diff --git a/CMakeLists.txt b/CMakeLists.txt
index b8e104817d805f806882a97faa224376f56f746f..79764dd5e0bc11dae912bbba1dcc6d8443b2ec70 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -53,11 +53,9 @@ option(MGE_WITH_DISTRIBUTED "Build with distributed support" ON)
 option(MGE_BUILD_IMPERATIVE_RT "Build _imperative_rt Python Module " ON)
 option(MGE_BUILD_SDK "Build load_and_run" ON)
 option(MGE_INFERENCE_ONLY "Build inference only library." OFF)
-option(MGE_WITH_PYTHON_MODULE "Build MegEngine legacy Python Module." OFF)
 option(MGE_WITH_MKLDNN "Enable Intel MKL_DNN support," ON)
 option(MGE_WITH_ROCM "Enable ROCM support" OFF)
 
-
 if(NOT ${MGE_BIN_REDUCE} STREQUAL "")
     message(STATUS "build with BIN REDUCE")
     if(MGE_WITH_MINIMUM_SIZE)
diff --git a/dnn/src/CMakeLists.txt b/dnn/src/CMakeLists.txt
index 7bc6934478b3df7c7ae0a020cb8f1f47e739aa80..430a6b946d1d22b470b8a2241d49a5aacae0003f 100644
--- a/dnn/src/CMakeLists.txt
+++ b/dnn/src/CMakeLists.txt
@@ -184,7 +184,6 @@ if(MGE_WITH_MKLDNN AND ${MGE_ARCH} STREQUAL "x86_64")
         target_link_libraries(megdnn PRIVATE dnnl)
     endif()
 endif()
-target_link_libraries(megdnn PRIVATE ${MGE_CUDA_LIBS})
 if (BUILD_SHARED_LIBS)
     target_link_libraries(megdnn PRIVATE $<BUILD_INTERFACE:${MGE_BLAS_LIBS}>)
 else()
diff --git a/dnn/test/CMakeLists.txt b/dnn/test/CMakeLists.txt
index b1ea35146903883fb976c77b9e22bb9566cafa7f..8cfeef2139e2ea08d9a48a71bcb0940c59f067ad 100644
--- a/dnn/test/CMakeLists.txt
+++ b/dnn/test/CMakeLists.txt
@@ -48,7 +48,7 @@ endif()
 add_executable(megdnn_test ${SOURCES})
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-narrowing")
 target_link_libraries(megdnn_test gtest)
-target_link_libraries(megdnn_test megdnn ${MGE_BLAS_LIBS})
+target_link_libraries(megdnn_test megdnn ${MGE_BLAS_LIBS} ${MGE_CUDA_LIBS})
 
 if (MGE_WITH_CUDA)
     target_link_libraries(megdnn_test cutlass)
diff --git a/imperative/CMakeLists.txt b/imperative/CMakeLists.txt
index 8f3759708df9713f6c2a811e57b564013ea9a069..82b36187c81ea78403bd2820e69d981fa1267a48 100644
--- a/imperative/CMakeLists.txt
+++ b/imperative/CMakeLists.txt
@@ -28,22 +28,12 @@ if (APPLE)
     target_link_libraries(${MODULE_NAME} PRIVATE megengine_export)
 elseif (MSVC OR WIN32)
     # Windows does not support implicitly importing data members from DLL.
-    target_link_libraries(${MODULE_NAME} PRIVATE megbrain megdnn)
+    target_link_libraries(${MODULE_NAME} PRIVATE megbrain megdnn ${MGE_CUDA_LIBS})
     message(STATUS "CMAKE_MSVC_RUNTIME_LIBRARY: ${CMAKE_MSVC_RUNTIME_LIBRARY}")
     set_target_properties(${MODULE_NAME} PROPERTIES MSVC_RUNTIME_LIBRARY "${CMAKE_MSVC_RUNTIME_LIBRARY}")
 else()
-    if (MGE_WITH_PYTHON_MODULE)
-        # use to fix runtime crash when build both mgb(MGE_WITH_PYTHON_MODULE) and imperative(MGE_BUILD_IMPERATIVE_RT)
-        target_link_libraries(${MODULE_NAME} PRIVATE megengine_export -Wl,--version-script=${MGE_VERSION_SCRIPT})
-    else()
-        # use to reduce whl size by depend on megbrain/dnn directly, caused by cmake create two cuda fatbin
-        # elf section on both megengine_export and target which depend on megengine_export
-        target_link_libraries(${MODULE_NAME} PRIVATE megbrain megdnn -Wl,--version-script=${MGE_VERSION_SCRIPT})
-        if (MGE_WITH_DISTRIBUTED)
-            message(VERBOSE "Imperative configured to link megray")
-            target_link_libraries(${MODULE_NAME} PRIVATE megray)
-        endif()
-    endif()
+    # use to fix runtime crash when build both mgb(MGE_WITH_PYTHON_MODULE) and imperative(MGE_BUILD_IMPERATIVE_RT)
+    target_link_libraries(${MODULE_NAME} PRIVATE megengine_export -Wl,--version-script=${MGE_VERSION_SCRIPT})
 endif()
 
 add_subdirectory(${PROJECT_SOURCE_DIR}/third_party/range-v3 ${PROJECT_BINARY_DIR}/third_party/range-v3)
diff --git a/imperative/test/CMakeLists.txt b/imperative/test/CMakeLists.txt
index 52aec90d40fb80f26f004aa39ec2659ce8b01403..68d0dcecb748ea7f2ed9cd85c8cf7c4447a74c23 100644
--- a/imperative/test/CMakeLists.txt
+++ b/imperative/test/CMakeLists.txt
@@ -18,7 +18,8 @@ target_include_directories(imperative_test PRIVATE ${MODULE_SRC_INCLUDE} ${PYTHO
 target_compile_definitions(imperative_test PRIVATE MODULE_NAME=C)
 target_compile_options(imperative_test PRIVATE -Wno-unused-parameter)
 
-set(LINK_LIBS megbrain megdnn gtest gmock pybind11::embed range-v3)
+set(LINK_LIBS megbrain megdnn ${MGE_CUDA_LIBS} gtest gmock pybind11::embed range-v3)
+
 if(MGE_WITH_CUDA)
     list(APPEND LINK_LIBS cudart)
 endif()
diff --git a/scripts/whl/manylinux2010/do_build.sh b/scripts/whl/manylinux2010/do_build.sh
index f79b8ba54394b5358837dbc5cf413c3aecf2c221..1a8e158bc10ce6d5f80697b5eac43dea481ff069 100755
--- a/scripts/whl/manylinux2010/do_build.sh
+++ b/scripts/whl/manylinux2010/do_build.sh
@@ -34,6 +34,9 @@ function patch_elf_depend_lib() {
 
     patchelf --remove-rpath ${BUILD_DIR}/staging/megengine/core/_imperative_rt.so
     patchelf --force-rpath --set-rpath '$ORIGIN/lib' ${BUILD_DIR}/staging/megengine/core/_imperative_rt.so
+    cp ${BUILD_DIR}/src/libmegengine_export.so ${LIBS_DIR}
+    patchelf --remove-rpath ${LIBS_DIR}/libmegengine_export.so
+    patchelf --force-rpath --set-rpath '$ORIGIN/.' ${LIBS_DIR}/libmegengine_export.so
 
 
     if [ ${BUILD_WHL_CPU_ONLY} = "OFF" ]; then
diff --git a/sdk/load-and-run/CMakeLists.txt b/sdk/load-and-run/CMakeLists.txt
index 159fa2d6d34a7c0ef16f75417bef0c3714baff6d..4d46b533214799f23ecff84f1378585c524e04ab 100755
--- a/sdk/load-and-run/CMakeLists.txt
+++ b/sdk/load-and-run/CMakeLists.txt
@@ -4,7 +4,7 @@ add_executable (load_and_run  ${SOURCES})
 
 if (WIN32)
     # Windows does not support implicitly importing data members from DLL.
-    target_link_libraries(load_and_run megbrain megdnn)
+    target_link_libraries(load_and_run megbrain megdnn ${MGE_CUDA_LIBS})
 else()
     target_link_libraries (load_and_run megengine)
 endif()
diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
index a804f99a79a5f127deba1b66da23cc09f3dfb984..8c0db789c1e2a518eaaed0961be66449e3d5e3c7 100644
--- a/src/CMakeLists.txt
+++ b/src/CMakeLists.txt
@@ -92,7 +92,6 @@ if(MGE_WITH_DISTRIBUTED)
     target_include_directories(megbrain PRIVATE ${CPPZMQ_INC} ${CMAKE_CURRENT_BINARY_DIR})
     target_link_libraries (megbrain PRIVATE megray)
 endif()
-target_link_libraries(megbrain PRIVATE ${MGE_CUDA_LIBS})
 target_link_libraries(megbrain PUBLIC ${MGE_CAMBRICON_LIBS})
 target_link_libraries(megbrain PUBLIC ${MGE_ATLAS_LIBS})
 if(MGE_WITH_JIT AND MGE_WITH_HALIDE)
@@ -189,6 +188,7 @@ if(MGE_BUILD_IMPERATIVE_RT
     message(VERBOSE "create a export SHARED lib for python use")
     add_library(megengine_export SHARED)
     target_link_libraries(megengine_export PUBLIC megbrain megdnn)
+    target_link_libraries(megengine_export PRIVATE ${MGE_CUDA_LIBS})
     if (MGE_WITH_DISTRIBUTED)
         message(VERBOSE "megengine_export configured to link megray")
         target_link_libraries(megengine_export PUBLIC megray)
@@ -196,6 +196,7 @@ if(MGE_BUILD_IMPERATIVE_RT
 endif()
 # Build as SHARED or STATIC depending on BUILD_SHARED_LIBS=ON/OFF
 add_library(megengine)
+target_link_libraries(megengine PRIVATE ${MGE_CUDA_LIBS})
 target_link_libraries(megengine PUBLIC megbrain megdnn)
 if (UNIX AND NOT APPLE)
     # TODO: Use target_link_options after upgrading to CMake 3.13
diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt
index eecc194160f6e14b4501367b5fe42d4426511a63..2387a109ce7dceab88f45239cf8ca4c006289bae 100644
--- a/test/CMakeLists.txt
+++ b/test/CMakeLists.txt
@@ -18,11 +18,7 @@ endif()
 
 add_executable(megbrain_test ${SOURCES})
 target_link_libraries(megbrain_test gtest gmock)
-if(MSVC OR WIN32)
-    target_link_libraries(megbrain_test megbrain megdnn)
-else()
-    target_link_libraries(megbrain_test megengine)
-endif()
+target_link_libraries(megbrain_test megbrain megdnn ${MGE_CUDA_LIBS})
 if(CXX_SUPPORT_WCLASS_MEMACCESS)
     if(MGE_WITH_CUDA)
         target_compile_options(megbrain_test PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:-Xcompiler=-Wno-class-memaccess>"