diff --git a/paddle/fluid/platform/profiler/CMakeLists.txt b/paddle/fluid/platform/profiler/CMakeLists.txt
old mode 100755
new mode 100644
diff --git a/paddle/fluid/platform/profiler/dump/test_serialization_logger.cc b/paddle/fluid/platform/profiler/dump/test_serialization_logger.cc
index dee1019da2b52f93a2c2383f8f38c2c1662e63cb..d294bfee58c2b5327ed709a795e756144ae1f3c1 100644
--- a/paddle/fluid/platform/profiler/dump/test_serialization_logger.cc
+++ b/paddle/fluid/platform/profiler/dump/test_serialization_logger.cc
@@ -152,7 +152,7 @@ TEST(SerializationLoggerTest, dump_case1) {
 TEST(DeserializationReaderTest, restore_case0) {
   DeserializationReader reader("test_serialization_logger_case0.pb");
   auto profiler_result = reader.Parse();
-  auto& tree = profiler_result->GetNodeTrees();
+  auto tree = profiler_result->GetNodeTrees();
   std::map<uint64_t, std::vector<HostTraceEventNode*>> nodes =
       tree->Traverse(true);
   EXPECT_EQ(nodes[10].size(), 4u);
@@ -179,7 +179,7 @@ TEST(DeserializationReaderTest, restore_case0) {
 TEST(DeserializationReaderTest, restore_case1) {
   DeserializationReader reader("test_serialization_logger_case1.pb");
   auto profiler_result = reader.Parse();
-  auto& tree = profiler_result->GetNodeTrees();
+  auto tree = profiler_result->GetNodeTrees();
   std::map<uint64_t, std::vector<HostTraceEventNode*>> nodes =
       tree->Traverse(true);
   EXPECT_EQ(nodes[10].size(), 1u);
diff --git a/paddle/fluid/platform/profiler/event_node.h b/paddle/fluid/platform/profiler/event_node.h
index dd8dfd32df4f7c075106e58019615eaeac567967..3e589b0be2e047e6d1005aebeabedd188e9c7973 100644
--- a/paddle/fluid/platform/profiler/event_node.h
+++ b/paddle/fluid/platform/profiler/event_node.h
@@ -103,7 +103,7 @@ class CudaRuntimeTraceEventNode {
     device_node_ptrs_.push_back(node);
   }
   void LogMe(BaseLogger* logger) { logger->LogRuntimeTraceEventNode(*this); }
-  std::vector<DeviceTraceEventNode*>& GetDeviceTraceEventNodes() {
+  const std::vector<DeviceTraceEventNode*>& GetDeviceTraceEventNodes() const {
     return device_node_ptrs_;
   }
 
@@ -139,8 +139,11 @@ class HostTraceEventNode {
   void AddCudaRuntimeNode(CudaRuntimeTraceEventNode* node) {
     runtime_node_ptrs_.push_back(node);
   }
-  std::vector<HostTraceEventNode*>& GetChildren() { return children_; }
-  std::vector<CudaRuntimeTraceEventNode*>& GetRuntimeTraceEventNodes() {
+  const std::vector<HostTraceEventNode*>& GetChildren() const {
+    return children_;
+  }
+  const std::vector<CudaRuntimeTraceEventNode*>& GetRuntimeTraceEventNodes()
+      const {
     return runtime_node_ptrs_;
   }
   void LogMe(BaseLogger* logger) { logger->LogHostTraceEventNode(*this); }
@@ -188,7 +191,7 @@ class NodeTrees {
   void HandleTrees(std::function<void(HostTraceEventNode*)>,
                    std::function<void(CudaRuntimeTraceEventNode*)>,
                    std::function<void(DeviceTraceEventNode*)>);
-  std::map<uint64_t, HostTraceEventNode*> GetNodeTrees() {
+  const std::map<uint64_t, HostTraceEventNode*>& GetNodeTrees() const {
     return thread_event_trees_map_;
   }
   std::map<uint64_t, std::vector<HostTraceEventNode*>> Traverse(bool bfs) const;
diff --git a/paddle/fluid/platform/profiler/event_python.cc b/paddle/fluid/platform/profiler/event_python.cc
index 1a6f19d2f93aff104e2100b7591982a7a1d6400f..5c42c8e8bf61ebcd798ed4203fbeb7eb0c47f621 100644
--- a/paddle/fluid/platform/profiler/event_python.cc
+++ b/paddle/fluid/platform/profiler/event_python.cc
@@ -81,7 +81,7 @@ HostPythonNode* ProfilerResult::CopyTree(HostTraceEventNode* root) {
 
 ProfilerResult::ProfilerResult(std::unique_ptr<NodeTrees> tree,
                                const ExtraInfo& extra_info)
-    : tree_(std::move(tree)), extra_info_(extra_info) {
+    : tree_(tree.release()), extra_info_(extra_info) {
   if (tree_ != nullptr) {
     std::map<uint64_t, HostTraceEventNode*> nodetrees = tree_->GetNodeTrees();
     for (auto it = nodetrees.begin(); it != nodetrees.end(); ++it) {
diff --git a/paddle/fluid/platform/profiler/event_python.h b/paddle/fluid/platform/profiler/event_python.h
index 12ecb9fde32aa0aedf9bf70e8e5bdf3d6c34b69f..172116dbb0edd745530cac6a5ef210a7c7d90475 100644
--- a/paddle/fluid/platform/profiler/event_python.h
+++ b/paddle/fluid/platform/profiler/event_python.h
@@ -82,11 +82,11 @@ class ProfilerResult {
   void Save(const std::string& file_name,
             const std::string format = std::string("json"));
 
-  std::unique_ptr<NodeTrees>& GetNodeTrees() { return tree_; }
+  std::shared_ptr<NodeTrees> GetNodeTrees() { return tree_; }
 
  private:
   std::map<uint64_t, HostPythonNode*> thread_event_trees_map_;
-  std::unique_ptr<NodeTrees> tree_;
+  std::shared_ptr<NodeTrees> tree_;
   ExtraInfo extra_info_;
   HostPythonNode* CopyTree(HostTraceEventNode* root);
 };
diff --git a/paddle/fluid/platform/profiler/profiler_test.cc b/paddle/fluid/platform/profiler/profiler_test.cc
index 32310b9e86228468a0b7e1e94f302a35bb834f8a..f2c867ffff21790719e3e70f9a9e9c86168f2a37 100644
--- a/paddle/fluid/platform/profiler/profiler_test.cc
+++ b/paddle/fluid/platform/profiler/profiler_test.cc
@@ -46,7 +46,7 @@ TEST(ProfilerTest, TestHostTracer) {
                        3);
   }
   auto profiler_result = profiler->Stop();
-  auto& nodetree = profiler_result->GetNodeTrees();
+  auto nodetree = profiler_result->GetNodeTrees();
   std::set<std::string> host_events;
   for (const auto pair : nodetree->Traverse(true)) {
     for (const auto evt : pair.second) {
@@ -79,7 +79,7 @@ TEST(ProfilerTest, TestCudaTracer) {
   hipStreamSynchronize(stream);
 #endif
   auto profiler_result = profiler->Stop();
-  auto& nodetree = profiler_result->GetNodeTrees();
+  auto nodetree = profiler_result->GetNodeTrees();
   std::vector<std::string> runtime_events;
   for (const auto pair : nodetree->Traverse(true)) {
     for (const auto host_node : pair.second) {