FIX avoid torch dlpack crash for negative strides (scikit-learn#34380)

Co-authored-by: Kevin-Li-2025 <2242139@qq.com>
Co-authored-by: Virgil Chan <virchan.math@gmail.com>
Co-authored-by: Olivier Grisel <olivier.grisel@ensta.org>

Author: 4 people

doc/whats_new/upcoming_changes/sklearn.utils/34380.fix.rst

@@ -0,0 +1,5 @@
+- Array API supporting functions and estimators that handle mixed input
+  namespaces now make negative-stride NumPy arrays contiguous before
+  transferring them to PyTorch through DLPack, avoiding a Python process
+  abort in ``torch.from_dlpack``.
+  By :user:`Yin Li <Kevin-Li-2025>`.

sklearn/utils/_array_api.py

@@ -580,6 +580,13 @@ def move_to(*arrays, xp, device):
             if xp == xp_array and device == device_array:
                 converted_arrays.append(array)
             else:
+                if _is_xp_namespace(xp, "torch") and _is_numpy_namespace(xp_array):
+                    if any(stride < 0 for stride in array.strides):
+                        # Work around PyTorch aborting the process when importing
+                        # negative-strided NumPy arrays with DLPack. Remove this once
+                        # https://github.com/pytorch/pytorch/issues/188023 is fixed.
+                        # See also https://github.com/scikit-learn/scikit-learn/issues/34307
+                        array = numpy.ascontiguousarray(array)
                 try:
                     # The dlpack protocol is the future proof and library agnostic
                     # method to transfer arrays across namespace and device boundaries

sklearn/utils/tests/test_array_api.py

@@ -53,6 +53,7 @@
     _array_api_for_tests,
     _convert_container,
     assert_array_equal,
+    assert_run_python_script_without_output,
     skip_if_array_api_compat_not_configured,
 )
 from sklearn.utils.fixes import _IS_32BIT, CSR_CONTAINERS, np_version, parse_version
@@ -192,6 +193,34 @@ def test_move_to_sparse():
             move_to(sparse1, None, xp=xp_torch, device=device_cpu)
 
 
+def test_move_to_numpy_negative_strides_to_torch():
+    """Check NumPy arrays with negative strides can be moved to torch."""
+    pytest.importorskip("torch")
+
+    code = """
+import os
+
+os.environ["SCIPY_ARRAY_API"] = "1"
+
+import numpy
+from numpy.testing import assert_allclose
+
+from sklearn._config import config_context
+from sklearn.utils._array_api import get_namespace_and_device, move_to
+
+import torch
+
+a = numpy.arange(12.0).reshape(3, 4)[:, ::-1]
+with config_context(array_api_dispatch=True):
+    xp, _, device = get_namespace_and_device(torch.asarray([1.0]))
+    result = move_to(a, xp=xp, device=device)
+    assert_allclose(result.cpu().numpy(), a)
+"""
+    # This must run in a subprocess because old PyTorch versions abort the
+    # Python process before a Python exception can be raised.
+    assert_run_python_script_without_output(code)
+
+
 @pytest.mark.parametrize("array_api", ["numpy", "array_api_strict"])
 def test_asarray_with_order(array_api):
     """Test _asarray_with_order passes along order for NumPy arrays."""