diff --git a/paddle/fluid/pybind/eager_method.cc b/paddle/fluid/pybind/eager_method.cc
index a8c1da2a8b8663f661d6a037afa2c7147ac5ebe7..46b56f27ff98e36f4e86a27a598f51dc7cbf63b4 100644
--- a/paddle/fluid/pybind/eager_method.cc
+++ b/paddle/fluid/pybind/eager_method.cc
@@ -189,7 +189,7 @@ static PyObject* eager_tensor__clear_gradient(EagerTensorObject* self,
             << " is initialized, will be released.";
     auto dense_tensor =
         std::dynamic_pointer_cast<pten::DenseTensor>(grad->impl());
-    dense_tensor->release();
+    dense_tensor->MoveMemoryHolder();
   }
   Py_INCREF(Py_None);
   return Py_None;
diff --git a/paddle/pten/api/lib/utils/tensor_utils.cc b/paddle/pten/api/lib/utils/tensor_utils.cc
index 0b6cb8d95cc1a5721caef9de05b97573e195d6bd..53d641896e43f8ba260b054f1469dab1c32cc635 100644
--- a/paddle/pten/api/lib/utils/tensor_utils.cc
+++ b/paddle/pten/api/lib/utils/tensor_utils.cc
@@ -306,10 +306,8 @@ void MovesStorage(pten::DenseTensor* src, paddle::framework::Tensor* dst) {
           "The destination Tensor is nullptr when move storage."));
   dst->Resize(src->dims());
   dst->set_type(pten::TransToProtoVarType(src->dtype()));
-  auto storage = src->release();
-  std::shared_ptr<pten::Allocation> holder(
-      new TensorStorage(std::move(storage)));
-  dst->ResetHolderWithType(holder, pten::TransToProtoVarType(src->dtype()));
+  auto storage = src->MoveMemoryHolder();
+  dst->ResetHolderWithType(storage, pten::TransToProtoVarType(src->dtype()));
   dst->set_offset(src->meta().offset);
 }
 
diff --git a/paddle/pten/core/dense_tensor.h b/paddle/pten/core/dense_tensor.h
index 1802a2461158fc9952857cce7d48ba42229405ed..4f25fc296724c86cd844d21c262681c9d24037b2 100644
--- a/paddle/pten/core/dense_tensor.h
+++ b/paddle/pten/core/dense_tensor.h
@@ -172,12 +172,6 @@ class DenseTensor : public TensorBase,
   /// \return The actual storage size occupied by tensor.
   size_t capacity() const { return storage_->size(); }
 
-  /// \brief Release the storage area for other purposes. Because of the
-  /// destruction of encapsulation, we do not support two dense tensors directly
-  /// sharing the same intrusive pointer.
-  /// \return The rvalue of instrusize pointer releated to the released storage.
-  intrusive_ptr<Storage> release() { return std::move(storage_); }
-
   /// \brief Get the mutable data pointer value of type T.
   /// Memory allocation may occur when calling this interface:
   /// 1. When the storage size is not enough to meet the current shape of the
diff --git a/paddle/pten/tests/core/test_dense_tensor.cc b/paddle/pten/tests/core/test_dense_tensor.cc
index c6db228c2b757d0c9aaf54f7d76dab97a0e0638c..8277c0d8dadb78edfe5b296d87f479fbae37dd2f 100644
--- a/paddle/pten/tests/core/test_dense_tensor.cc
+++ b/paddle/pten/tests/core/test_dense_tensor.cc
@@ -116,9 +116,6 @@ TEST(dense_tensor, resize) {
   CHECK_EQ(tensor_0.capacity(), 6u);
   tensor_0.mutable_data<int8_t>();
   CHECK_EQ(tensor_0.capacity(), 6u);
-
-  auto storage = tensor_0.release();
-  CHECK_EQ(storage->size(), 6u);
 }
 
 TEST(dense_tensor, shallow_copy) {
@@ -133,10 +130,6 @@ TEST(dense_tensor, shallow_copy) {
 
   DenseTensor tensor_1(tensor_0);
   CHECK(tensor_0.meta() == tensor_1.meta());
-
-  // Copy constructor: Now shares the underlying shared_ptr<Allocation> instead
-  // of Storage
-  CHECK(tensor_0.release() != tensor_1.release());
 }
 
 }  // namespace tests