diff --git a/paddle/phi/api/yaml/legacy_backward.yaml b/paddle/phi/api/yaml/legacy_backward.yaml
index 234885a9d7ee5dc7265325e5582325433fed02f1..2f4eff98356044343bee115cf169a050be117417 100755
--- a/paddle/phi/api/yaml/legacy_backward.yaml
+++ b/paddle/phi/api/yaml/legacy_backward.yaml
@@ -2096,6 +2096,12 @@
   optional : grad_grad_out_grad
   inplace : (grad_grad_x -> fwd_grad_out_grad)
 
+- backward_op : sign_grad
+  forward : sign (Tensor x) -> Tensor(out)
+  args : (Tensor out_grad)
+  output : Tensor(x_grad)
+  invoke : scale(out_grad, 0.0, 0.0, true)
+
 - backward_op : silu_grad
   forward : silu (Tensor x) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)
diff --git a/paddle/phi/api/yaml/legacy_ops.yaml b/paddle/phi/api/yaml/legacy_ops.yaml
index ee251b8d9997e2e9d4779cba8a81ae8f1b3008cf..eddc9117f05188a1a834d5a6585c27f06cbda233 100755
--- a/paddle/phi/api/yaml/legacy_ops.yaml
+++ b/paddle/phi/api/yaml/legacy_ops.yaml
@@ -2377,6 +2377,7 @@
     func : UnchangedInferMeta
   kernel :
     func : sign
+  backward : sign_grad
 
 - op : silu
   args : (Tensor x)
diff --git a/python/paddle/fluid/tests/unittests/test_sign_op.py b/python/paddle/fluid/tests/unittests/test_sign_op.py
index 444675a4bb5c22c701177f052751e5efb7753e38..3eda8b286c84021a84ffb9d6d6d434337f36ad42 100644
--- a/python/paddle/fluid/tests/unittests/test_sign_op.py
+++ b/python/paddle/fluid/tests/unittests/test_sign_op.py
@@ -19,7 +19,11 @@ import numpy as np
 from op_test import OpTest
 import paddle
 import paddle.fluid as fluid
+import paddle.fluid.core as core
 from paddle.fluid import Program, program_guard
+import gradient_checker
+from decorator_helper import prog_scope
+import paddle.fluid.layers as layers
 
 
 class TestSignOp(OpTest):
@@ -91,6 +95,80 @@ class TestSignAPI(unittest.TestCase):
             paddle.sign(input4)
 
 
+class TestSignDoubleGradCheck(unittest.TestCase):
+
+    def sign_wrapper(self, x):
+        return paddle.sign(x[0])
+
+    @prog_scope()
+    def func(self, place):
+        # the shape of input variable should be clearly specified, not inlcude -1.
+        eps = 0.005
+        dtype = np.float32
+
+        data = layers.data('data', [1, 4], False, dtype)
+        data.persistable = True
+        out = paddle.sign(data)
+        data_arr = np.random.uniform(-1, 1, data.shape).astype(dtype)
+
+        gradient_checker.double_grad_check([data],
+                                           out,
+                                           x_init=[data_arr],
+                                           place=place,
+                                           eps=eps)
+        fluid.set_flags({"FLAGS_retain_grad_for_all_tensor": True})
+        gradient_checker.double_grad_check_for_dygraph(self.sign_wrapper,
+                                                       [data],
+                                                       out,
+                                                       x_init=[data_arr],
+                                                       place=place)
+
+    def test_grad(self):
+        paddle.enable_static()
+        places = [fluid.CPUPlace()]
+        if core.is_compiled_with_cuda():
+            places.append(fluid.CUDAPlace(0))
+        for p in places:
+            self.func(p)
+
+
+class TestSignTripleGradCheck(unittest.TestCase):
+
+    def sign_wrapper(self, x):
+        return paddle.sign(x[0])
+
+    @prog_scope()
+    def func(self, place):
+        # the shape of input variable should be clearly specified, not inlcude -1.
+        eps = 0.005
+        dtype = np.float32
+
+        data = layers.data('data', [1, 4], False, dtype)
+        data.persistable = True
+        out = paddle.sign(data)
+        data_arr = np.random.uniform(-1, 1, data.shape).astype(dtype)
+
+        gradient_checker.triple_grad_check([data],
+                                           out,
+                                           x_init=[data_arr],
+                                           place=place,
+                                           eps=eps)
+        fluid.set_flags({"FLAGS_retain_grad_for_all_tensor": True})
+        gradient_checker.triple_grad_check_for_dygraph(self.sign_wrapper,
+                                                       [data],
+                                                       out,
+                                                       x_init=[data_arr],
+                                                       place=place)
+
+    def test_grad(self):
+        paddle.enable_static()
+        places = [fluid.CPUPlace()]
+        if core.is_compiled_with_cuda():
+            places.append(fluid.CUDAPlace(0))
+        for p in places:
+            self.func(p)
+
+
 if __name__ == "__main__":
     paddle.enable_static()
     unittest.main()