diff --git a/cmake/cudnn.cmake b/cmake/cudnn.cmake
index 842b94d47e75b4bab577a1150cb3d198eb42ebaf..c0cb4ccea67cd493a30a6be43ee6ee48f70c36bf 100644
--- a/cmake/cudnn.cmake
+++ b/cmake/cudnn.cmake
@@ -26,7 +26,8 @@ list(APPEND CUDNN_CHECK_LIBRARY_DIRS
     ${CUDNN_ROOT}/lib64
     ${CUDNN_ROOT}/lib
     ${CUDNN_ROOT}/lib/${TARGET_ARCH}-linux-gnu
-    ${CUDNN_ROOT}/local/cuda-${CUDA_VERSION}/targets/${TARGET_ARCH}-linux/lib/
+    /usr/local/cuda-${CUDA_VERSION}/targets/${TARGET_ARCH}-linux/lib/
+    /usr/lib/${TARGET_ARCH}-linux-gnu/
     $ENV{CUDNN_ROOT}
     $ENV{CUDNN_ROOT}/lib64
     $ENV{CUDNN_ROOT}/lib
diff --git a/lite/kernels/cuda/concat_compute.cu b/lite/kernels/cuda/concat_compute.cu
index 9ec693667252e76a99f305bc3ec9d6062cc2e840..72d0af459b26b6864dd308fece7131224b212a05 100644
--- a/lite/kernels/cuda/concat_compute.cu
+++ b/lite/kernels/cuda/concat_compute.cu
@@ -51,9 +51,9 @@ void ConcatCompute<Dtype>::Run() {
   Tensor* output = param.output;
   auto* output_data = output->mutable_data<Dtype>(TARGET(kCUDA));
   int axis = param.axis;
-  auto* axis_tensor = param.axis_tensor;
+  Tensor* axis_tensor = param.axis_tensor;
   if (axis_tensor != nullptr) {
-    auto* axis_tensor_data = axis_tensor->data<int>();
+    const int* axis_tensor_data = axis_tensor->data<int>();
     axis = axis_tensor_data[0];
   }
   int inner_size = 1;
diff --git a/lite/tools/ci_build.sh b/lite/tools/ci_build.sh
index 6c3750e69fa6eb1ecc36a1fc7e42fe32bead377a..8b5741a7a68bee3e783dff68e4bd4a8fc7cd8527 100755
--- a/lite/tools/ci_build.sh
+++ b/lite/tools/ci_build.sh
@@ -1,9 +1,10 @@
 #!/bin/bash
+# The git version of CI is 2.7.4. This script is not compatible with git version 1.7.1.
 set -ex
 
 TESTS_FILE="./lite_tests.txt"
 LIBS_FILE="./lite_libs.txt"
-
+CUDNN_ROOT="/usr/local/cudnn"
 
 readonly ADB_WORK_DIR="/data/local/tmp"
 readonly common_flags="-DWITH_LITE=ON -DLITE_WITH_LIGHT_WEIGHT_FRAMEWORK=OFF -DWITH_PYTHON=OFF -DWITH_TESTING=ON -DLITE_WITH_ARM=OFF"
@@ -162,6 +163,12 @@ function cmake_x86_for_CI {
     # make test_generated_code -j$NUM_CORES_FOR_COMPILE
 }
 
+function cmake_cuda_for_CI {
+    prepare_workspace # fake an empty __generated_code__.cc to pass cmake.
+    cmake ..  -DLITE_WITH_CUDA=ON -DWITH_MKLDNN=OFF -DLITE_WITH_X86=OFF ${common_flags} -DLITE_WITH_PROFILE=ON -DWITH_MKL=OFF \
+        -DLITE_BUILD_EXTRA=ON -DCUDNN_ROOT=${CUDNN_ROOT}
+}
+
 function cmake_gpu {
     prepare_workspace
     cmake .. " -DWITH_GPU=ON {common_flags} -DLITE_WITH_GPU=ON"
@@ -226,6 +233,16 @@ function build_test_server {
     test_model_optimize_tool_compile
 }
 
+# The CUDA version of CI is cuda_10.1.243_418.87.00_linux.
+# The cuDNN version is cudnn-10.1-linux-x64-v7.5.0.56.
+function build_test_cuda_server {
+    mkdir -p ./build
+    cd ./build
+    export LD_LIBRARY_PATH="$LD_LIBRARY_PATH:$PWD/third_party/install/mklml/lib"
+    cmake_cuda_for_CI
+    build
+}
+
 function build_test_train {
     mkdir -p ./build
     cd ./build
@@ -950,6 +967,10 @@ function main {
                 test_arm_android $TEST_NAME $ARM_PORT
                 shift
                 ;;
+            build_test_cuda_server)
+                build_test_cuda_server
+                shift
+                ;;
             build_test_server)
                 build_test_server
                 shift