Use _raise_if_invalid_cast and cover both same-kind out dtypes

cupy/_core/_routines_indexing.pyx

@@ -14,6 +14,7 @@ from libcpp cimport vector
 
 from cupy._core._carray cimport shape_t
 from cupy._core._carray cimport strides_t
+from cupy._core._dtype cimport _raise_if_invalid_cast
 from cupy._core cimport core
 from cupy._core cimport _routines_math as _math
 from cupy._core cimport _routines_manipulation as _manipulation
@@ -576,13 +577,13 @@ out = a[out_i];
 _take_kernel = ElementwiseKernel(
     'raw T a, S indices, int64 ldim_, int64 cdim_, int64 rdim_, '
     'int64 index_range_',
-    'T out', _take_kernel_core, 'cupy_take')
+    'U out', _take_kernel_core, 'cupy_take')
 
 
 _take_kernel_scalar = ElementwiseKernel(
     'raw T a, int64 indices, int64 ldim_, int64 cdim_, int64 rdim_, '
     'int64 index_range_',
-    'T out', _take_kernel_core, 'cupy_take_scalar')
+    'U out', _take_kernel_core, 'cupy_take_scalar')
 
 
 _choose_kernel = ElementwiseKernel(
@@ -881,8 +882,8 @@ cdef _ndarray_base _take(
     if out is None:
         out = core.ndarray(out_shape, dtype=a.dtype)
     else:
-        if out.dtype != a.dtype:
-            raise TypeError('Output dtype mismatch')
+        _raise_if_invalid_cast(
+            a.dtype, out.dtype, 'same_kind', 'output operand')
         if out.shape != out_shape:
             raise ValueError('Output shape mismatch')
     if a.size == 0 and out.size != 0:

tests/cupy_tests/core_tests/test_ndarray.py

@@ -507,22 +507,39 @@ def test_shape_mismatch(self):
     {'shape': (3, 4, 5), 'indices': (2, 3), 'out_shape': (2, 3)},
     {'shape': (), 'indices': (), 'out_shape': ()}
 )
-class TestNdarrayTakeErrorTypeMismatch(unittest.TestCase):
+class TestNdarrayTakeTypeMismatch(unittest.TestCase):
+    # NOTE(seberg): Historically cupy was always fully restrictive
+    # while NumPy was just wrong: https://github.com/numpy/numpy/pull/30615
+    # As of CuPy 14.2, CuPy uses 2.5+ (future) behavior.
 
-    def test_output_type_mismatch(self):
+    @testing.with_requires('numpy>=2.5')
+    @testing.numpy_cupy_array_equal()
+    # incorrectly given by numpy (presumably until Deprecation is finalized)
+    @pytest.mark.filterwarnings('ignore::numpy.exceptions.ComplexWarning')
+    def test_output_dtype_same_kind_ok(self, xp):
+        # After NumPy 2.5, NumPy gets the cast safety right, the following
+        # is OK under same-kind casting rules.
+        a = testing.shaped_arange(self.shape, xp, numpy.int64)
+        i = testing.shaped_arange(self.indices, xp, numpy.int32) % 3
+        results = []
+        for out_dtype in (numpy.complex64, numpy.int32):
+            o = testing.shaped_arange(self.out_shape, xp, out_dtype)
+            results.append(wrap_take(a, i, out=o))
+        return results
+
+    @pytest.mark.filterwarnings(
+        'error:Implicit casting of output dtype:DeprecationWarning')
+    def test_output_dtype_unsafe_rejected(self):
         for xp in (numpy, cupy):
-            a = testing.shaped_arange(self.shape, xp, numpy.int32)
+            a = testing.shaped_arange(self.shape, xp, numpy.float32)
             i = testing.shaped_arange(self.indices, xp, numpy.int32) % 3
-            o = testing.shaped_arange(self.out_shape, xp, numpy.float32)
-            if (xp is numpy
-                    and numpy.lib.NumpyVersion(numpy.__version__)
-                    >= '2.5.0'):
-                # numpy>=2.5 permits safe casting of the take() out= dtype,
-                # so it no longer raises; cupy still enforces a match.
+            o = testing.shaped_arange(self.out_shape, xp, numpy.int32)
+            # As of NumPy 2.5 this is a deprecation warning, but CuPy never
+            # allowed it (so no deprecation required)
+            if xp is numpy and not testing.numpy_satisfies('>=2.5'):
+                continue
+            with pytest.raises((TypeError, DeprecationWarning)):
                 wrap_take(a, i, out=o)
-            else:
-                with pytest.raises(TypeError):
-                    wrap_take(a, i, out=o)
 
 
 @testing.parameterize(