diff --git a/python/paddle/fluid/framework.py b/python/paddle/fluid/framework.py
index c207c5282f3b524e8deab9857b9fe0eddc9b4f55..1409acd5ca0aa3bef7e4c916296b99484929449a 100644
--- a/python/paddle/fluid/framework.py
+++ b/python/paddle/fluid/framework.py
@@ -2773,6 +2773,9 @@ class OpProtoHolder:
 
         return custom_op_names
 
+    def has_op_proto(self, type):
+        return type in self.op_proto_map
+
     @staticmethod
     def generated_op_attr_names():
         return {
diff --git a/python/paddle/jit/translated_layer.py b/python/paddle/jit/translated_layer.py
index 45563584f166d00128858c10c40b56068f2ada3a..be9e1bee3255894d4c934e1ce6269b45dd5d44cd 100644
--- a/python/paddle/jit/translated_layer.py
+++ b/python/paddle/jit/translated_layer.py
@@ -563,6 +563,11 @@ class _ProgramHolder:
                         op.desc.set_output("ReserveSpace", [reserve_space.name])
                     continue
 
+                # There are some situations that users will add backward op in Forward
+                # function of Layer. And because backward op doesn't have proto. So, we
+                # should skip it when we meet it.
+                if not OpProtoHolder.instance().has_op_proto(op.type):
+                    continue
                 proto = OpProtoHolder.instance().get_op_proto(op.type)
                 has_create_intermediate_out = False
                 for output_proto in proto.outputs: