diff --git a/paddle/fluid/framework/details/build_strategy.cc b/paddle/fluid/framework/details/build_strategy.cc
index f8030c53f72bc8a6f007c1eb6a3072abd8037de2..010c8dee6c414027c0a5665f69692f21b298d297 100644
--- a/paddle/fluid/framework/details/build_strategy.cc
+++ b/paddle/fluid/framework/details/build_strategy.cc
@@ -133,12 +133,15 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
   void AppendMultiDevPass(const BuildStrategy &strategy) {
     ir::Pass *multi_devices_pass;
     if (strategy_.is_distribution_) {
+      VLOG(3) << "multi device dist train mode";
       multi_devices_pass = AppendPass("dist_multi_devices_pass").get();
     } else {
       if (strategy.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce) {
+        VLOG(3) << "multi device allreduce mode";
         multi_devices_pass =
             AppendPass("allreduce_mode_multi_devices_pass").get();
       } else if (strategy.reduce_ == BuildStrategy::ReduceStrategy::kReduce) {
+        VLOG(3) << "multi device reduce mode";
         multi_devices_pass = AppendPass("reduce_mode_multi_devices_pass").get();
       } else {
         PADDLE_THROW("Unknown reduce strategy.");
diff --git a/paddle/fluid/framework/details/multi_devices_graph_pass.cc b/paddle/fluid/framework/details/multi_devices_graph_pass.cc
index 75f922d2cca6855a67be7284ae407e549a1a1afb..24977aabdacb265694d0130ccc28b5f2e05ce4f5 100644
--- a/paddle/fluid/framework/details/multi_devices_graph_pass.cc
+++ b/paddle/fluid/framework/details/multi_devices_graph_pass.cc
@@ -731,7 +731,6 @@ bool DistSSAGraphBuilder::DealWithSpecialOp(ir::Graph *result,
       }
     }
     insert_op = true;
-    need_broadcast_var_ = true;
   } else if (OpHaveRole(*node, OpRole::kDist)) {
     int op_dev_id = CreateDistTrainOp(result, node);
     if (node->Op()->Type() == "concat") {
@@ -925,9 +924,8 @@ void DistSSAGraphBuilder::InsertCollectiveOp(ir::Graph *result,
 }
 
 void DistSSAGraphBuilder::InsertPostprocessOps(ir::Graph *result) const {
-  if (need_broadcast_var_ ||
-      (UseGPU() &&
-       strategy_.reduce_ == BuildStrategy::ReduceStrategy::kReduce)) {
+  // only GPU reduce mode need to broadcast parameters to each device.
+  if (UseGPU() && strategy_.reduce_ == BuildStrategy::ReduceStrategy::kReduce) {
     if (strategy_.fuse_broadcast_op_) {
       CreateFusedBroadcastOp(result, bcast_var_name_set_);
     } else {
diff --git a/paddle/fluid/framework/details/multi_devices_graph_pass.h b/paddle/fluid/framework/details/multi_devices_graph_pass.h
index 6d4386538ea7d0cc318647c92282af9d598fa699..21f85dc828687ba6224e5f353133546bedaa4a0a 100644
--- a/paddle/fluid/framework/details/multi_devices_graph_pass.h
+++ b/paddle/fluid/framework/details/multi_devices_graph_pass.h
@@ -174,7 +174,6 @@ class DistSSAGraphBuilder : public BalanceVarSSAGraphBuilder {
   int CreateDistTrainOp(ir::Graph *result, ir::Node *node) const;
 
   mutable std::vector<std::unordered_set<std::string>> bcast_var_name_set_;
-  mutable bool need_broadcast_var_{false};
 };
 
 std::unordered_set<std::string> &MultiDevSSAGraphBuilder();
diff --git a/python/paddle/fluid/compiler.py b/python/paddle/fluid/compiler.py
index ef0242942838fcca737a10fafbafa61bf520b532..2b69fd89a2c19f933a7ff7bb3bcc4bbb2c49e7e9 100644
--- a/python/paddle/fluid/compiler.py
+++ b/python/paddle/fluid/compiler.py
@@ -19,6 +19,7 @@ import sys
 from .. import compat as cpt
 
 from . import core
+from . import framework
 
 __all__ = ['CompiledProgram', 'ExecutionStrategy', 'BuildStrategy']
 
@@ -34,6 +35,15 @@ def _place_obj(place):
     return p
 
 
+def _is_pserver_mode(main_program):
+    main = main_program if main_program \
+        else framework.default_main_program()
+    for op in main.global_block().ops:
+        if op.type in ["send", "recv"]:
+            return True
+    return False
+
+
 class CompiledProgram(object):
     """
     Compiles a Program for execution.
@@ -110,6 +120,7 @@ class CompiledProgram(object):
             self._exec_strategy = ExecutionStrategy()
         if self._build_strategy is None:
             self._build_strategy = BuildStrategy()
+        self._build_strategy.is_distribution = _is_pserver_mode(self._program)
         return self
 
     def with_inference_optimize(self, config):