distribute lookup table work with parallel executor

Author: jacquesqiao

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);
 

python/paddle/fluid/transpiler/distribute_transpiler.py

@@ -779,7 +779,9 @@ def _replace_lookup_table_op_with_prefetch(self, program,
                         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