fix(cuda.core): preserve DLPack C exchange errors

cuda_core/cuda/core/_dlpack.pxd

@@ -1,4 +1,4 @@
-# SPDX-FileCopyrightText: Copyright (c) 2024-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-FileCopyrightText: Copyright (c) 2024-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 #
 # SPDX-License-Identifier: Apache-2.0
 
@@ -90,26 +90,28 @@ cdef extern from "_include/dlpack.h":
         void (*SetError)(void* error_ctx, const char* kind, const char* message) noexcept
     )
 
+    # dlpack.h: these entry points return -1 on failure with a Python
+    # exception set. `except -1` tells Cython so; it does not change the C type.
     ctypedef int (*DLPackManagedTensorFromPyObjectNoSync)(
         void* py_object,
         DLManagedTensorVersioned** out
-    )
+    ) except -1
 
     ctypedef int (*DLPackManagedTensorToPyObjectNoSync)(
         DLManagedTensorVersioned* tensor,
         void** out_py_object
-    )
+    ) except -1
 
     ctypedef int (*DLPackDLTensorFromPyObjectNoSync)(
         void* py_object,
         DLTensor* out
-    )
+    ) except -1
 
     ctypedef int (*DLPackCurrentWorkStream)(
         _DLDeviceType device_type,
         int32_t device_id,
         void** out_current_stream
-    )
+    ) except -1
 
     ctypedef struct DLPackExchangeAPIHeader:
         DLPackVersion version

cuda_core/cuda/core/_memoryview.pyx

@@ -899,79 +899,65 @@ cdef int _smv_managed_tensor_allocator(
 cdef int _smv_managed_tensor_from_py_object_no_sync(
     void* py_object,
     DLManagedTensorVersioned** out,
-) noexcept with gil:
+) except -1 with gil:
     cdef DLManagedTensorVersioned* dlm_tensor_ver = NULL
     if out == NULL:
-        cpython.PyErr_SetString(RuntimeError, b"out cannot be NULL")
-        return -1
+        raise RuntimeError("out cannot be NULL")
     out[0] = NULL
     cdef object obj = <object>py_object
     if not isinstance(obj, StridedMemoryView):
-        cpython.PyErr_SetString(TypeError, b"py_object must be a StridedMemoryView")
-        return -1
+        raise TypeError("py_object must be a StridedMemoryView")
     try:
         dlm_tensor_ver = _smv_allocate_dlm_tensor_versioned()
         _smv_fill_managed_tensor_versioned(dlm_tensor_ver, <StridedMemoryView>obj)
-    except Exception:
+    except BaseException:
+        # Rollback must also run for KeyboardInterrupt.
         _smv_versioned_deleter(dlm_tensor_ver)
-        return -1
+        raise
     out[0] = dlm_tensor_ver
     return 0
 
 
 cdef int _smv_managed_tensor_to_py_object_no_sync(
     DLManagedTensorVersioned* tensor,
     void** out_py_object,
-) noexcept with gil:
-    cdef object capsule
-    cdef object py_view
+) except -1 with gil:
     if out_py_object == NULL:
-        cpython.PyErr_SetString(RuntimeError, b"out_py_object cannot be NULL")
-        return -1
+        raise RuntimeError("out_py_object cannot be NULL")
     out_py_object[0] = NULL
     if tensor == NULL:
-        cpython.PyErr_SetString(RuntimeError, b"tensor cannot be NULL")
-        return -1
-    try:
-        capsule = cpython.PyCapsule_New(
-            <void*>tensor,
-            DLPACK_VERSIONED_TENSOR_UNUSED_NAME,
-            _smv_pycapsule_deleter,
-        )
-        py_view = _smv_from_dlpack_capsule(capsule, capsule)
-        cpython.Py_INCREF(py_view)
-        out_py_object[0] = <void*>py_view
-    except Exception:
-        return -1
+        raise RuntimeError("tensor cannot be NULL")
+    cdef object capsule = cpython.PyCapsule_New(
+        <void*>tensor,
+        DLPACK_VERSIONED_TENSOR_UNUSED_NAME,
+        _smv_pycapsule_deleter,
+    )
+    cdef object py_view = _smv_from_dlpack_capsule(capsule, capsule)
+    cpython.Py_INCREF(py_view)
+    out_py_object[0] = <void*>py_view
     return 0
 
 
 cdef int _smv_dltensor_from_py_object_no_sync(
     void* py_object,
     DLTensor* out,
-) noexcept with gil:
+) except -1 with gil:
     if out == NULL:
-        cpython.PyErr_SetString(RuntimeError, b"out cannot be NULL")
-        return -1
+        raise RuntimeError("out cannot be NULL")
     cdef object obj = <object>py_object
     if not isinstance(obj, StridedMemoryView):
-        cpython.PyErr_SetString(TypeError, b"py_object must be a StridedMemoryView")
-        return -1
-    try:
-        _smv_setup_dltensor_borrowed(out, <StridedMemoryView>obj)
-    except Exception:
-        return -1
+        raise TypeError("py_object must be a StridedMemoryView")
+    _smv_setup_dltensor_borrowed(out, <StridedMemoryView>obj)
     return 0
 
 
 cdef int _smv_current_work_stream(
     _DLDeviceType device_type,
     int32_t device_id,
     void** out_current_stream,
-) noexcept with gil:
+) except -1 with gil:
     if out_current_stream == NULL:
-        cpython.PyErr_SetString(RuntimeError, b"out_current_stream cannot be NULL")
-        return -1
+        raise RuntimeError("out_current_stream cannot be NULL")
     # cuda.core has no global/current stream state today.
     out_current_stream[0] = NULL
     return 0

cuda_core/docs/source/release/1.3.0-notes.rst

@@ -169,3 +169,7 @@ Fixes and enhancements
   :attr:`graph.GraphNode.succ` views with an object that is not a
   :class:`~graph.GraphNode` is now a no-op instead of raising ``TypeError``,
   consistent with :class:`collections.abc.MutableSet` semantics.
+
+- The ``StridedMemoryView`` DLPack C exchange API now preserves the underlying
+  Python exception when an export or import fails, as required by the DLPack
+  contract, instead of returning ``-1`` with no exception set.

cuda_core/tests/test_utils_dlpack.py

@@ -501,6 +501,33 @@ def test_dlpack_c_exchange_api_dltensor_from_py_object_scalar():
     assert not out.strides
 
 
+@pytest.mark.agent_authored(model="gpt-5.6-sol")
+def test_dlpack_c_exchange_api_managed_tensor_export_failure_sets_exception():
+    """A view that DLPack cannot describe reports its BufferError through the
+    managed-tensor export entry point and leaves the output tensor NULL."""
+    api = _get_exchange_api()
+    swapped = np.dtype(np.int32).newbyteorder("S")
+    view = StridedMemoryView.from_array_interface(np.zeros(3, dtype=swapped))
+    out = ctypes.c_void_p(123)
+
+    with pytest.raises(BufferError, match="Non-native-endian"):
+        api.managed_tensor_from_py_object_no_sync(id(view), ctypes.byref(out))
+    assert not out.value
+
+
+@pytest.mark.agent_authored(model="gpt-5.6-sol")
+def test_dlpack_c_exchange_api_dltensor_export_failure_sets_exception():
+    """A view that DLPack cannot describe reports its BufferError through the
+    borrowed DLTensor export entry point."""
+    api = _get_exchange_api()
+    swapped = np.dtype(np.int32).newbyteorder("S")
+    view = StridedMemoryView.from_array_interface(np.zeros(3, dtype=swapped))
+    out = _DLTensor()
+
+    with pytest.raises(BufferError, match="Non-native-endian"):
+        api.dltensor_from_py_object_no_sync(id(view), ctypes.byref(out))
+
+
 def test_dlpack_c_exchange_api_managed_tensor_roundtrip():
     """``managed_tensor_from_py_object_no_sync`` produces a managed tensor that
     ``managed_tensor_to_py_object_no_sync`` turns back into a StridedMemoryView.
@@ -562,6 +589,21 @@ def test_dlpack_c_exchange_api_to_py_object_null_tensor():
     assert not out_obj.value  # set to NULL before the error
 
 
+@pytest.mark.agent_authored(model="gpt-5.6-sol")
+def test_dlpack_c_exchange_api_import_failure_sets_exception():
+    """Rejecting an unsupported DLPack device reports the underlying BufferError."""
+    api = _get_exchange_api()
+    tensor = _DLManagedTensorVersioned()
+    tensor.version = _DLPackVersion(1, 0)
+    tensor.dl_tensor.device = _DLDevice(7, 0)  # kDLVulkan is unsupported by cuda.core
+    tensor.dl_tensor.dtype = _DLDataType(0, 32, 1)
+    out_obj = ctypes.c_void_p(123)
+
+    with pytest.raises(BufferError, match="device not supported"):
+        api.managed_tensor_to_py_object_no_sync(ctypes.byref(tensor), ctypes.byref(out_obj))
+    assert not out_obj.value
+
+
 @pytest.mark.parametrize(
     "device_type",
     [