diff --git a/paddle/fluid/eager/backward.cc b/paddle/fluid/eager/backward.cc
index 3f56c2d01c76eca3065dff5c23ecf16419ecb734..7a4e7f81611d12f01e01ce94072b56533bc1dfbc 100644
--- a/paddle/fluid/eager/backward.cc
+++ b/paddle/fluid/eager/backward.cc
@@ -553,7 +553,13 @@ std::vector<paddle::experimental::Tensor> RunBackward(
   for (size_t i = 0; i < tensors.size(); i++) {
     const paddle::experimental::Tensor& tensor = tensors[i];
 
-    AutogradMeta* auto_grad_meta = EagerUtils::unsafe_autograd_meta(tensor);
+    AutogradMeta* auto_grad_meta = EagerUtils::nullable_autograd_meta(tensor);
+    if (auto_grad_meta == nullptr) {
+      VLOG(3) << "Skip auto grad since there is no grad op for var or loss is "
+                 "stop_gradient=True: "
+              << tensor.name();
+      continue;
+    }
     // Get grad input info from target tensors
     auto input_info = auto_grad_meta->OutRankInfo();
 
diff --git a/python/paddle/fluid/tests/unittests/test_diff_op.py b/python/paddle/fluid/tests/unittests/test_diff_op.py
index 99a46bfd9584d9bfb5e5cb42d24482cceced4eb5..b4359754520091d80f945adc47487be5f46d93c0 100644
--- a/python/paddle/fluid/tests/unittests/test_diff_op.py
+++ b/python/paddle/fluid/tests/unittests/test_diff_op.py
@@ -19,8 +19,7 @@ import paddle
 import paddle.fluid as fluid
 import paddle.fluid.layers as layers
 import paddle.fluid.core as core
-from paddle.fluid.framework import _enable_legacy_dygraph
-_enable_legacy_dygraph()
+from paddle.fluid.framework import _test_eager_guard
 
 
 class TestDiffOp(unittest.TestCase):
@@ -55,7 +54,7 @@ class TestDiffOp(unittest.TestCase):
         if core.is_compiled_with_cuda():
             self.places.append(paddle.CUDAPlace(0))
 
-    def test_dygraph(self):
+    def func_dygraph(self):
         for place in self.places:
             paddle.disable_static()
             x = paddle.to_tensor(self.input, place=place)
@@ -71,6 +70,13 @@ class TestDiffOp(unittest.TestCase):
                 append=self.append)
             self.assertTrue((out.numpy() == self.output).all(), True)
 
+    def test_dygraph(self):
+        with _test_eager_guard():
+            self.setUp()
+            self.func_dygraph()
+        self.setUp()
+        self.func_dygraph()
+
     def test_static(self):
         paddle.enable_static()
         places = [fluid.CPUPlace()]
@@ -110,7 +116,7 @@ class TestDiffOp(unittest.TestCase):
                                   fetch_list=[out])
                 self.assertTrue((fetches[0] == self.output).all(), True)
 
-    def test_grad(self):
+    def func_grad(self):
         for place in self.places:
             x = paddle.to_tensor(self.input, place=place, stop_gradient=False)
             if self.prepend is not None:
@@ -129,6 +135,13 @@ class TestDiffOp(unittest.TestCase):
             except:
                 raise RuntimeError("Check Diff Gradient Failed")
 
+    def test_grad(self):
+        with _test_eager_guard():
+            self.setUp()
+            self.func_grad()
+        self.setUp()
+        self.func_grad()
+
 
 class TestDiffOpAxis(TestDiffOp):
     def set_args(self):
diff --git a/python/paddle/tensor/math.py b/python/paddle/tensor/math.py
index b7b08af9e60bc70164b6f6cbb3166d74cb1312c9..83501b039949288866d0873d433058b72268e367 100644
--- a/python/paddle/tensor/math.py
+++ b/python/paddle/tensor/math.py
@@ -4260,18 +4260,19 @@ def diff(x, n=1, axis=-1, prepend=None, append=None, name=None):
         ends_2 = [dim_len]
         attrs_2 += ('ends', ends_2)
         if in_dygraph_mode():
-            input_back = input_front = _C_ops.final_state_slice(new_input, axes, starts_2, ends_2, infer_flags,
+            input_back = _C_ops.final_state_slice(new_input, axes, starts_2, ends_2, infer_flags,
                                             [])
         else:
             input_back = _C_ops.slice(new_input, None, None, None, None, 'axes', axes, \
                 'infer_flags', infer_flags, *attrs_2)
 
         if x.dtype == paddle.bool:
-            op = getattr(_C_ops, "logical_xor")
-            out = op(input_back, input_front)
+            if in_dygraph_mode():
+                return _C_ops.final_state_logical_xor(input_back, input_front)
+            else:
+                return _C_ops.logical_xor(input_back, input_front)
         else:
-            out = elementwise_sub(input_back, input_front, axis=axis)
-        return out
+            return elementwise_sub(input_back, input_front, axis=axis)
 
     else:
         check_variable_and_dtype(x, 'x', ['float32', 'float64', 'bool', 'int32', 'int64'], 'diff')