diff --git a/paddle/fluid/pybind/eager_method.cc b/paddle/fluid/pybind/eager_method.cc
index 895d1db43cbf4f316d0b5d4db4880f01894b66f2..1f2ab94675205af1d8c8341c7c08d4a6553b7536 100644
--- a/paddle/fluid/pybind/eager_method.cc
+++ b/paddle/fluid/pybind/eager_method.cc
@@ -1308,6 +1308,27 @@ static PyObject* tensor_method_get_rows(TensorObject* self, PyObject* args,
   EAGER_CATCH_AND_THROW_RETURN_NULL
 }
 
+static PyObject* tensor__reset_grad_inplace_version(TensorObject* self,
+                                                    PyObject* args,
+                                                    PyObject* kwargs) {
+  EAGER_TRY
+  Py_ssize_t args_num = PyTuple_Size(args);
+  bool set_to_zero = true;
+  if (args_num == (Py_ssize_t)1) {
+    set_to_zero = CastPyArg2AttrBoolean(PyTuple_GET_ITEM(args, 0), 0);
+  }
+
+  paddle::experimental::Tensor* grad =
+      egr::EagerUtils::mutable_grad(self->tensor);
+  if (grad && grad->defined() && grad->is_dense_tensor() &&
+      grad->initialized()) {
+    grad->reset_inplace_version(set_to_zero);
+  }
+  Py_INCREF(Py_None);
+  return Py_None;
+  EAGER_CATCH_AND_THROW_RETURN_NULL
+}
+
 PyMethodDef variable_methods[] = {
     {"numpy", (PyCFunction)(void (*)(void))tensor_method_numpy,
      METH_VARARGS | METH_KEYWORDS, NULL},
@@ -1407,6 +1428,9 @@ PyMethodDef variable_methods[] = {
      METH_VARARGS | METH_KEYWORDS, NULL},
     {"rows", (PyCFunction)(void (*)(void))tensor_method_get_rows,
      METH_VARARGS | METH_KEYWORDS, NULL},
+    {"_reset_grad_inplace_version",
+     (PyCFunction)(void (*)(void))tensor__reset_grad_inplace_version,
+     METH_VARARGS | METH_KEYWORDS, NULL},
     {NULL, NULL, 0, NULL}};
 
 }  // namespace pybind
diff --git a/paddle/phi/api/include/tensor.h b/paddle/phi/api/include/tensor.h
index 934c89fc927e8cfcec1b3af717172e7aabfca4a2..9b0371fc380b9d7a807a31e8fd6f0ba2751e38d4 100644
--- a/paddle/phi/api/include/tensor.h
+++ b/paddle/phi/api/include/tensor.h
@@ -516,6 +516,11 @@ class PADDLE_API Tensor final {
    */
   uint32_t current_inplace_version();
 
+  /**
+   * @brief Reset inplace version
+   */
+  void reset_inplace_version(bool set_to_zero = false);
+
   /* Part 10: Auto generated Tensor methods */
 
   /* Part 11: Methods of converting SparseTensor and DenseTensor to each other
diff --git a/paddle/phi/api/lib/tensor.cc b/paddle/phi/api/lib/tensor.cc
index cfa8e80e84541b044611e1f17b93d4ccb1fdc9dd..5cd1fcb919638897ad17fdec496fdcc6c678e154 100644
--- a/paddle/phi/api/lib/tensor.cc
+++ b/paddle/phi/api/lib/tensor.cc
@@ -384,5 +384,16 @@ uint32_t Tensor::current_inplace_version() {
   return 0;
 }
 
+void Tensor::reset_inplace_version(bool set_to_zero) {
+  if (set_to_zero) {
+    if (is_dense_tensor()) {
+      auto &inplace_version_counter =
+          std::dynamic_pointer_cast<phi::DenseTensor>(impl_)
+              ->InplaceVersionCounter();
+      inplace_version_counter.SetInplaceVersionToZero();
+    }
+  }
+}
+
 }  // namespace experimental
 }  // namespace paddle
diff --git a/python/paddle/fluid/tests/unittests/test_reset_grad_inplace_version.py b/python/paddle/fluid/tests/unittests/test_reset_grad_inplace_version.py
index fee5bb8f47f260e8f4675c19fd194ec80a9dd4a6..84e22024f76cf53c245c2a744277674d274d61a9 100644
--- a/python/paddle/fluid/tests/unittests/test_reset_grad_inplace_version.py
+++ b/python/paddle/fluid/tests/unittests/test_reset_grad_inplace_version.py
@@ -16,6 +16,7 @@ import paddle
 import paddle.fluid as fluid
 from paddle import _C_ops
 from paddle.fluid import framework
+from paddle.fluid.framework import _test_eager_guard
 import unittest
 paddle.set_device('cpu')
 
@@ -32,7 +33,7 @@ def clear_grad_test_0(w, a):
 
 
 class TestInplaceAndClearGradient(unittest.TestCase):
-    def test(self):
+    def func_test(self):
         input_data = np.ones([1, 1])
         w = paddle.to_tensor(input_data, 'float32', stop_gradient=False)
 
@@ -45,6 +46,11 @@ class TestInplaceAndClearGradient(unittest.TestCase):
             out.backward()
         assert w.grad[0] == 0.15
 
+    def test(self):
+        with _test_eager_guard():
+            self.func_test()
+        self.func_test()
+
 
 # Test 2
 class Counter:
@@ -67,7 +73,7 @@ def clear_grad_test_1(w, c):
 
 
 class TestInplaceClearGradAccumulation(unittest.TestCase):
-    def test(self):
+    def func_test(self):
         input_data = np.ones([1, 1])
         w = paddle.to_tensor(input_data, 'float32', stop_gradient=False)
         c = Counter()
@@ -87,9 +93,14 @@ class TestInplaceClearGradAccumulation(unittest.TestCase):
             assert c.num_calls == 1
             c.num_calls = 0
 
+    def test(self):
+        with _test_eager_guard():
+            self.func_test()
+        self.func_test()
+
 
 class TestInplaceClearGradAccumulationAlt(unittest.TestCase):
-    def test(self):
+    def func_test(self):
         input_data = np.ones([1, 1])
         w = paddle.to_tensor(input_data, 'float32', stop_gradient=False)
         out = _C_ops.scale(w, 'scale', 0.1)
@@ -100,6 +111,11 @@ class TestInplaceClearGradAccumulationAlt(unittest.TestCase):
 
         assert w.grad._inplace_version() == 1
 
+    def test(self):
+        with _test_eager_guard():
+            self.func_test()
+        self.func_test()
+
 
 if __name__ == '__main__':
     unittest.main()