broadcast_arrays et al: change the return type into a tuple

array_api_compat/cupy/_aliases.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 from builtins import bool as py_bool
+from typing import Literal
 
 import cupy as cp
 
@@ -139,6 +140,15 @@ def take_along_axis(x: Array, indices: Array, /, *, axis: int = -1) -> Array:
     return cp.take_along_axis(x, indices, axis=axis)
 
 
+# https://github.com/cupy/cupy/pull/9582
+def broadcast_arrays(*arrays: Array) -> tuple[Array, ...]:
+    return tuple(cp.broadcast_arrays(*arrays))
+
+
+def meshgrid(*arrays: Array, indexing: Literal['xy', 'ij'] = 'xy') -> tuple[Array, ...]:
+    return tuple(cp.meshgrid(*arrays, indexing=indexing))
+
+
 # These functions are completely new here. If the library already has them
 # (i.e., numpy 2.0), use the library version instead of our wrapper.
 if hasattr(cp, 'vecdot'):
@@ -161,7 +171,8 @@ def take_along_axis(x: Array, indices: Array, /, *, axis: int = -1) -> Array:
                               'atan2', 'atanh', 'bitwise_left_shift',
                               'bitwise_invert', 'bitwise_right_shift',
                               'bool', 'concat', 'count_nonzero', 'pow', 'sign',
-                              'ceil', 'floor', 'trunc', 'take_along_axis']
+                              'ceil', 'floor', 'trunc', 'take_along_axis',
+                              'broadcast_arrays', 'meshgrid']
 
 
 def __dir__() -> list[str]:

array_api_compat/cupy/_info.py

@@ -333,4 +333,4 @@ def devices(self):
         __array_namespace_info__.dtypes
 
         """
-        return [cuda.Device(i) for i in range(cuda.runtime.getDeviceCount())]
+        return tuple(cuda.Device(i) for i in range(cuda.runtime.getDeviceCount()))

array_api_compat/dask/array/_info.py

@@ -379,7 +379,7 @@ def dtypes(
             return res
         raise ValueError(f"unsupported kind: {kind!r}")
 
-    def devices(self) -> list[Device]:
+    def devices(self) -> tuple[Device]:
         """
         The devices supported by Dask.
 
@@ -404,4 +404,4 @@ def devices(self) -> list[Device]:
         ['cpu', DASK_DEVICE]
 
         """
-        return ["cpu", _DASK_DEVICE]
+        return ("cpu", _DASK_DEVICE)

array_api_compat/numpy/_info.py

@@ -332,7 +332,7 @@ def dtypes(
             return res
         raise ValueError(f"unsupported kind: {kind!r}")
 
-    def devices(self) -> list[Device]:
+    def devices(self) -> tuple[Device]:
         """
         The devices supported by NumPy.
 
@@ -357,7 +357,7 @@ def devices(self) -> list[Device]:
         ['cpu']
 
         """
-        return ["cpu"]
+        return ("cpu",)
 
 
 __all__ = ["__array_namespace_info__"]

array_api_compat/torch/_aliases.py

@@ -711,9 +711,9 @@ def astype(
     return x.to(dtype=dtype, copy=copy)
 
 
-def broadcast_arrays(*arrays: Array) -> list[Array]:
+def broadcast_arrays(*arrays: Array) -> tuple[Array, ...]:
     shape = torch.broadcast_shapes(*[a.shape for a in arrays])
-    return [torch.broadcast_to(a, shape) for a in arrays]
+    return tuple(torch.broadcast_to(a, shape) for a in arrays)
 
 # Note that these named tuples aren't actually part of the standard namespace,
 # but I don't see any issue with exporting the names here regardless.
@@ -897,10 +897,11 @@ def sign(x: Array, /) -> Array:
         return out
 
 
-def meshgrid(*arrays: Array, indexing: Literal['xy', 'ij'] = 'xy') -> list[Array]:
-    # enforce the default of 'xy'
-    # TODO: is the return type a list or a tuple
-    return list(torch.meshgrid(*arrays, indexing=indexing))
+def meshgrid(*arrays: Array, indexing: Literal['xy', 'ij'] = 'xy') -> tuple[Array, ...]:
+    # torch <= 2.9 emits a UserWarning: "torch.meshgrid: in an upcoming release, it
+    # will be required to pass the indexing argument."
+    # Thus always pass it explicitly.
+    return torch.meshgrid(*arrays, indexing=indexing)
 
 
 __all__ = ['asarray', 'result_type', 'can_cast',

array_api_compat/torch/_info.py

@@ -366,4 +366,4 @@ def devices(self):
                     break
                 i += 1
 
-        return devices
+        return tuple(devices)

dask-xfails.txt

@@ -129,6 +129,13 @@ array_api_tests/test_linalg.py::test_matrix_norm
 array_api_tests/test_linalg.py::test_qr
 array_api_tests/test_manipulation_functions.py::test_roll
 
+# 2025.12 support
+array_api_tests/test_has_names.py::test_has_names[manipulation-broadcast_shapes]
+array_api_tests/test_signatures.py::test_func_signature[broadcast_shapes]
+array_api_tests/test_data_type_functions.py::TestBroadcastShapes::test_broadcast_shapes
+array_api_tests/test_data_type_functions.py::TestBroadcastShapes::test_empty
+array_api_tests/test_data_type_functions.py::TestBroadcastShapes::test_error
+
 # Stubs have a comment: (**note**: libraries may return ``NaN`` to match Python behavior.)
 array_api_tests/test_special_cases.py::test_binary[floor_divide(x1_i is +infinity and isfinite(x2_i) and x2_i > 0) -> +infinity]
 array_api_tests/test_special_cases.py::test_binary[floor_divide(x1_i is +infinity and isfinite(x2_i) and x2_i < 0) -> -infinity]

numpy-1-22-xfails.txt

@@ -152,6 +152,13 @@ array_api_tests/test_signatures.py::test_func_signature[bitwise_right_shift]
 array_api_tests/test_signatures.py::test_func_signature[bitwise_xor]
 array_api_tests/test_data_type_functions.py::TestResultType::test_with_scalars
 
+# 2025.12 support
+
+# older numpies return lists not tuples
+array_api_tests/test_creation_functions.py::test_meshgrid
+array_api_tests/test_data_type_functions.py::test_broadcast_arrays
+
+
 # Stubs have a comment: (**note**: libraries may return ``NaN`` to match Python behavior.); Apparently,NumPy does just that
 array_api_tests/test_special_cases.py::test_binary[floor_divide(x1_i is +infinity and isfinite(x2_i) and x2_i > 0) -> +infinity]
 array_api_tests/test_special_cases.py::test_binary[floor_divide(x1_i is +infinity and isfinite(x2_i) and x2_i < 0) -> -infinity]

numpy-1-26-xfails.txt

@@ -42,9 +42,15 @@ array_api_tests/test_signatures.py::test_array_method_signature[__dlpack__]
 
 # 2024.12 support
 array_api_tests/test_data_type_functions.py::TestResultType::test_with_scalars
-
 array_api_tests/test_operators_and_elementwise_functions.py::test_where_with_scalars
 
+# 2025.12 support
+
+# older numpies return lists not tuples
+array_api_tests/test_creation_functions.py::test_meshgrid
+array_api_tests/test_data_type_functions.py::test_broadcast_arrays
+
+
 # Stubs have a comment: (**note**: libraries may return ``NaN`` to match Python behavior.); Apparently, NumPy does just that
 array_api_tests/test_special_cases.py::test_binary[floor_divide(x1_i is +infinity and isfinite(x2_i) and x2_i > 0) -> +infinity]
 array_api_tests/test_special_cases.py::test_binary[floor_divide(x1_i is +infinity and isfinite(x2_i) and x2_i < 0) -> -infinity]

torch-xfails.txt

@@ -144,6 +144,11 @@ array_api_tests/test_signatures.py::test_func_signature[from_dlpack]
 # Argument 'max_version' missing from signature
 array_api_tests/test_signatures.py::test_array_method_signature[__dlpack__]
 
+# 2025.12 support
+
+# broadcast_shapes emits a RuntimeError where the spec says ValueError
+array_api_tests/test_data_type_functions.py::TestBroadcastShapes::test_error
+
 
 # 2024.12 support: binary functions reject python scalar arguments
 array_api_tests/test_operators_and_elementwise_functions.py::test_binary_with_scalars_real[atan2]