diff --git a/deploy/cpp/demo/classifier.cpp b/deploy/cpp/demo/classifier.cpp
index 8b78d7e8975642b8d10d6c641c0245cdf661d3d9..6fd354d3f9cb6a366f0efb0b31e7ae073a90b4ad 100644
--- a/deploy/cpp/demo/classifier.cpp
+++ b/deploy/cpp/demo/classifier.cpp
@@ -37,6 +37,7 @@ DEFINE_int32(batch_size, 1, "Batch size of infering");
 DEFINE_int32(thread_num,
              omp_get_num_procs(),
              "Number of preprocessing threads");
+DEFINE_bool(use_ir_optim, true, "use ir optimization");
 
 int main(int argc, char** argv) {
   // Parsing command-line
@@ -57,7 +58,8 @@ int main(int argc, char** argv) {
              FLAGS_use_gpu,
              FLAGS_use_trt,
              FLAGS_gpu_id,
-             FLAGS_key);
+             FLAGS_key,
+             FLAGS_use_ir_optim);
 
   // 进行预测
   double total_running_time_s = 0.0;
diff --git a/deploy/cpp/demo/detector.cpp b/deploy/cpp/demo/detector.cpp
index 5b4e3a2ba9d2c921cf23774a17e34e0c8e26cc2a..54f93d2995fa24af73bba2855b6b26466129fa20 100644
--- a/deploy/cpp/demo/detector.cpp
+++ b/deploy/cpp/demo/detector.cpp
@@ -43,6 +43,7 @@ DEFINE_double(threshold,
 DEFINE_int32(thread_num,
              omp_get_num_procs(),
              "Number of preprocessing threads");
+DEFINE_bool(use_ir_optim, true, "use ir optimization");
 
 int main(int argc, char** argv) {
   // 解析命令行参数
@@ -62,7 +63,8 @@ int main(int argc, char** argv) {
              FLAGS_use_gpu,
              FLAGS_use_trt,
              FLAGS_gpu_id,
-             FLAGS_key);
+             FLAGS_key,
+             FLAGS_use_ir_optim);
 
   double total_running_time_s = 0.0;
   double total_imread_time_s = 0.0;
diff --git a/deploy/cpp/demo/segmenter.cpp b/deploy/cpp/demo/segmenter.cpp
index 7dd48e551890f3c8e4550694c45bc3f84088ec0a..90adb5aea860bf5ad9f6cb9019990a188c37f871 100644
--- a/deploy/cpp/demo/segmenter.cpp
+++ b/deploy/cpp/demo/segmenter.cpp
@@ -39,6 +39,7 @@ DEFINE_int32(batch_size, 1, "Batch size of infering");
 DEFINE_int32(thread_num,
              omp_get_num_procs(),
              "Number of preprocessing threads");
+DEFINE_bool(use_ir_optim, false, "use ir optimization");
 
 int main(int argc, char** argv) {
   // 解析命令行参数
@@ -59,7 +60,8 @@ int main(int argc, char** argv) {
              FLAGS_use_gpu,
              FLAGS_use_trt,
              FLAGS_gpu_id,
-             FLAGS_key);
+             FLAGS_key,
+             FLAGS_use_ir_optim);
 
   double total_running_time_s = 0.0;
   double total_imread_time_s = 0.0;
diff --git a/deploy/cpp/include/paddlex/paddlex.h b/deploy/cpp/include/paddlex/paddlex.h
index af4d8898496fee47ed9b5c74599536ddf1fe9f6c..e0d0569341198d0a0b2a8c6d0637c3f5a61e1f3f 100644
--- a/deploy/cpp/include/paddlex/paddlex.h
+++ b/deploy/cpp/include/paddlex/paddlex.h
@@ -72,20 +72,23 @@ class Model {
    * @param use_trt: use Tensor RT or not when infering
    * @param gpu_id: the id of gpu when infering with using gpu
    * @param key: the key of encryption when using encrypted model
+   * @param use_ir_optim: use ir optimization when infering
    * */
   void Init(const std::string& model_dir,
             bool use_gpu = false,
             bool use_trt = false,
             int gpu_id = 0,
-            std::string key = "") {
-    create_predictor(model_dir, use_gpu, use_trt, gpu_id, key);
+            std::string key = "",
+            bool use_ir_optim = true) {
+    create_predictor(model_dir, use_gpu, use_trt, gpu_id, key, use_ir_optim);
   }
 
   void create_predictor(const std::string& model_dir,
                         bool use_gpu = false,
                         bool use_trt = false,
                         int gpu_id = 0,
-                        std::string key = "");
+                        std::string key = "",
+                        bool use_ir_optim = true);
 
   /*
    * @brief
diff --git a/deploy/cpp/include/paddlex/visualize.h b/deploy/cpp/include/paddlex/visualize.h
index 5a8e39a762ff5bbde966aff213f1751f520789e2..9ddba5387b427c60645db7c96a54bcba76fa9898 100644
--- a/deploy/cpp/include/paddlex/visualize.h
+++ b/deploy/cpp/include/paddlex/visualize.h
@@ -22,9 +22,14 @@
 #include <io.h>
 #else  // Linux/Unix
 #include <dirent.h>
-#include <sys/io.h>
+// #include <sys/io.h>
+#ifdef __arm__  // for arm
+#include <aarch64-linux-gpu/sys/stat.h>
+#include <aarch64-linux-gpu/sys/types.h>
+#else
 #include <sys/stat.h>
 #include <sys/types.h>
+#endif
 #include <unistd.h>
 #endif
 #include <string>
diff --git a/deploy/cpp/src/paddlex.cpp b/deploy/cpp/src/paddlex.cpp
index bedd83b356baff41d7f9d16ac6de855e982332b2..cf1dfc955c43f9a61539e93a34c77c6ab4b198a9 100644
--- a/deploy/cpp/src/paddlex.cpp
+++ b/deploy/cpp/src/paddlex.cpp
@@ -22,7 +22,8 @@ void Model::create_predictor(const std::string& model_dir,
                              bool use_gpu,
                              bool use_trt,
                              int gpu_id,
-                             std::string key) {
+                             std::string key,
+                             bool use_ir_optim) {
   paddle::AnalysisConfig config;
   std::string model_file = model_dir + OS_PATH_SEP + "__model__";
   std::string params_file = model_dir + OS_PATH_SEP + "__params__";
@@ -63,6 +64,8 @@ void Model::create_predictor(const std::string& model_dir,
   }
   config.SwitchUseFeedFetchOps(false);
   config.SwitchSpecifyInputNames(true);
+  // 开启图优化
+  config.SwitchIrOptim(use_ir_optim);
   // 开启内存优化
   config.EnableMemoryOptim();
   if (use_trt) {