Acquire the GIL when releasing NumPy-backed Core ML inputs

coremlpython/CoreMLPythonArray.mm

@@ -65,4 +65,15 @@ - (PybindCompatibleArray *)initWithArray:(py::array)array {
     return self;
 }
 
+- (void)dealloc {
+    // Core ML may release the multi-array on one of its private queues. Clear
+    // the Python owner while holding the GIL so py::array does not decrement
+    // its reference count from a non-Python thread.
+    py::handle array = m_array.release();
+    if (array) {
+        py::gil_scoped_acquire gil;
+        array.dec_ref();
+    }
+}
+
 @end

coremltools/test/api/test_python_bindings.py

@@ -0,0 +1,83 @@
+# Copyright (c) 2026, Apple Inc. All rights reserved.
+#
+# Use of this source code is governed by a BSD-3-clause license that can be
+# found in the LICENSE.txt file or at https://opensource.org/licenses/BSD-3-Clause
+
+import subprocess
+import sys
+
+import pytest
+
+import coremltools as ct
+from coremltools.converters.mil import Builder as mb
+
+
+_NUMPY_OWNER_RELEASE_SCRIPT = """
+import ctypes
+import sys
+import time
+import weakref
+
+import coremltools as ct
+import numpy as np
+
+model = ct.models.MLModel(sys.argv[1], compute_units=ct.ComputeUnit.CPU_ONLY)
+input_name = model.get_spec().description.input[0].name
+
+check_gil = ctypes.pythonapi.PyGILState_Check
+check_gil.restype = ctypes.c_int
+gil_states = []
+
+input_array = np.zeros((1, 4), dtype=np.float32)
+input_ref = weakref.ref(
+    input_array,
+    lambda _: gil_states.append(bool(check_gil())),
+)
+model.predict({input_name: input_array})
+del input_array
+
+deadline = time.monotonic() + 10
+while input_ref() is not None and time.monotonic() < deadline:
+    time.sleep(0.05)
+
+assert input_ref() is None, "Core ML did not release the NumPy input owner"
+assert gil_states == [True], f"NumPy input owner release GIL states: {gil_states}"
+"""
+
+
+@pytest.mark.skipif(
+    ct.utils._macos_version() < (12, 0),
+    reason="ML Program prediction is available only on macOS 12+",
+)
+def test_numpy_input_owner_is_released_with_gil(tmp_path):
+    @mb.program(
+        input_specs=[mb.TensorSpec(shape=(1, 4))],
+        opset_version=ct.target.iOS16,
+    )
+    def program(x):
+        return mb.add(x=x, y=1.0)
+
+    model = ct.convert(
+        program,
+        convert_to="mlprogram",
+        compute_precision=ct.precision.FLOAT32,
+    )
+    model_path = tmp_path / "add_one.mlpackage"
+    model.save(str(model_path))
+
+    # Core ML may release the NumPy owner asynchronously. Use a subprocess so
+    # an unsafe release is reported as a test failure instead of terminating
+    # the entire pytest process.
+    result = subprocess.run(
+        [sys.executable, "-c", _NUMPY_OWNER_RELEASE_SCRIPT, str(model_path)],
+        capture_output=True,
+        check=False,
+        text=True,
+        timeout=30,
+    )
+
+    assert result.returncode == 0, (
+        f"Prediction subprocess exited with {result.returncode}\n"
+        f"stdout:\n{result.stdout}\n"
+        f"stderr:\n{result.stderr}"
+    )