diff --git a/paddle/fluid/pybind/eager_method.cc b/paddle/fluid/pybind/eager_method.cc
index 5b9b91ef89b25a195cd877c137baddc12146b8d1..bb638ffd3a1e4177934d225e4025484c7a3efd67 100644
--- a/paddle/fluid/pybind/eager_method.cc
+++ b/paddle/fluid/pybind/eager_method.cc
@@ -868,16 +868,22 @@ static PyObject* tensor_register_grad_hook(TensorObject* self, PyObject* args,
   int64_t hook_id;
   if (egr::egr_utils_api::IsLeafTensor(self->tensor)) {
     VLOG(6) << "Register hook for leaf tensor: " << self->tensor.name();
+
+    auto autograd_meta = egr::EagerUtils::unsafe_autograd_meta(self->tensor);
+
+    if (autograd_meta && !autograd_meta->StopGradient()) {
+      if (!autograd_meta->GetMutableGradNode()) {
+        VLOG(6) << "Detected NULL grad_node, Leaf tensor should have had "
+                   "grad_node with type: GradNodeAccumulation.";
+        autograd_meta->SetGradNode(
+            std::make_shared<egr::GradNodeAccumulation>(autograd_meta));
+      }
+    }
+
     std::shared_ptr<egr::GradNodeBase> grad_node =
         egr::EagerUtils::grad_node(self->tensor);
-    PADDLE_ENFORCE(
-        grad_node.get() != nullptr,
-        paddle::platform::errors::Fatal("Detected NULL grad_node,"
-                                        "Leaf tensor should have had grad_node "
-                                        "with type: GradNodeAccumulation."));
     auto rank_info =
         egr::EagerUtils::unsafe_autograd_meta(self->tensor)->OutRankInfo();
-
     PyObject* hook_func = PyTuple_GET_ITEM(args, 0);
 
     auto accumulation_grad_node =
diff --git a/python/paddle/fluid/tests/unittests/test_inplace_eager_fluid.py b/python/paddle/fluid/tests/unittests/test_inplace_eager_fluid.py
index 33f55e0d518815b92b37377475d6bbc0514f0765..45232ae4e4600efec192f35191e5ccd25cfe47f0 100644
--- a/python/paddle/fluid/tests/unittests/test_inplace_eager_fluid.py
+++ b/python/paddle/fluid/tests/unittests/test_inplace_eager_fluid.py
@@ -171,6 +171,180 @@ class TestDygraphInplace(unittest.TestCase):
                 grad_var_a = var_a.grad.numpy()
         self.assertTrue(np.array_equal(grad_var_a_inplace, grad_var_a))
 
+    # inplace + hook
+    def test_backward_success_3(self):
+        # var_b is modified inplace before using it, the inplace operator doesn't result
+        # in incorrect gradient computation.
+        def double_hook(grad):
+            grad = grad * 2
+            return grad
+
+        grad_var_a, grad_var_a_inplace = 0, 1
+        with paddle.fluid.dygraph.guard():
+            with _test_eager_guard():
+                var_a = paddle.to_tensor(self.input_var_numpy).astype(
+                    self.dtype)
+                var_a.stop_gradient = False
+                helper = var_a.register_hook(double_hook)
+
+                var_b = var_a**2
+                var_c = self.inplace_api_processing(
+                    var_b)  # var_b is modified inplace before using it
+
+                # Here, the gradient computation will use the value of var_b
+                var_d = var_c**2
+                loss = var_d.sum()
+                loss.backward()
+                grad_var_a_inplace = var_a.grad.numpy()
+
+        with paddle.fluid.dygraph.guard():
+            with _test_eager_guard():
+                var_a = paddle.to_tensor(self.input_var_numpy).astype(
+                    self.dtype)
+                var_a.stop_gradient = False
+                helper = var_a.register_hook(double_hook)
+
+                var_b = var_a**2
+                var_c = self.non_inplace_api_processing(var_b)
+                var_d = var_c**2
+                loss = var_d.sum()
+                loss.backward()
+                grad_var_a = var_a.grad.numpy()
+
+        self.assertTrue(self.np_compare(grad_var_a_inplace, grad_var_a))
+
+    # inplace + hook
+    def test_backward_success_4(self):
+        # Although var_b is modified inplace after using it, it does not used in gradient computation.
+        # The inplace operator doesn't result in incorrect gradient computation.
+        def double_hook(grad):
+            grad = grad * 2
+            return grad
+
+        grad_var_a, grad_var_a_inplace = 0, 1
+        with paddle.fluid.dygraph.guard():
+            with _test_eager_guard():
+                var_a = paddle.to_tensor(self.input_var_numpy).astype(
+                    self.dtype)
+                var_a.stop_gradient = False
+                var_a.register_hook(double_hook)
+
+                var_b = var_a**2
+
+                var_c = self.inplace_api_processing(
+                    var_b)  # var_b is modified inplace before using it
+
+                var_d = var_c + var_c  # Here, the grad op of sum doesn't use the value of var_b
+                loss = var_d.sum()
+
+                loss.backward()
+                grad_var_a_inplace = var_a.grad.numpy()
+
+        with paddle.fluid.dygraph.guard():
+            with _test_eager_guard():
+                var_a = paddle.to_tensor(self.input_var_numpy).astype(
+                    self.dtype)
+                var_a.stop_gradient = False
+                var_a.register_hook(double_hook)
+
+                var_b = var_a**2
+
+                var_c = self.non_inplace_api_processing(
+                    var_b)  # var_b is modified inplace before using it
+
+                var_d = var_c + var_c  # Here, the grad op of sum doesn't use the value of var_b
+                loss = var_d.sum()
+
+                loss.backward()
+                grad_var_a = var_a.grad.numpy()
+        self.assertTrue(np.array_equal(grad_var_a_inplace, grad_var_a))
+
+    # inplace + hook
+    def test_backward_success_5(self):
+        # var_b is modified inplace before using it, the inplace operator doesn't result
+        # in incorrect gradient computation.
+        def double_hook(grad):
+            grad = grad * 2
+            return grad
+
+        grad_var_a, grad_var_a_inplace = 0, 1
+        with paddle.fluid.dygraph.guard():
+            with _test_eager_guard():
+                var_a = paddle.to_tensor(self.input_var_numpy).astype(
+                    self.dtype)
+                var_a.stop_gradient = False
+
+                var_b = var_a**2
+                var_b.register_hook(double_hook)
+                var_c = self.inplace_api_processing(
+                    var_b)  # var_b is modified inplace before using it
+
+                # Here, the gradient computation will use the value of var_b
+                var_d = var_c**2
+                loss = var_d.sum()
+                loss.backward()
+                grad_var_a_inplace = var_a.grad.numpy()
+
+        with paddle.fluid.dygraph.guard():
+            with _test_eager_guard():
+                var_a = paddle.to_tensor(self.input_var_numpy).astype(
+                    self.dtype)
+                var_a.stop_gradient = False
+
+                var_b = var_a**2
+                var_b.register_hook(double_hook)
+                var_c = self.non_inplace_api_processing(var_b)
+                var_d = var_c**2
+                loss = var_d.sum()
+                loss.backward()
+                grad_var_a = var_a.grad.numpy()
+
+        self.assertTrue(self.np_compare(grad_var_a_inplace, grad_var_a))
+
+    # inplace + hook
+    def test_backward_success_6(self):
+        # Although var_b is modified inplace before using it, it does not used in gradient computation.
+        # The inplace operator doesn't result in incorrect gradient computation.
+        def double_hook(grad):
+            grad = grad * 2
+            return grad
+
+        grad_var_a, grad_var_a_inplace = 0, 1
+        with paddle.fluid.dygraph.guard():
+            with _test_eager_guard():
+                var_a = paddle.to_tensor(self.input_var_numpy).astype(
+                    self.dtype)
+                var_a.stop_gradient = False
+
+                var_b = var_a**2
+                var_b.register_hook(double_hook)
+                var_c = self.inplace_api_processing(
+                    var_b)  # var_b is modified inplace before using it
+
+                var_d = var_c + var_c  # Here, the grad op of sum doesn't use the value of var_b
+                loss = var_d.sum()
+
+                loss.backward()
+                grad_var_a_inplace = var_a.grad.numpy()
+
+        with paddle.fluid.dygraph.guard():
+            with _test_eager_guard():
+                var_a = paddle.to_tensor(self.input_var_numpy).astype(
+                    self.dtype)
+                var_a.stop_gradient = False
+
+                var_b = var_a**2
+                var_b.register_hook(double_hook)
+                var_c = self.non_inplace_api_processing(
+                    var_b)  # var_b is modified inplace before using it
+
+                var_d = var_c + var_c  # Here, the grad op of sum doesn't use the value of var_b
+                loss = var_d.sum()
+
+                loss.backward()
+                grad_var_a = var_a.grad.numpy()
+        self.assertTrue(np.array_equal(grad_var_a_inplace, grad_var_a))
+
 
 class TestDygraphInplaceUnsqueeze(TestDygraphInplace):
     def non_inplace_api_processing(self, var):