diff --git a/paddle/fluid/framework/details/multi_devices_graph_builder.cc b/paddle/fluid/framework/details/multi_devices_graph_builder.cc
index f1f8674caf663ce38df5a2eecbcf690b5ca87dc4..e0fb92cc3c73ecc094d53f5878a224cd3136907a 100644
--- a/paddle/fluid/framework/details/multi_devices_graph_builder.cc
+++ b/paddle/fluid/framework/details/multi_devices_graph_builder.cc
@@ -207,7 +207,7 @@ std::unique_ptr<Graph> MultiDevSSAGraphBuilder::Apply(
   result.Set("ops", new GraphOps);
 
   // find send/recv vars so that we can place the distributed training
-  // realted op in the place 0
+  // related op in the place 0
   auto send_vars = FindDistTrainSendVars(nodes);
   auto recv_vars = FindDistTrainRecvVars(nodes);
 
diff --git a/python/paddle/fluid/transpiler/distribute_transpiler.py b/python/paddle/fluid/transpiler/distribute_transpiler.py
index e7698d8c52b411fd0afe919625034107081726b5..4a9ea6af747c36e5817ede5fafbadeea79fb07ac 100644
--- a/python/paddle/fluid/transpiler/distribute_transpiler.py
+++ b/python/paddle/fluid/transpiler/distribute_transpiler.py
@@ -779,7 +779,9 @@ class DistributeTranspiler(object):
                         outputs={"Out": prefetch_output_vars},
                         attrs={
                             "epmap": pserver_endpoints,
-                            RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE
+                            # FIXME(qiao) temporarily disable this config because prefetch
+                            # is not act as other rpc op, it's more like a forward op
+                            # RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE
                         })
 
                     # insert concat_op