diff --git a/paddle/fluid/operators/activation_op_npu.cc b/paddle/fluid/operators/activation_op_npu.cc
old mode 100755
new mode 100644
index d815a3eeb4d81c70f7eb6ab729afee6b04ffe12f..5f2925784e42442a1d42dbf1fcb34ee2c39a41c4
--- a/paddle/fluid/operators/activation_op_npu.cc
+++ b/paddle/fluid/operators/activation_op_npu.cc
@@ -386,6 +386,35 @@ class SquareNPUKernel : public framework::OpKernel<T> {
   }
 };
 
+template <typename DeviceContext, typename T>
+class SquareGradNPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* x = ctx.Input<Tensor>("X");
+    auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
+    auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
+
+    auto factor = static_cast<float>(2.0);
+
+    auto place = ctx.GetPlace();
+    auto stream =
+        ctx.template device_context<paddle::platform::NPUDeviceContext>()
+            .stream();
+    // Step 1: Compute x_muls_factor = factor * x
+    Tensor x_muls_factor(x->type());
+    x_muls_factor.mutable_data<T>(x->dims(), place);
+    const auto& runner_muls_1 =
+        NpuOpRunner("Muls", {*x}, {x_muls_factor}, {{"value", factor}});
+    runner_muls_1.Run(stream);
+
+    // Step 2: Compute dx = dout * factor * x
+    dx->mutable_data<T>(place);
+    const auto& runner_mul_2 =
+        NpuOpRunner("Mul", {*dout, x_muls_factor}, {*dx}, {});
+    runner_mul_2.Run(stream);
+  }
+};
+
 template <typename DeviceContext, typename T>
 class SigmoidNPUKernel : public framework::OpKernel<T> {
  public:
@@ -869,6 +898,12 @@ REGISTER_OP_NPU_KERNEL(
                          paddle::platform::float16>,
     ops::SquareNPUKernel<paddle::platform::NPUDeviceContext, int>);
 
+REGISTER_OP_NPU_KERNEL(
+    square_grad,
+    ops::SquareGradNPUKernel<paddle::platform::NPUDeviceContext, float>,
+    ops::SquareNPUKernel<paddle::platform::NPUDeviceContext,
+                         paddle::platform::float16>);
+
 REGISTER_OP_NPU_KERNEL(
     sigmoid, ops::SigmoidNPUKernel<paddle::platform::NPUDeviceContext, float>,
     ops::SigmoidNPUKernel<paddle::platform::NPUDeviceContext,
diff --git a/python/paddle/fluid/tests/unittests/npu/test_square_op_npu.py b/python/paddle/fluid/tests/unittests/npu/test_square_op_npu.py
index 8262c3b94d6a8531fe104eed6c4c447d0e9edef1..caf55b4850f0b18f0fb20ed5692119c2b4ceccc2 100644
--- a/python/paddle/fluid/tests/unittests/npu/test_square_op_npu.py
+++ b/python/paddle/fluid/tests/unittests/npu/test_square_op_npu.py
@@ -50,12 +50,10 @@ class TestSquare(OpTest):
     def test_check_output(self):
         self.check_output_with_place(self.place)
 
-    # TODO(ascendrc): Add grad test
-    # def test_check_grad(self):
-    #     if self.dtype == np.float16:
-    #         return
-    #     self.check_grad(['X'], 'Out')
-    #
+    def test_check_grad(self):
+        if self.dtype == np.float16:
+            return
+        self.check_grad_with_place(self.place, ['X'], 'Out')
 
 
 class TestSquareFp16(OpTest):