Fix UVA Tensor (#46371)

paddle/fluid/pybind/tensor_py.h

@@ -608,9 +608,10 @@ void SetStringTensorFromPyArray(phi::StringTensor *self,
 }
 
 template <typename T>
-void SetUVATensorFromPyArrayImpl(framework::LoDTensor *self_tensor,
-                                 const py::array_t<T> &array,
-                                 int device_id) {
+void SetUVATensorFromPyArrayImpl(
+    framework::LoDTensor *self_tensor,
+    const py::array_t<T, py::array::c_style | py::array::forcecast> &array,
+    int device_id) {
 #if defined(PADDLE_WITH_CUDA)
   VLOG(4) << "Running in SetUVATensorFromPyArrayImpl.";
   std::vector<int64_t> dims;
@@ -647,7 +648,7 @@ void SetUVATensorFromPyArrayImpl(framework::LoDTensor *self_tensor,
 template <typename T>
 void SetUVATensorFromPyArray(
     const std::shared_ptr<paddle::imperative::VarBase> &self,
-    const py::array_t<T> &array,
+    const py::array_t<T, py::array::c_style | py::array::forcecast> &array,
     int device_id) {
 #if defined(PADDLE_WITH_CUDA)
   VLOG(4) << "Running in SetUVATensorFromPyArray for VarBase.";

python/paddle/fluid/tests/unittests/test_tensor_uva.py

@@ -57,6 +57,17 @@ class TestUVATensorFromNumpy(unittest.TestCase):
                 np.testing.assert_allclose(tensor.numpy(), data, rtol=1e-05)
                 np.testing.assert_allclose(tensor2.numpy(), data, rtol=1e-05)
 
+    def test_uva_tensor_corectness(self):
+        if paddle.fluid.core.is_compiled_with_cuda():
+            a = np.arange(0, 100, dtype="int32")
+            a = a.reshape([10, 10])
+            slice_a = a[:, 5]
+            tensor1 = paddle.to_tensor(slice_a)
+            tensor2 = core.eager.to_uva_tensor(slice_a)
+            np.testing.assert_allclose(tensor1.numpy(),
+                                       tensor2.numpy(),
+                                       rtol=1e-05)
+
     def test_uva_tensor_creation(self):
         with _test_eager_guard():
             self.func_uva_tensor_creation()