Merge pull request #25919 from mhvk/logspace-broadcast-only-arrays

BUG: Ensure non-array logspace base does not influence dtype of output.

numpy/_core/function_base.py

@@ -298,13 +298,16 @@ def logspace(start, stop, num=50, endpoint=True, base=10.0, dtype=None,
     >>> plt.show()
 
     """
-    ndmax = np.broadcast(start, stop, base).ndim
-    start, stop, base = (
-        np.array(a, copy=None, subok=True, ndmin=ndmax)
-        for a in (start, stop, base)
-    )
+    if not isinstance(base, (float, int)) and np.ndim(base):
+        # If base is non-scalar, broadcast it with the others, since it
+        # may influence how axis is interpreted.
+        ndmax = np.broadcast(start, stop, base).ndim
+        start, stop, base = (
+            np.array(a, copy=None, subok=True, ndmin=ndmax)
+            for a in (start, stop, base)
+        )
+        base = np.expand_dims(base, axis=axis)
     y = linspace(start, stop, num=num, endpoint=endpoint, axis=axis)
-    base = np.expand_dims(base, axis=axis)
     if dtype is None:
         return _nx.power(base, y)
     return _nx.power(base, y).astype(dtype, copy=False)

numpy/_core/tests/test_regression.py

@@ -2598,3 +2598,19 @@ def test_replace_regression(self):
         expected = np.char.chararray((2,), itemsize=25)
         expected[:] = [s.replace(b"E", b"D") for s in test_strings]
         assert_array_equal(out, expected)
+
+    def test_logspace_base_does_not_determine_dtype(self):
+        # gh-24957 and cupy/cupy/issues/7946
+        start = np.array([0, 2], dtype=np.float16)
+        stop = np.array([2, 0], dtype=np.float16)
+        out = np.logspace(start, stop, num=5, axis=1, dtype=np.float32)
+        expected = np.array([[1., 3.1621094, 10., 31.625, 100.],
+                             [100., 31.625, 10., 3.1621094, 1.]],
+                            dtype=np.float32)
+        assert_almost_equal(out, expected)
+        # Check test fails if the calculation is done in float64, as happened
+        # before when a python float base incorrectly influenced the dtype.
+        out2 = np.logspace(start, stop, num=5, axis=1, dtype=np.float32,
+                           base=np.array([10.0]))
+        with pytest.raises(AssertionError, match="not almost equal"):
+            assert_almost_equal(out2, expected)