* if you also require support for numpy < 2.1. To work around this, the numpy_typing_compat.ABCPolyBase is a type alias for numpy.polynomial._polybase.ABCPolyBase[LiteralString | None] on numpy==2.1.*, and a direct re-export of numpy.polynomial._polybase.ABCPolyBase otherwise. This way, type checkers (when configured in strict mode) won't report an error on numpy==2.1.* when using ABCPolyBase. Note that numpy_typing_compat.ABCPolyBase does not accept a type parameter, even on numpy >= 2.1, for the sake of consistency.

LiteralTrue and LiteralFalse

In NumPy 2.2, the np.bool scalar type became a generic type that accepts a type parameter of either True or False. For compatibility with numpy < 2.2, the numpy_typing_compat.LiteralTrue and numpy_typing_compat.LiteralFalse types are provided. These are type aliases for Literal[True] and Literal[False] on numpy < 2.2, and Literal[True] | np.bool[Literal[True]] and Literal[False] | np.bool[Literal[False]] on numpy >= 2.2, respectively.

Version constants

The following low-level boolean version constants are available:

Constant	True when
NUMPY_GE_1_22	numpy >= 1.22
NUMPY_GE_1_23	numpy >= 1.23
NUMPY_GE_1_25	numpy >= 1.25
NUMPY_GE_2_0	numpy >= 2.0
NUMPY_GE_2_1	numpy >= 2.1
NUMPY_GE_2_2	numpy >= 2.2
NUMPY_GE_2_3	numpy >= 2.3

Each constant is typed as Literal[True] or Literal[False] depending on your NumPy version, so that static type-checkers are able to understand the NumPy version being used.

[!WARNING] At the moment, mypy and pyright only have limited support for Literal type conditions, and will not treat if NUMPY_GE_2_0: ... in the same way as if sys.version_info >= (3, 12): ... or if sys.platform == "win32": ....