diff --git a/cmake/generic.cmake b/cmake/generic.cmake
index 415eb451a986cd7e59829b9a8f2c744ecf464bd6..225a3c19a16435c4df6403ff7d1bdd01e628dd72 100644
--- a/cmake/generic.cmake
+++ b/cmake/generic.cmake
@@ -490,6 +490,9 @@ function(nv_binary TARGET_NAME)
     set(multiValueArgs SRCS DEPS)
     cmake_parse_arguments(nv_binary "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
     cuda_add_executable(${TARGET_NAME} ${nv_binary_SRCS})
+    target_link_libraries(${TARGET_NAME} ${CUDNN_LIBRARY} ${CUBLAS_LIBRARIES})
+    get_property(os_dependency_modules GLOBAL PROPERTY OS_DEPENDENCY_MODULES)
+    target_link_libraries(${TARGET_NAME} ${os_dependency_modules})
     if(nv_binary_DEPS)
       target_link_libraries(${TARGET_NAME} ${nv_binary_DEPS})
       add_dependencies(${TARGET_NAME} ${nv_binary_DEPS})
@@ -507,7 +510,7 @@ function(nv_test TARGET_NAME)
     cuda_add_executable(${TARGET_NAME} ${nv_test_SRCS})
     get_property(os_dependency_modules GLOBAL PROPERTY OS_DEPENDENCY_MODULES)
     target_link_libraries(${TARGET_NAME} ${nv_test_DEPS} lite_gtest_main gtest
-gflags glog ${os_dependency_modules} ${CUDNN_LIBRARY} ${CUBLAS_LIBRARIES} )
+       gflags glog ${os_dependency_modules} ${CUDNN_LIBRARY} ${CUBLAS_LIBRARIES} )
     add_dependencies(${TARGET_NAME} ${nv_test_DEPS} lite_gtest_main gtest gflags glog)
     common_link(${TARGET_NAME})
     add_test(${TARGET_NAME} ${TARGET_NAME})
diff --git a/cmake/lite.cmake b/cmake/lite.cmake
index 0061ddea32bf4b15a53b1e2742bae70f38c1e040..00d5cfdd8ee5e9947a331454f7ff4d78c93936b4 100644
--- a/cmake/lite.cmake
+++ b/cmake/lite.cmake
@@ -234,6 +234,7 @@ endfunction()
 
 set(arm_kernels CACHE INTERNAL "arm kernels")
 set(x86_kernels CACHE INTERNAL "x86 kernels")
+set(cuda_kernels CACHE INTERNAL "cuda kernels")
 set(fpga_kernels CACHE INTERNAL "fpga kernels")
 set(npu_kernels CACHE INTERNAL "npu kernels")
 set(opencl_kernels CACHE INTERNAL "opencl kernels")
diff --git a/lite/CMakeLists.txt b/lite/CMakeLists.txt
index 2da710e563c17360285924c9ed20c4bf4db08648..4597dd6219e51dc3ed054fc5b47844b32457c961 100644
--- a/lite/CMakeLists.txt
+++ b/lite/CMakeLists.txt
@@ -82,12 +82,10 @@ if (LITE_WITH_LIGHT_WEIGHT_FRAMEWORK AND LITE_WITH_ARM)
                 add_dependencies(publish_inference_cxx_lib bundle_full_api)
                 add_dependencies(publish_inference_cxx_lib bundle_light_api)
                 add_dependencies(publish_inference_cxx_lib test_model_bin)
-                if (ARM_TARGET_OS STREQUAL "android") 
                     add_dependencies(publish_inference_cxx_lib paddle_full_api_shared)
                     add_dependencies(publish_inference paddle_light_api_shared)
                     add_custom_command(TARGET publish_inference_cxx_lib 
                           COMMAND cp ${CMAKE_BINARY_DIR}/lite/api/*.so ${INFER_LITE_PUBLISH_ROOT}/cxx/lib)
-                endif()
                 add_dependencies(publish_inference publish_inference_cxx_lib)
                 add_custom_command(TARGET publish_inference_cxx_lib POST_BUILD
                         COMMAND ${CMAKE_STRIP} "--strip-debug" ${INFER_LITE_PUBLISH_ROOT}/cxx/lib/*.a)
@@ -103,7 +101,6 @@ if (LITE_WITH_LIGHT_WEIGHT_FRAMEWORK AND LITE_WITH_ARM)
             add_dependencies(tiny_publish_lib bundle_light_api)
             add_dependencies(publish_inference tiny_publish_lib)
         else()
-            if (ARM_TARGET_OS STREQUAL "android")
                 add_custom_target(tiny_publish_cxx_lib ${TARGET}
                     COMMAND mkdir -p "${INFER_LITE_PUBLISH_ROOT}/cxx"
                     COMMAND mkdir -p "${INFER_LITE_PUBLISH_ROOT}/cxx/include"
@@ -113,7 +110,6 @@ if (LITE_WITH_LIGHT_WEIGHT_FRAMEWORK AND LITE_WITH_ARM)
                 )
                 add_dependencies(tiny_publish_cxx_lib paddle_light_api_shared)
                 add_dependencies(publish_inference tiny_publish_cxx_lib)  
-            endif() 
         endif()
     endif()
 
diff --git a/lite/api/CMakeLists.txt b/lite/api/CMakeLists.txt
index 438d0f9c9744839848bd6021e595d91c9142baaf..093195b63a20f5a4c2e279a1e0d634be32aa69a3 100644
--- a/lite/api/CMakeLists.txt
+++ b/lite/api/CMakeLists.txt
@@ -9,13 +9,15 @@ if (LITE_ON_TINY_PUBLISH)
     set(CMAKE_C_FLAGS_RELEASE "-Os -DNDEBUG")
 endif()
 set(light_lib_DEPS light_api paddle_api paddle_api_light optimizer)
-if ((NOT LITE_ON_TINY_PUBLISH) AND (ARM_TARGET_OS STREQUAL "android"))
+#if ((NOT LITE_ON_TINY_PUBLISH) AND (ARM_TARGET_OS STREQUAL "android"))
+if (NOT LITE_ON_TINY_PUBLISH) 
     #full api dynamic library
     add_library(paddle_full_api_shared SHARED "")
     target_sources(paddle_full_api_shared PUBLIC ${__lite_cc_files} paddle_api.cc light_api.cc cxx_api.cc cxx_api_impl.cc light_api_impl.cc)
-    add_dependencies(paddle_full_api_shared op_list_h kernel_list_h framework_proto)
-    target_link_libraries(paddle_full_api_shared framework_proto)
-    
+    add_dependencies(paddle_full_api_shared op_list_h kernel_list_h framework_proto xxhash )
+    target_link_libraries(paddle_full_api_shared framework_proto xxhash ${math_cuda})# ${cuda_kernels})
+    target_link_libraries(paddle_full_api_shared "-Wl,--whole-archive"  ${cuda_kernels} "-Wl,--no-whole-archive")
+    add_dependencies(lite_compile_deps paddle_full_api_shared)
     #light api dynamic library
     lite_cc_library(paddle_light_api_shared MODULE
     SRCS light_api_shared.cc
@@ -23,11 +25,11 @@ if ((NOT LITE_ON_TINY_PUBLISH) AND (ARM_TARGET_OS STREQUAL "android"))
     ARM_DEPS ${arm_kernels} NPU_DEPS ${npu_kernels})
     target_link_libraries(paddle_light_api_shared ${light_lib_DEPS} ${arm_kernels} ${npu_kernels})
 else()
-    if (ARM_TARGET_OS STREQUAL "android")
+#    if (ARM_TARGET_OS STREQUAL "android")
         add_library(paddle_light_api_shared SHARED "")
         target_sources(paddle_light_api_shared PUBLIC ${__lite_cc_files} paddle_api.cc light_api.cc light_api_impl.cc)
         add_dependencies(paddle_light_api_shared op_list_h kernel_list_h)
-    endif()
+#    endif()
 endif()
 
 if (WITH_TESTING)
@@ -44,6 +46,7 @@ endif()
 
 message(STATUS "get ops ${ops}")
 message(STATUS "get X86 kernels ${x86_kernels}")
+message(STATUS "get CUDA kernels ${cuda_kernels}")
 message(STATUS "get Host kernels ${host_kernels}")
 message(STATUS "get ARM kernels ${arm_kernels}")
 message(STATUS "get NPU kernels ${npu_kernels}")
@@ -245,20 +248,28 @@ endif()
 
 # Some bins
 if(NOT IOS)
-  lite_cc_binary(test_model_bin SRCS model_test.cc DEPS paddle_api_full paddle_api_light gflags utils
-    ${ops} ${host_kernels}
-    ARM_DEPS ${arm_kernels}
-    NPU_DEPS ${npu_kernels}
-    CL_DEPS ${opencl_kernels}
-    FPGA_DEPS ${fpga_kernels}
-    X86_DEPS ${x86_kernels})
-  lite_cc_binary(benchmark_bin SRCS benchmark.cc DEPS paddle_api_full paddle_api_light gflags utils
-    ${ops} ${host_kernels}
-    ARM_DEPS ${arm_kernels}
-    NPU_DEPS ${npu_kernels}
-    CL_DEPS ${opencl_kernels}
-    FPGA_DEPS ${fpga_kernels}
-    X86_DEPS ${x86_kernels})
+  if(LITE_WITH_CUDA)
+
+    nv_binary(test_model_bin SRCS model_test.cc DEPS paddle_api_full paddle_api_light gflags utils
+        ${ops} ${host_kernels} ${cuda_kernels})
+    nv_binary(benchmark_bin SRCS benchmark.cc DEPS paddle_api_full paddle_api_light gflags utils
+         ${ops} ${host_kernels} ${cuda_kernels})
+  else()  
+     lite_cc_binary(test_model_bin SRCS model_test.cc DEPS paddle_api_full paddle_api_light gflags utils
+        ${ops} ${host_kernels}
+        ARM_DEPS ${arm_kernels}
+        NPU_DEPS ${npu_kernels}
+        CL_DEPS ${opencl_kernels}
+        FPGA_DEPS ${fpga_kernels}
+        X86_DEPS ${x86_kernels})
+      lite_cc_binary(benchmark_bin SRCS benchmark.cc DEPS paddle_api_full paddle_api_light gflags utils
+        ${ops} ${host_kernels}
+        ARM_DEPS ${arm_kernels}
+        NPU_DEPS ${npu_kernels}
+        CL_DEPS ${opencl_kernels}
+        FPGA_DEPS ${fpga_kernels}
+        X86_DEPS ${x86_kernels})
+  endif(LITE_WITH_CUDA)
 endif()
 
 #lite_cc_binary(cxx_api_bin SRCS cxx_api_bin.cc
diff --git a/lite/tests/kernels/CMakeLists.txt b/lite/tests/kernels/CMakeLists.txt
index 1e5fdbb34de3fd0b986e6ec635545fd114f42e5f..77d6b72a296715f0e08880bdaa334d81a550800d 100644
--- a/lite/tests/kernels/CMakeLists.txt
+++ b/lite/tests/kernels/CMakeLists.txt
@@ -1,54 +1,54 @@
-if((NOT LITE_WITH_OPENCL AND NOT LITE_WITH_FPGA) AND (LITE_WITH_X86 OR LITE_WITH_ARM))
-    lite_cc_test(test_kernel_scale_compute SRCS scale_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    lite_cc_test(test_kernel_power_compute SRCS power_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    lite_cc_test(test_kernel_shuffle_channel_compute SRCS shuffle_channel_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    lite_cc_test(test_kernel_yolo_box_compute SRCS yolo_box_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    lite_cc_test(test_fc SRCS fc_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    lite_cc_test(test_kernel_elementwise_compute SRCS elementwise_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    lite_cc_test(test_kernel_lrn_compute SRCS lrn_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    lite_cc_test(test_kernel_decode_bboxes_compute SRCS decode_bboxes_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    lite_cc_test(test_kernel_box_coder_compute SRCS box_coder_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    lite_cc_test(test_kernel_activation_compute SRCS activation_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    lite_cc_test(test_kernel_argmax_compute SRCS argmax_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    lite_cc_test(test_kernel_axpy_compute SRCS axpy_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    lite_cc_test(test_kernel_conv2d_transpose_compute SRCS conv2d_transpose_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    lite_cc_test(test_kernel_norm_compute SRCS norm_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    lite_cc_test(test_kernel_cast_compute SRCS cast_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    lite_cc_test(test_kernel_assign_compute SRCS assign_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    lite_cc_test(test_kernel_assign_value_compute SRCS assign_value_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    #lite_cc_test(test_kernel_sequence_softmax_compute SRCS sequence_softmax_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    #lite_cc_test(test_kernel_im2sequence_compute SRCS im2sequence_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    #lite_cc_test(test_kernel_compare_compute SRCS compare_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    #lite_cc_test(test_kernel_logical_xor_compute SRCS logical_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    #lite_cc_test(test_kernel_topk_compute SRCS topk_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    #lite_cc_test(test_kernel_increment_compute SRCS increment_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    #lite_cc_test(test_kernel_write_to_array_compute SRCS write_to_array_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    #lite_cc_test(test_kernel_read_from_array_compute SRCS read_from_array_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    lite_cc_test(test_kernel_box_clip_compute SRCS box_clip_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
+if((NOT LITE_WITH_OPENCL AND NOT LITE_WITH_FPGA) AND (LITE_WITH_CUDA OR LITE_WITH_X86 OR LITE_WITH_ARM))
+    lite_cc_test(test_kernel_scale_compute SRCS scale_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    lite_cc_test(test_kernel_power_compute SRCS power_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    lite_cc_test(test_kernel_shuffle_channel_compute SRCS shuffle_channel_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    lite_cc_test(test_kernel_yolo_box_compute SRCS yolo_box_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    lite_cc_test(test_fc SRCS fc_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    lite_cc_test(test_kernel_elementwise_compute SRCS elementwise_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    lite_cc_test(test_kernel_lrn_compute SRCS lrn_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    lite_cc_test(test_kernel_decode_bboxes_compute SRCS decode_bboxes_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    lite_cc_test(test_kernel_box_coder_compute SRCS box_coder_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    lite_cc_test(test_kernel_activation_compute SRCS activation_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    lite_cc_test(test_kernel_argmax_compute SRCS argmax_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    lite_cc_test(test_kernel_axpy_compute SRCS axpy_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    lite_cc_test(test_kernel_conv2d_transpose_compute SRCS conv2d_transpose_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    lite_cc_test(test_kernel_norm_compute SRCS norm_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    lite_cc_test(test_kernel_cast_compute SRCS cast_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    lite_cc_test(test_kernel_assign_compute SRCS assign_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    lite_cc_test(test_kernel_assign_value_compute SRCS assign_value_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    #lite_cc_test(test_kernel_sequence_softmax_compute SRCS sequence_softmax_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    #lite_cc_test(test_kernel_im2sequence_compute SRCS im2sequence_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    #lite_cc_test(test_kernel_compare_compute SRCS compare_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    #lite_cc_test(test_kernel_logical_xor_compute SRCS logical_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    #lite_cc_test(test_kernel_topk_compute SRCS topk_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    #lite_cc_test(test_kernel_increment_compute SRCS increment_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    #lite_cc_test(test_kernel_write_to_array_compute SRCS write_to_array_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    #lite_cc_test(test_kernel_read_from_array_compute SRCS read_from_array_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    lite_cc_test(test_kernel_box_clip_compute SRCS box_clip_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
 
 if(LITE_BUILD_EXTRA)
-    lite_cc_test(test_gru_unit SRCS gru_unit_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    lite_cc_test(test_kernel_sequence_pool_compute SRCS sequence_pool_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    lite_cc_test(test_kernel_reduce_max_compute SRCS reduce_max_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
+    lite_cc_test(test_gru_unit SRCS gru_unit_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    lite_cc_test(test_kernel_sequence_pool_compute SRCS sequence_pool_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    lite_cc_test(test_kernel_reduce_max_compute SRCS reduce_max_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
 endif()
-    lite_cc_test(test_kernel_pad2d_compute SRCS  pad2d_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    lite_cc_test(test_kernel_prior_box_compute SRCS  prior_box_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    lite_cc_test(test_kernel_negative_compute SRCS negative_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    lite_cc_test(test_kernel_bilinear_interp_compute SRCS bilinear_interp_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    lite_cc_test(test_kernel_nearest_interp_compute SRCS nearest_interp_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    lite_cc_test(test_kernel_shape_compute SRCS shape_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    lite_cc_test(test_kernel_crop_compute SRCS crop_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    lite_cc_test(test_kernel_sequence_expand_compute SRCS sequence_expand_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    lite_cc_test(test_kernel_squeeze_compute SRCS squeeze_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    lite_cc_test(test_kernel_unsqueeze_compute SRCS unsqueeze_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    lite_cc_test(test_kernel_slice_compute SRCS slice_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    lite_cc_test(test_kernel_expand_compute SRCS expand_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    lite_cc_test(test_kernel_matmul_compute SRCS matmul_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    lite_cc_test(test_kernel_reduce_mean_compute SRCS reduce_mean_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    lite_cc_test(test_kernel_stack_compute SRCS stack_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    lite_cc_test(test_kernel_range_compute SRCS range_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    lite_cc_test(test_kernel_affine_channel_compute SRCS affine_channel_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    lite_cc_test(test_kernel_anchor_generator_compute SRCS anchor_generator_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    #lite_cc_test(test_kernel_generate_proposals_compute SRCS generate_proposals_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
-    #lite_cc_test(test_kernel_roi_align_compute SRCS roi_align_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
+    lite_cc_test(test_kernel_pad2d_compute SRCS  pad2d_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    lite_cc_test(test_kernel_prior_box_compute SRCS  prior_box_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    lite_cc_test(test_kernel_negative_compute SRCS negative_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    lite_cc_test(test_kernel_bilinear_interp_compute SRCS bilinear_interp_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    lite_cc_test(test_kernel_nearest_interp_compute SRCS nearest_interp_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    lite_cc_test(test_kernel_shape_compute SRCS shape_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    lite_cc_test(test_kernel_crop_compute SRCS crop_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    lite_cc_test(test_kernel_sequence_expand_compute SRCS sequence_expand_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    lite_cc_test(test_kernel_squeeze_compute SRCS squeeze_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    lite_cc_test(test_kernel_unsqueeze_compute SRCS unsqueeze_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    lite_cc_test(test_kernel_slice_compute SRCS slice_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    lite_cc_test(test_kernel_expand_compute SRCS expand_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    lite_cc_test(test_kernel_matmul_compute SRCS matmul_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    lite_cc_test(test_kernel_reduce_mean_compute SRCS reduce_mean_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    lite_cc_test(test_kernel_stack_compute SRCS stack_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    lite_cc_test(test_kernel_range_compute SRCS range_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    lite_cc_test(test_kernel_affine_channel_compute SRCS affine_channel_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    lite_cc_test(test_kernel_anchor_generator_compute SRCS anchor_generator_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    #lite_cc_test(test_kernel_generate_proposals_compute SRCS generate_proposals_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
+    #lite_cc_test(test_kernel_roi_align_compute SRCS roi_align_compute_test.cc DEPS arena_framework ${x86_kernels} ${arm_kernels} ${lite_ops} ${host_kernels} ${cuda_kernels})
 endif()
diff --git a/lite/tools/ci_build.sh b/lite/tools/ci_build.sh
index 17ec0fdccd09fe357d78303afc4b2d71c7bd0309..b14f573ad1a5de6fb9658ba1b88c81fc24d40c22 100755
--- a/lite/tools/ci_build.sh
+++ b/lite/tools/ci_build.sh
@@ -162,6 +162,20 @@ function cmake_x86_for_CI {
     # make test_generated_code -j$NUM_CORES_FOR_COMPILE
 }
 
+# This method is only called in CI.
+function cmake_nv_for_CI {
+    prepare_workspace # fake an empty __generated_code__.cc to pass cmake.
+    cmake ..  -DCMAKE_BUILD_TYPE=RelWithDebInfo -DWITH_MKL=OFF -DLITE_WITH_CUDA=ON -DWITH_MKLDNN=OFF -DLITE_WITH_X86=OFF \
+          -DWITH_LITE=ON  -DWITH_PYTHON=OFF -DWITH_TESTING=OFF -DLITE_WITH_ARM=OFF
+
+    # Compile and execute the gen_code related test, so it will generate some code, and make the compilation reasonable.
+    # make test_gen_code -j$NUM_CORES_FOR_COMPILE
+    # make test_cxx_api -j$NUM_CORES_FOR_COMPILE
+    # ctest -R test_cxx_api
+    # ctest -R test_gen_code
+    # make test_generated_code -j$NUM_CORES_FOR_COMPILE
+}
+
 function cmake_gpu {
     prepare_workspace
     cmake .. " -DWITH_GPU=ON {common_flags} -DLITE_WITH_GPU=ON"
@@ -227,6 +241,18 @@ function build_test_server {
     test_model_optimize_tool_compile
 }
 
+# Build the code and run lite server tests. This is executed in the CI system.
+function build_test_server_nv {
+    mkdir -p ./build_nv
+    cd ./build_nv
+    export LD_LIBRARY_PATH="$LD_LIBRARY_PATH:$PWD/third_party/install/mklml/lib"
+    cmake_nv_for_CI
+    build
+
+    #test_server
+    #test_model_optimize_tool_compile
+}
+
 function build_test_train {
     mkdir -p ./build
     cd ./build
@@ -890,6 +916,10 @@ function main {
                 build_test_server
                 shift
                 ;;
+            build_test_server_nv)
+                build_test_server_nv
+                shift
+                ;;
             build_test_train)
                 build_test_train
                 shift