diff --git a/paddle/fluid/framework/ir/graph_helper.cc b/paddle/fluid/framework/ir/graph_helper.cc
index 98112c1ed317c230cb5150e7cbc6d0d173256601..963179192fa6cc959db66f76e0f48393143be0da 100644
--- a/paddle/fluid/framework/ir/graph_helper.cc
+++ b/paddle/fluid/framework/ir/graph_helper.cc
@@ -15,8 +15,15 @@ limitations under the License. */
 #include "paddle/fluid/framework/ir/graph_helper.h"
 #include <algorithm>
 #include <deque>
+#include <fstream>
+#include <iosfwd>
+#include <ostream>
 #include <unordered_set>
 
+DEFINE_string(print_sub_graph_dir, "",
+              "FLAGS_print_sub_graph_dir is used "
+              "to print the nodes of sub_graphs.");
+
 namespace paddle {
 namespace framework {
 namespace ir {
@@ -164,12 +171,15 @@ size_t GraphNum(const Graph &graph) {
     graph_nodes.emplace_back(g_nodes);
   }
 
-  if (VLOG_IS_ON(100)) {
-    VLOG(100) << "graph_num: " << graph_nodes.size();
-    for (auto &g_n : graph_nodes) {
-      VLOG(100) << "graph_nodes: " << g_n.size();
-      if (g_n.size() < 10) {
-        std::stringstream out;
+  if (FLAGS_print_sub_graph_dir.size()) {
+    if (graph_nodes.size() > 1) {
+      std::stringstream out;
+      for (auto &g_n : graph_nodes) {
+        out << "graph_nodes: " << g_n.size() << "\n";
+      }
+      out << "\n\n";
+      for (auto &g_n : graph_nodes) {
+        out << "graph_nodes: " << g_n.size();
         for (auto &node : g_n) {
           out << "\nNode: " << node->Name() << " in [";
           for (auto &n : node->inputs) {
@@ -181,8 +191,12 @@ size_t GraphNum(const Graph &graph) {
           }
           out << "]";
         }
-        VLOG(100) << out.str();
+        out << "\n\n\n";
       }
+      std::unique_ptr<std::ostream> fout(
+          new std::ofstream(FLAGS_print_sub_graph_dir));
+      PADDLE_ENFORCE(fout->good());
+      *fout << out.str();
     }
   }
 
diff --git a/paddle/fluid/framework/parallel_executor.cc b/paddle/fluid/framework/parallel_executor.cc
index 39b47415ff7e378cabc79e668fe2be63eb71d87f..2c6e337568306502fbaa362015e51f81efc0a5ff 100644
--- a/paddle/fluid/framework/parallel_executor.cc
+++ b/paddle/fluid/framework/parallel_executor.cc
@@ -171,8 +171,17 @@ ParallelExecutor::ParallelExecutor(
   }
   // If the loss_var_name is given, the number of graph should be only one.
   if (loss_var_name.size()) {
-    PADDLE_ENFORCE_EQ(ir::GraphNum(*graph), 1,
-                      "The number of graph should be only one");
+    size_t graph_num = ir::GraphNum(*graph);
+    if (graph_num > 1) {
+      LOG(WARNING)
+          << "The number of graph should be only one, "
+             "but the current graph has "
+          << ir::GraphNum(*graph)
+          << " sub_graphs. If you want to see the nodes of the "
+             "sub_graphs, you should use 'FLAGS_print_sub_graph_dir' "
+             "to specify the output dir. NOTES: if you not do training, "
+             "please don't pass loss_var_name.";
+    }
   }
 
   if (exec_strategy.type_ == ExecutionStrategy::kDefault) {
diff --git a/python/paddle/fluid/__init__.py b/python/paddle/fluid/__init__.py
index b99197492870e9886e8a29f8c46723401a2a5ce1..f2f49f813a1840897024d851f2810786a153fb49 100644
--- a/python/paddle/fluid/__init__.py
+++ b/python/paddle/fluid/__init__.py
@@ -116,7 +116,8 @@ def __bootstrap__():
         'use_mkldnn', 'use_ngraph', 'initial_cpu_memory_in_mb',
         'init_allocated_mem', 'free_idle_memory', 'paddle_num_threads',
         "dist_threadpool_size", 'cpu_deterministic', 'eager_delete_tensor_gb',
-        'allocator_strategy', 'reader_queue_speed_test_mode'
+        'allocator_strategy', 'reader_queue_speed_test_mode',
+        'print_sub_graph_dir'
     ]
     if os.name != 'nt':
         read_env_flags.append('warpctc_dir')