diff --git a/demo/csrc/cpp/classifier.cxx b/demo/csrc/cpp/classifier.cxx
index 79094c6d89..3f55d9524d 100644
--- a/demo/csrc/cpp/classifier.cxx
+++ b/demo/csrc/cpp/classifier.cxx
@@ -28,8 +28,10 @@ int main(int argc, char* argv[]) {
 
   // construct a classifier instance
   mmdeploy::Classifier classifier(mmdeploy::Model{ARGS_model}, context);
-
-
+  // warmup
+  for (int i = 0; i < 20; ++i) {
+    classifier.Apply(img);
+  }
 
   // apply the classifier; the result is an array-like class holding references to
   // `mmdeploy_classification_t`, will be released automatically on destruction
diff --git a/demo/csrc/cpp/detector.cxx b/demo/csrc/cpp/detector.cxx
index f6e77c9689..b91ea1b280 100644
--- a/demo/csrc/cpp/detector.cxx
+++ b/demo/csrc/cpp/detector.cxx
@@ -29,6 +29,11 @@ int main(int argc, char* argv[]) {
   // construct a detector instance
   mmdeploy::Detector detector(mmdeploy::Model{ARGS_model}, context);
 
+  // warmup
+  for (int i = 0; i < 20; ++i) {
+    detector.Apply(img);
+  }
+
   // apply the detector, the result is an array-like class holding references to
   // `mmdeploy_detection_t`, will be released automatically on destruction
   mmdeploy::Detector::Result dets = detector.Apply(img);
diff --git a/demo/csrc/cpp/pose_detector.cxx b/demo/csrc/cpp/pose_detector.cxx
index 577c2b6383..025afdd794 100644
--- a/demo/csrc/cpp/pose_detector.cxx
+++ b/demo/csrc/cpp/pose_detector.cxx
@@ -27,6 +27,12 @@ int main(int argc, char *argv[]) {
   context.Add(profiler);
 
   PoseDetector detector{Model(model_path), context};
+
+  // warmup
+  for (int i = 0; i < 20; ++i) {
+    detector.Apply(img);
+  }
+
   auto res = detector.Apply(img);
 
   for (int i = 0; i < res[0].length; i++) {
diff --git a/demo/csrc/cpp/restorer.cxx b/demo/csrc/cpp/restorer.cxx
index 639b29fd44..0b788a2514 100644
--- a/demo/csrc/cpp/restorer.cxx
+++ b/demo/csrc/cpp/restorer.cxx
@@ -30,6 +30,11 @@ int main(int argc, char* argv[]) {
   // construct a restorer instance
   mmdeploy::Restorer restorer{mmdeploy::Model{ARGS_model}, context};
 
+  // warmup
+  for (int i = 0; i < 20; ++i) {
+    restorer.Apply(img);
+  }
+
   // apply restorer to the image
   mmdeploy::Restorer::Result result = restorer.Apply(img);
 
diff --git a/demo/csrc/cpp/rotated_detector.cxx b/demo/csrc/cpp/rotated_detector.cxx
index 24bf1f99af..6c5b9f1d37 100644
--- a/demo/csrc/cpp/rotated_detector.cxx
+++ b/demo/csrc/cpp/rotated_detector.cxx
@@ -30,6 +30,11 @@ int main(int argc, char* argv[]) {
   // construct a detector instance
   mmdeploy::RotatedDetector detector(mmdeploy::Model{ARGS_model}, context);
 
+  // warmup
+  for (int i = 0; i < 20; ++i) {
+    detector.Apply(img);
+  }
+
   // apply the detector, the result is an array-like class holding references to
   // `mmdeploy_rotated_detection_t`, will be released automatically on destruction
   mmdeploy::RotatedDetector::Result dets = detector.Apply(img);
diff --git a/demo/csrc/cpp/segmentor.cxx b/demo/csrc/cpp/segmentor.cxx
index a2aca27bca..c12f8f4153 100644
--- a/demo/csrc/cpp/segmentor.cxx
+++ b/demo/csrc/cpp/segmentor.cxx
@@ -33,6 +33,11 @@ int main(int argc, char* argv[]) {
   context.Add(profiler);
   mmdeploy::Segmentor segmentor{mmdeploy::Model{ARGS_model}, context};
 
+  // warmup
+  for (int i = 0; i < 20; ++i) {
+    segmentor.Apply(img);
+  }
+
   // apply the detector, the result is an array-like class holding a reference to
   // `mmdeploy_segmentation_t`, will be released automatically on destruction
   mmdeploy::Segmentor::Result seg = segmentor.Apply(img);
diff --git a/tools/scripts/build_ubuntu_x64_ncnn.py b/tools/scripts/build_ubuntu_x64_ncnn.py
index 015381ebac..2f7ab48cac 100644
--- a/tools/scripts/build_ubuntu_x64_ncnn.py
+++ b/tools/scripts/build_ubuntu_x64_ncnn.py
@@ -60,7 +60,7 @@ def install_pyncnn(dep_dir):
     # git clone
     if not os.path.exists('ncnn'):
         os.system(
-            'git clone --depth 1 --branch 20230517 https://github.com/tencent/ncnn && cd ncnn'  # noqa: E501
+            'git clone --depth 1 --branch 20230816 https://github.com/tencent/ncnn && cd ncnn'  # noqa: E501
         )
 
     ncnn_dir = os.path.join(dep_dir, 'ncnn')
@@ -90,7 +90,7 @@ def install_pyncnn(dep_dir):
 
     # install
     os.chdir(ncnn_dir)
-    os.system('cd python && python -m pip install -e .')
+    os.system('cd python && python -m pip install -e .  --user --no-cache-dir')
     ncnn_cmake_dir = os.path.join(ncnn_dir, 'build', 'install', 'lib', 'cmake',
                                   'ncnn')
     assert (os.path.exists(ncnn_cmake_dir))
@@ -130,7 +130,7 @@ def install_mmdeploy(work_dir, dep_dir, ncnn_cmake_dir):
     os.system(cmd)
 
     os.system('cd build && make -j {} && make install'.format(g_jobs))
-    os.system('python3 -m pip install -v -e .')
+    os.system('python3 -m pip install -v -e . --user --no-cache-dir')
     os.system(""" echo 'export PATH={}:$PATH' >> ~/mmdeploy.env """.format(
         os.path.join(work_dir, 'mmdeploy', 'backend', 'ncnn')))
     try: