diff --git a/paddle/phi/kernels/gpu/argsort_grad_kernel.cu b/paddle/phi/kernels/gpu/argsort_grad_kernel.cu
index f28da8704cbffb06a3fb76f11f36f3abcfd6a2cc..b8d9df64c23efb54fcc81a3ff43fb26051edd850 100644
--- a/paddle/phi/kernels/gpu/argsort_grad_kernel.cu
+++ b/paddle/phi/kernels/gpu/argsort_grad_kernel.cu
@@ -150,7 +150,7 @@ void ArgsortGradKernel(const Context& dev_ctx,
   int64_t size = in_grad->numel();
 
   if (rank == 0) {
-    phi::funcs::set_constant(dev_ctx, in_grad, 1.0);
+    phi::Copy<Context>(dev_ctx, out_grad, dev_ctx.GetPlace(), false, in_grad);
     return;
   }
 
diff --git a/paddle/phi/kernels/xpu/argsort_grad_kernel.cc b/paddle/phi/kernels/xpu/argsort_grad_kernel.cc
index 00c679f0ab999e4539d4fc8886ddec1cb475c301..96cce0461783444e71459e50a1aa775d4b9efb90 100644
--- a/paddle/phi/kernels/xpu/argsort_grad_kernel.cc
+++ b/paddle/phi/kernels/xpu/argsort_grad_kernel.cc
@@ -43,7 +43,7 @@ void ArgsortGradKernel(const Context& dev_ctx,
   if (out_grad.numel() == 0) return;
 
   if (rank == 0) {
-    phi::funcs::set_constant(dev_ctx, in_grad, 1.0);
+    phi::Copy<Context>(dev_ctx, out_grad, dev_ctx.GetPlace(), false, in_grad);
     return;
   }
 
diff --git a/python/paddle/fluid/tests/unittests/test_zero_dim_tensor.py b/python/paddle/fluid/tests/unittests/test_zero_dim_tensor.py
index 5aec0c8010f3f350fd648a1be7546a1902f31829..546c0c48f9b585c6b3a1549a9288fe386a1392d2 100644
--- a/python/paddle/fluid/tests/unittests/test_zero_dim_tensor.py
+++ b/python/paddle/fluid/tests/unittests/test_zero_dim_tensor.py
@@ -868,10 +868,14 @@ class TestSundryAPI(unittest.TestCase):
 
         self.assertEqual(out1.shape, [])
         self.assertEqual(out2.shape, [])
+        self.assertEqual(out1.numpy(), x1.numpy())
+        self.assertEqual(out2.numpy(), x2.numpy())
         self.assertEqual(out1.grad.shape, [])
         self.assertEqual(out2.grad.shape, [])
         self.assertEqual(x1.grad.shape, [])
         self.assertEqual(x2.grad.shape, [])
+        self.assertEqual(x1.grad.numpy(), 1)
+        self.assertEqual(x2.grad.numpy(), 1)
 
     def test_argsort(self):
         x1 = paddle.rand([])
@@ -886,10 +890,14 @@ class TestSundryAPI(unittest.TestCase):
 
         self.assertEqual(out1.shape, [])
         self.assertEqual(out2.shape, [])
+        self.assertEqual(out1.numpy(), 0)
+        self.assertEqual(out2.numpy(), 0)
         self.assertEqual(out1.grad.shape, [])
         self.assertEqual(out2.grad.shape, [])
         self.assertEqual(x1.grad.shape, [])
         self.assertEqual(x2.grad.shape, [])
+        self.assertEqual(x1.grad.numpy(), 0)
+        self.assertEqual(x2.grad.numpy(), 0)
 
 
 class TestSundryAPIStatic(unittest.TestCase):
diff --git a/python/paddle/fluid/tests/unittests/xpu/test_zero_dim_tensor_xpu.py b/python/paddle/fluid/tests/unittests/xpu/test_zero_dim_tensor_xpu.py
index 8cb27ecf0992ad534b590946a1734d4947340e5e..221c46228ea9a3f4625a608b58b2f9b80684f26a 100644
--- a/python/paddle/fluid/tests/unittests/xpu/test_zero_dim_tensor_xpu.py
+++ b/python/paddle/fluid/tests/unittests/xpu/test_zero_dim_tensor_xpu.py
@@ -659,10 +659,14 @@ class TestSundryAPI(unittest.TestCase):
 
         self.assertEqual(out1.shape, [])
         self.assertEqual(out2.shape, [])
+        self.assertEqual(out1.numpy(), x1.numpy())
+        self.assertEqual(out2.numpy(), x2.numpy())
         self.assertEqual(out1.grad.shape, [])
         self.assertEqual(out2.grad.shape, [])
         self.assertEqual(x1.grad.shape, [])
         self.assertEqual(x2.grad.shape, [])
+        self.assertEqual(x1.grad.numpy(), 1)
+        self.assertEqual(x2.grad.numpy(), 1)
 
     def test_argsort(self):
         x1 = paddle.rand([])