diff --git a/python/paddle/autograd/backward_utils.py b/python/paddle/autograd/backward_utils.py
index ad530c76d1d2f..51b5781d5134f 100644
--- a/python/paddle/autograd/backward_utils.py
+++ b/python/paddle/autograd/backward_utils.py
@@ -63,6 +63,7 @@
     "pd_op.reduce_as",
     "pd_op.relu",
     "pd_op.reshape",
+    "pd_op.roll",
     "pd_op.rsqrt",
     "pd_op.scale",
     "pd_op.scatter",
diff --git a/test/prim/pir_prim/test_prim_sub_graph_pqrst_backward_dynamic_shape.py b/test/prim/pir_prim/test_prim_sub_graph_pqrst_backward_dynamic_shape.py
index 018e6b019d77a..d3b09a6e21020 100644
--- a/test/prim/pir_prim/test_prim_sub_graph_pqrst_backward_dynamic_shape.py
+++ b/test/prim/pir_prim/test_prim_sub_graph_pqrst_backward_dynamic_shape.py
@@ -64,6 +64,10 @@ def reshape_net(x):
     return paddle.reshape(x, [30, 200 * 40])
 
 
+def roll_net(x):
+    return paddle.roll(x, shifts=[101, -1], axis=[0, -2])
+
+
 def scale_net(x):
     return paddle.scale(x, scale=-2.3)
 
@@ -334,6 +338,32 @@ def setUp(self):
         self.tol = 1e-6
 
 
+class TestPrimRollWithGrad1(TestPrimBaseWithGrad):
+    def setUp(self):
+        np.random.seed(2024)
+        self.op_name = "pd_op.roll_grad"
+        self.dtype = "float32"
+        self.x_shape = [100, 4, 5]
+        self.init_x_shape = [None, None, 5]
+        self.x = np.random.random(self.x_shape).astype(self.dtype)
+        self.net = roll_net
+        self.enable_cinn = False
+        self.tol = 1e-6
+
+
+class TestPrimRollWithGrad2(TestPrimBaseWithGrad):
+    def setUp(self):
+        np.random.seed(2024)
+        self.op_name = "pd_op.roll_grad"
+        self.dtype = "float32"
+        self.x_shape = [100, 4, 5]
+        self.init_x_shape = [100, None, None]
+        self.x = np.random.random(self.x_shape).astype(self.dtype)
+        self.net = roll_net
+        self.enable_cinn = False
+        self.tol = 1e-6
+
+
 class TestPrimScaleWithGrad(TestPrimBaseWithGrad):
     def setUp(self):
         np.random.seed(2023)