diff --git a/unyt/array.py b/unyt/array.py
index fb5c0fde..6317732a 100644
--- a/unyt/array.py
+++ b/unyt/array.py
@@ -505,6 +505,25 @@ def wrapper(self, ufunc, method, *inputs, **kwargs):
 LARGE_INPUT = {4: 16777217, 8: 9007199254740993}
 
 
+def _update_array_dtype_inplace(array, *, dtype) -> None:
+    # change the dtypes in-place, this does not change the
+    # underlying memory buffer
+    dt = np.dtype(dtype)
+    assert dt.kind == "f"
+    if NUMPY_VERSION >= Version("2.5.0dev0"):
+        array._set_dtype(dt)
+    else:
+        array.dtype = dt
+
+
+def _update_array_values_inplace(array, *, dtype) -> None:
+    new_values = array.astype(dtype)
+    _update_array_dtype_inplace(array, dtype=dtype)
+    # actually fill in the new float values now that our
+    # dtype is correct
+    np.copyto(array, new_values)
+
+
 class unyt_array(np.ndarray):
     """
     An ndarray subclass that attaches a symbolic unit object to the array data.
@@ -795,7 +814,6 @@ def convert_to_units(self, units, equivalence=None, **kwargs):
                         f"Input dtype ({self.dtype}) has a smaller itemsize than the "
                         "smallest floating point representation possible."
                     )
-                new_dtype = "f" + str(dsize)
                 large = LARGE_INPUT.get(dsize, 0)
                 if large and np.any(np.abs(values) > large):
                     warnings.warn(
@@ -803,14 +821,8 @@ def convert_to_units(self, units, equivalence=None, **kwargs):
                         RuntimeWarning,
                         stacklevel=2,
                     )
-                float_values = values.astype(new_dtype)
-                # change the dtypes in-place, this does not change the
-                # underlying memory buffer
-                values.dtype = new_dtype
-                self.dtype = new_dtype
-                # actually fill in the new float values now that our
-                # dtype is correct
-                np.copyto(values, float_values)
+                _update_array_values_inplace(values, dtype=f"f{dsize}")
+                _update_array_dtype_inplace(self, dtype=values.dtype)
             values *= conv_factor
 
             if offset:
@@ -1884,10 +1896,7 @@ def __array_ufunc__(self, ufunc, method, *inputs, **kwargs):
             else:
                 out = out[0]
                 if out.dtype.kind in ("u", "i"):
-                    new_dtype = "f" + str(out.dtype.itemsize)
-                    float_values = out.astype(new_dtype)
-                    out.dtype = new_dtype
-                    np.copyto(out, float_values)
+                    _update_array_values_inplace(out, dtype=f"f{out.dtype.itemsize}")
                 out_func = out.view(np.ndarray)
         if len(inputs) == 1:
             # Unary ufuncs