optimize broadcast logic test=develop

paddle/fluid/framework/details/multi_devices_graph_pass.cc

@@ -925,10 +925,13 @@ void DistSSAGraphBuilder::InsertCollectiveOp(ir::Graph *result,
 }
 
 void DistSSAGraphBuilder::InsertPostprocessOps(ir::Graph *result) const {
-  // only GPU reduce mode need to broadcast parameters to each device.
-  if (UseGPU()) {
-    if (need_broadcast_var_ ||
+  // broad cast received parameters when training in parameter server mode.
+  if (need_broadcast_var_) {
+    // cpu reduce mode did not need to broadcast received parameters.
+    if (!UseGPU() &&
         strategy_.reduce_ == BuildStrategy::ReduceStrategy::kReduce) {
+      return;
+    }
     if (strategy_.fuse_broadcast_op_) {
       CreateFusedBroadcastOp(result, bcast_var_name_set_);
     } else {
@@ -940,7 +943,6 @@ void DistSSAGraphBuilder::InsertPostprocessOps(ir::Graph *result) const {
       }
     }
   }
-  }
 }
 
 std::unordered_set<std::string> &MultiDevSSAGraphBuilder() {