Array API 2025.12

.github/workflows/array-api-tests.yml

@@ -33,7 +33,7 @@ jobs:
         uses: actions/checkout@v3
         with:
           repository: data-apis/array-api-tests
-          ref: '2025.05.23'
+          ref: '2026.02.26'
           path: array-api-tests
           submodules: "true"
       - name: Set up Python ${{ matrix.python-version }}
@@ -136,9 +136,12 @@ jobs:
           array_api_tests/test_array_object.py::test_getitem_arrays_and_ints_2
           # test_searchsorted depends on sort which is not implemented
           array_api_tests/test_searching_functions.py::test_searchsorted
+          array_api_tests/test_searching_functions.py::test_searchsorted_with_scalars
           # cumulative_* functions with include_initial=True are not implemented
           array_api_tests/test_statistical_functions.py::test_cumulative_prod
           array_api_tests/test_statistical_functions.py::test_cumulative_sum
+          # repeated axis not supported
+          array_api_tests/test_manipulation_functions.py::TestExpandDims::test_expand_dims_tuples
 
           # not implemented
           array_api_tests/test_array_object.py::test_setitem
@@ -160,6 +163,24 @@ jobs:
           # https://github.com/numpy/numpy/issues/20870
           #array_api_tests/test_data_type_functions.py::test_can_cast
 
+          # complex special cases, see also https://github.com/data-apis/array-api-compat/blob/main/numpy-xfails.txt
+          array_api_tests/test_special_cases.py::test_unary[acosh(real(x_i) is +0 and imag(x_i) is NaN) -> NaN \xb1 \u03c0j/2]
+          array_api_tests/test_special_cases.py::test_unary[atanh(isfinite(real(x_i)) and real(x_i) != 0 and imag(x_i) is NaN) -> NaN + NaN j]
+          array_api_tests/test_special_cases.py::test_unary[atanh(real(x_i) is +0 and imag(x_i) is NaN) -> +0 + NaN j]
+          array_api_tests/test_special_cases.py::test_unary[atanh(real(x_i) is 1 and imag(x_i) is +0) -> +infinity + 0j]
+          array_api_tests/test_special_cases.py::test_unary[atanh(real(x_i) is +infinity and imag(x_i) is NaN) -> +0 + NaN j]
+          array_api_tests/test_special_cases.py::test_unary[atanh(real(x_i) is NaN and imag(x_i) is +infinity) -> \xb10 + \u03c0j/2]
+          array_api_tests/test_special_cases.py::test_unary[expm1(real(x_i) is +infinity and imag(x_i) is +0) -> +infinity + 0j]
+          array_api_tests/test_special_cases.py::test_unary[expm1(real(x_i) is -infinity and imag(x_i) is +infinity) -> -1 + 0j]
+          array_api_tests/test_special_cases.py::test_unary[expm1(real(x_i) is +infinity and imag(x_i) is +infinity) -> infinity + NaN j]
+          array_api_tests/test_special_cases.py::test_unary[expm1(real(x_i) is -infinity and imag(x_i) is NaN) -> -1 + 0j]
+          array_api_tests/test_special_cases.py::test_unary[expm1(real(x_i) is +infinity and imag(x_i) is NaN) -> infinity + NaN j]
+          array_api_tests/test_special_cases.py::test_unary[expm1(real(x_i) is NaN and imag(x_i) is +0) -> NaN + 0j]
+          array_api_tests/test_special_cases.py::test_unary[expm1((real(x_i) is +0 or real(x_i) == -0) and imag(x_i) is +0) -> 0 + 0j]
+          array_api_tests/test_special_cases.py::test_unary[tanh(real(x_i) is +0 and imag(x_i) is +infinity) -> +0 + NaN j]
+          array_api_tests/test_special_cases.py::test_unary[tanh(real(x_i) is +0 and imag(x_i) is NaN) -> +0 + NaN j]
+          array_api_tests/test_special_cases.py::test_unary[tanh(real(x_i) is +infinity and isfinite(imag(x_i)) and imag(x_i) > 0) -> 1 + 0j]
+
           EOF
 
           CUBED_BACKEND_ARRAY_API_MODULE=array_api_compat.numpy pytest -v -rxXfEA --hypothesis-max-examples=2 --disable-data-dependent-shapes --disable-extension linalg --hypothesis-disable-deadline

api_status.md

@@ -1,6 +1,6 @@
 ## Array API Coverage Implementation Status
 
-Cubed supports version [2024.12](https://data-apis.org/array-api/2024.12/index.html) of the Python array API standard, with a few exceptions noted below. The [Fourier transform functions](https://data-apis.org/array-api/2024.12/extensions/fourier_transform_functions.html) are *not* supported.
+Cubed supports version [2025.12](https://data-apis.org/array-api/2025.12/index.html) of the Python array API standard, with a few exceptions noted below. The [Fourier transform functions](https://data-apis.org/array-api/2025.12/extensions/fourier_transform_functions.html) are *not* supported.
 
 This table shows which parts of the the [Array API](https://data-apis.org/array-api/latest/API_specification/index.html) have been implemented in Cubed, and which ones are missing. The version column shows the version when the feature was added to the standard, for version 2022.12 or later.
 
@@ -55,6 +55,7 @@ This table shows which parts of the the [Array API](https://data-apis.org/array-
 |                          | `tensordot`         | :white_check_mark: |            |                              |
 |                          | `vecdot`            | :white_check_mark: |            |                              |
 | Manipulation Functions   | `broadcast_arrays`  | :white_check_mark: |            |                              |
+|                          | `broadcast_shapes`  | :white_check_mark: | 2025.12    |                              |
 |                          | `broadcast_to`      | :white_check_mark: |            |                              |
 |                          | `concat`            | :white_check_mark: |            |                              |
 |                          | `expand_dims`       | :white_check_mark: |            |                              |
@@ -73,7 +74,8 @@ This table shows which parts of the the [Array API](https://data-apis.org/array-
 |                          | `nonzero`           | :x:                |            | Shape is data dependent      |
 |                          | `searchsorted`      | :white_check_mark: | 2023.12    |                              |
 |                          | `where`             | :white_check_mark: |            |                              |
-| Set Functions            | `unique_all`        | :x:                |            | Shape is data dependent      |
+| Set Functions            | `isin`              | :white_check_mark: | 2025.12    |                              |
+|                          | `unique_all`        | :x:                |            | Shape is data dependent      |
 |                          | `unique_counts`     | :x:                |            | Shape is data dependent      |
 |                          | `unique_inverse`    | :x:                |            | Shape is data dependent      |
 |                          | `unique_values`     | :x:                |            | Shape is data dependent      |
@@ -94,15 +96,17 @@ This table shows which parts of the the [Array API](https://data-apis.org/array-
 
 ### Linear Algebra Extension
 
-A few of the [linear algebra extension](https://data-apis.org/array-api/2022.12/extensions/linear_algebra_functions.html) functions are supported, as indicated in this table.
+A few of the [linear algebra extension](https://data-apis.org/array-api/2025.12/extensions/linear_algebra_functions.html) functions are supported, as indicated in this table.
 
 | Category                 | Object/Function     | Implemented        | Version    | Notes                        |
 | ------------------------ | ------------------- | ------------------ | ---------- | ---------------------------- |
 | Linear Algebra Functions | `cholesky`          | :x:                |            |                              |
 |                          | `cross`             | :x:                |            |                              |
 |                          | `det`               | :x:                |            |                              |
 |                          | `diagonal`          | :x:                |            |                              |
+|                          | `eig`               | :x:                |            |                              |
 |                          | `eigh`              | :x:                |            |                              |
+|                          | `eigvals`           | :x:                |            |                              |
 |                          | `eigvalsh`          | :x:                |            |                              |
 |                          | `inv`               | :x:                |            |                              |
 |                          | `matmul`            | :white_check_mark: |            |                              |

cubed/__init__.py

@@ -52,7 +52,7 @@
 
 # Array API
 
-__array_api_version__ = "2024.12"
+__array_api_version__ = "2025.12"
 
 from .array_api.inspection import __array_namespace_info__
 
@@ -301,6 +301,7 @@
 
 from .array_api.manipulation_functions import (
     broadcast_arrays,
+    broadcast_shapes,
     broadcast_to,
     concat,
     expand_dims,
@@ -318,6 +319,7 @@
 
 __all__ += [
     "broadcast_arrays",
+    "broadcast_shapes",
     "broadcast_to",
     "concat",
     "expand_dims",

cubed/array_api/__init__.py

@@ -1,6 +1,6 @@
 __all__ = []
 
-__array_api_version__ = "2024.12"
+__array_api_version__ = "2025.12"
 
 from .inspection import __array_namespace_info__
 
@@ -236,6 +236,7 @@
 
 from .manipulation_functions import (
     broadcast_arrays,
+    broadcast_shapes,
     broadcast_to,
     concat,
     expand_dims,
@@ -253,6 +254,7 @@
 
 __all__ += [
     "broadcast_arrays",
+    "broadcast_shapes",
     "broadcast_to",
     "concat",
     "expand_dims",

cubed/array_api/array_object.py

@@ -367,6 +367,7 @@ def __array_namespace__(self, /, *, api_version=None):
             "2022.12",
             "2023.12",
             "2024.12",
+            "2025.12",
         ):
             raise ValueError(f"Unrecognized array API version: {api_version!r}")
         import cubed

cubed/array_api/creation_functions.py

@@ -267,7 +267,7 @@ def meshgrid(*arrays, indexing="xy") -> List["Array"]:
     if indexing == "xy" and len(arrs) > 1:
         grid[0], grid[1] = grid[1], grid[0]
 
-    return grid
+    return tuple(grid)
 
 
 def ones(shape, *, dtype=None, device=None, chunks="auto", spec=None) -> "Array":

cubed/array_api/indexing_functions.py

@@ -1,2 +1,7 @@
-def take(x, indices, /, *, axis):
+def take(x, indices, /, *, axis=None):
+    from cubed.array_api.manipulation_functions import flatten
+
+    if axis is None:
+        x = flatten(x)
+        axis = 0
     return x[(slice(None),) * axis + (indices,)]

cubed/array_api/manipulation_functions.py

@@ -1,7 +1,6 @@
 from bisect import bisect
 from operator import add, mul
 
-import numpy as np
 import tlz
 from toolz import reduce
 
@@ -41,14 +40,18 @@ def broadcast_arrays(*arrays):
     uc_args = tlz.concat(zip(arrays, inds))
     _, args = unify_chunks(*uc_args, warn=False)
 
-    shape = np.broadcast_shapes(*(e.shape for e in args))
+    shape = broadcast_shapes(*(e.shape for e in args))
     chunks = broadcast_chunks(*(e.chunks for e in args))
 
     result = tuple(broadcast_to(e, shape=shape, chunks=chunks) for e in args)
 
     return result
 
 
+def broadcast_shapes(*shapes):
+    return nxp.broadcast_shapes(*shapes)
+
+
 def broadcast_to(x, /, shape, *, chunks=None):
     if x.shape == shape and (chunks is None or chunks == x.chunks):
         return x
@@ -272,7 +275,7 @@ def _chunk_slices(
         arr_sel_offset += cp.out_selection[axis].stop
 
 
-def expand_dims(x, /, *, axis):
+def expand_dims(x, /, axis):
     if not isinstance(axis, tuple):
         axis = (axis,)
     ndim_new = len(axis) + x.ndim

cubed/array_api/searching_functions.py

@@ -55,6 +55,9 @@ def searchsorted(x1, x2, /, *, side="left", sorter=None):
     if x1.ndim != 1:
         raise ValueError("Input array x1 must be one dimensional")
 
+    if isinstance(x2, int | float | complex):
+        x2 = asarray(x2, spec=x1.spec)
+
     if sorter is not None:
         raise NotImplementedError(
             "searchsorted with a sorter argument is not supported"
@@ -97,7 +100,7 @@ def _searchsorted(x, y, side):
     # of telling which block is being operated on (unlike map_blocks),
     # so set all 0 values to a special value and set back at the end of searchsorted
     res = nxp.where(res == 0, -1, res)
-    return res[nxp.newaxis, :]
+    return res[nxp.newaxis, ...]
 
 
 def where(condition, x1, x2, /):

cubed/core/ops.py

@@ -591,8 +591,10 @@ def _general_blockwise(
 
 def elemwise(func, *args: "Array", dtype=None) -> "Array":
     """Apply a function elementwise to array arguments, respecting broadcasting."""
+    from cubed.array_api.manipulation_functions import broadcast_shapes
+
     shapes = [arg.shape for arg in args]
-    out_ndim = len(np.broadcast_shapes(*shapes))
+    out_ndim = len(broadcast_shapes(*shapes))
     expr_inds = tuple(range(out_ndim))[::-1]
     if dtype is None:
         raise ValueError("dtype must be specified for elemwise")

cubed/tests/test_array_api.py

@@ -581,10 +581,11 @@ def test_concat_incompatible_shapes(spec):
     xp.concat([a, b], axis=1)  # OK
 
 
-def test_expand_dims(spec, executor):
+@pytest.mark.parametrize("axis", [0, 1, (0, 1), (2, 0)])
+def test_expand_dims(spec, axis):
     a = xp.asarray([1, 2, 3], chunks=(2,), spec=spec)
-    b = xp.expand_dims(a, axis=0)
-    assert_array_equal(b.compute(executor=executor), np.expand_dims([1, 2, 3], 0))
+    b = xp.expand_dims(a, axis=axis)
+    assert_array_equal(b.compute(), np.expand_dims([1, 2, 3], axis))
 
 
 @pytest.mark.parametrize(
@@ -826,6 +827,21 @@ def test_searchsorted(x1, x1_chunks, x2, x2_chunks, side):
     assert_array_equal(out.compute(), np.searchsorted(x1, x2, side=side))
 
 
+@pytest.mark.parametrize("side", ["left", "right"])
+def test_searchsorted_scalar(side):
+    x1 = np.array([-10, 0, 10, 20, 30])
+    x2 = 11
+
+    x1d = xp.asarray(x1, chunks=3)
+    x2d = x2
+
+    out = xp.searchsorted(x1d, x2d, side=side)
+
+    assert out.shape == ()
+    assert out.chunks == ()
+    assert_array_equal(out.compute(), np.searchsorted(x1, x2, side=side))
+
+
 def test_searchsorted_sorter_not_implemented():
     with pytest.raises(NotImplementedError):
         xp.searchsorted(xp.asarray([1, 0]), xp.asarray([1]), sorter=xp.asarray([1, 0]))

cubed/vendor/dask/array/utils.py

@@ -1,7 +1,7 @@
 import numbers
 
 import numpy as np
-
+from numpy.exceptions import AxisError
 
 def validate_axis(axis, ndim):
     """Validate an input to axis= keywords"""
@@ -10,7 +10,7 @@ def validate_axis(axis, ndim):
     if not isinstance(axis, numbers.Integral):
         raise TypeError("Axis value must be an integer, got %s" % axis)
     if axis < -ndim or axis >= ndim:
-        raise np.AxisError(
+        raise AxisError(
             "Axis %d is out of bounds for array of dimension %d" % (axis, ndim)
         )
     if axis < 0:

docs/array-api.md

@@ -1,6 +1,6 @@
 # Python Array API
 
-Cubed implements version 2024.12 of the [Python Array API standard](https://data-apis.org/array-api/2024.12/index.html) in `cubed.array_api`, with a few exceptions listed on the [coverage status](https://github.com/cubed-dev/cubed/blob/main/api_status.md) page. The [Fourier transform functions](https://data-apis.org/array-api/2024.12/extensions/fourier_transform_functions.html) are *not* supported.
+Cubed implements version 2025.12 of the [Python Array API standard](https://data-apis.org/array-api/2025.12/index.html) in `cubed.array_api`, with a few exceptions listed on the [coverage status](https://github.com/cubed-dev/cubed/blob/main/api_status.md) page. The [Fourier transform functions](https://data-apis.org/array-api/2025.12/extensions/fourier_transform_functions.html) are *not* supported.
 
 ## Differences between Cubed and the standard