diff --git a/paddle/fluid/operators/tril_triu_op_xpu.cc b/paddle/fluid/operators/tril_triu_op_xpu.cc
index a44ea8ff689b85d9f718572c45b4f8fafaf1565d..70200fe733a5ada57041b2369015502619423a85 100644
--- a/paddle/fluid/operators/tril_triu_op_xpu.cc
+++ b/paddle/fluid/operators/tril_triu_op_xpu.cc
@@ -43,6 +43,34 @@ class TrilTriuXPUKernel : public framework::OpKernel<T> {
   }
 };
 
+template <typename DeviceContext, typename T>
+class TrilTriuGradXPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    const auto* d_out =
+        context.Input<framework::Tensor>(framework::GradVarName("Out"));
+    const auto* dout_data = d_out->data<T>();
+    auto* d_x = context.Output<framework::Tensor>(framework::GradVarName("X"));
+    auto* dx_data = d_x->mutable_data<T>(context.GetPlace());
+
+    const int diagonal = context.Attr<int>("diagonal");
+    const bool lower = context.Attr<bool>("lower");
+
+    auto dy_shape = phi::vectorize<int>(d_out->dims());
+    auto& dev_ctx = context.template device_context<DeviceContext>();
+    int r = 0;
+    if (lower) {
+      r = xpu::tril(dev_ctx.x_context(), dout_data, dx_data, dy_shape,
+                    diagonal);
+      PADDLE_ENFORCE_XDNN_SUCCESS(r, "tril_op");
+    } else {
+      r = xpu::triu(dev_ctx.x_context(), dout_data, dx_data, dy_shape,
+                    diagonal);
+      PADDLE_ENFORCE_XDNN_SUCCESS(r, "triu_op");
+    }
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle
 
@@ -50,4 +78,8 @@ namespace ops = paddle::operators;
 REGISTER_OP_XPU_KERNEL(
     tril_triu, ops::TrilTriuXPUKernel<paddle::platform::XPUDeviceContext, int>,
     ops::TrilTriuXPUKernel<paddle::platform::XPUDeviceContext, float>);
+REGISTER_OP_XPU_KERNEL(
+    tril_triu_grad,
+    ops::TrilTriuGradXPUKernel<paddle::platform::XPUDeviceContext, int>,
+    ops::TrilTriuGradXPUKernel<paddle::platform::XPUDeviceContext, float>);
 #endif
diff --git a/paddle/fluid/platform/device/xpu/xpu2_op_list.h b/paddle/fluid/platform/device/xpu/xpu2_op_list.h
index 6f4826bd8c39a4b4669a595352138df67c6f5149..7b88f261d5a4fc3fd9a0ab135beb41dd31e4aa8b 100644
--- a/paddle/fluid/platform/device/xpu/xpu2_op_list.h
+++ b/paddle/fluid/platform/device/xpu/xpu2_op_list.h
@@ -380,6 +380,9 @@ XPUOpMap& get_kl2_ops() {
                              pOpKernelType(vartype::FP16, XPUPlace())})},
       {"tril_triu", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace()),
                                   pOpKernelType(vartype::INT32, XPUPlace())})},
+      {"tril_triu_grad",
+       XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace()),
+                     pOpKernelType(vartype::INT32, XPUPlace())})},
       {"tile", XPUKernelSet({pOpKernelType(vartype::INT32, XPUPlace()),
                              pOpKernelType(vartype::INT64, XPUPlace()),
                              pOpKernelType(vartype::BOOL, XPUPlace()),
diff --git a/python/paddle/fluid/tests/unittests/xpu/test_tril_triu_op_xpu.py b/python/paddle/fluid/tests/unittests/xpu/test_tril_triu_op_xpu.py
index fb6b28d9c282512caf3405d56a4321df9e2668dd..ee689efbb38a05ffe02f6d36db298346ff7ed551 100644
--- a/python/paddle/fluid/tests/unittests/xpu/test_tril_triu_op_xpu.py
+++ b/python/paddle/fluid/tests/unittests/xpu/test_tril_triu_op_xpu.py
@@ -42,6 +42,7 @@ class XPUTestTrilTriuOp(XPUOpTestWrapper):
             self.real_np_op = getattr(np, self.real_op_type)
             self.set_xpu()
             self.op_type = "tril_triu"
+            self.place = paddle.XPUPlace(0)
             if self.dtype == np.int32:
                 self.X = np.arange(
                     1, self.get_Xshape_prod() + 1,
@@ -69,13 +70,22 @@ class XPUTestTrilTriuOp(XPUOpTestWrapper):
 
         def set_xpu(self):
             self.__class__.use_xpu = True
-            self.__class__.no_need_check_grad = True
+            self.__class__.no_need_check_grad = False
             self.__class__.op_type = self.real_op_type
 
         def test_check_output(self):
-            if paddle.is_compiled_with_xpu():
-                place = paddle.XPUPlace(0)
-                self.check_output_with_place(place)
+            self.check_output_with_place(self.place)
+
+        def test_check_grad_normal(self):
+            if self.dtype == np.int32:
+                user_defined_grad_outputs = np.random.random(
+                    self.Xshape).astype('float32')
+                self.check_grad_with_place(
+                    self.place, ['X'],
+                    'Out',
+                    user_defined_grad_outputs=user_defined_grad_outputs)
+            else:
+                self.check_grad_with_place(self.place, ['X'], 'Out')
 
         def initTestCase(self):
             self.diagonal = None