[Eager] Avoid redundant memcpy in to_tensor (#68164)

SigureMo · web-flow · commit 5e0c20842385 · 2024-09-12T12:08:03.000+08:00
diff --git a/python/paddle/tensor/creation.py b/python/paddle/tensor/creation.py
@@ -726,16 +726,11 @@ def _handle_np_dtype(
             data = _handle_tensor_dtype(data, dtype)
             data.stop_gradient = stop_gradient
             return data
-        elif isinstance(data, (core.LoDTensor, core.Tensor)):
+        elif isinstance(data, core.Tensor):
             # should't expose it to users, just for internal use.
             # convert core.Tensor/core.LoDTensor to Tensor first
             # Currently, there is no copy when places are same
-            if in_dynamic_mode():
-                data = core.eager.Tensor(data)
-            else:
-                data = paddle.Tensor(data)
-            if not data.place._equals(place):
-                data = data._copy_to(place, False)
+            data = paddle.Tensor(data, place=place)
             data = _handle_tensor_dtype(data, dtype)
             data.stop_gradient = stop_gradient
             return data
diff --git a/test/legacy_test/test_eager_tensor.py b/test/legacy_test/test_eager_tensor.py
@@ -1534,5 +1534,18 @@ def test_eager_tensor_grad_name_value(self):
         self.assertIsNotNone(a._grad_value())
 
 
+class TestDenseTensorToTensor(unittest.TestCase):
+    def test_same_place_data_ptr_consistency(self):
+        places = [paddle.CPUPlace()]
+        if paddle.is_compiled_with_cuda():
+            places.append(paddle.CUDAPlace(0))
+        for place in places:
+            x = paddle.rand([3, 5]).to(device=place)
+            x_dense = x.get_tensor()
+            y = paddle.to_tensor(x_dense, place=place)
+
+            self.assertEqual(x.data_ptr(), y.data_ptr())
+
+
 if __name__ == "__main__":
     unittest.main()
