diff --git a/.github/workflows/release_packages.yml b/.github/workflows/release_packages.yml
index 54f7d856..01a67c32 100644
--- a/.github/workflows/release_packages.yml
+++ b/.github/workflows/release_packages.yml
@@ -191,6 +191,6 @@ jobs:
         path: release/
 
     - name: Release all packages
-      uses: softprops/action-gh-release@v2
+      uses: softprops/action-gh-release@v1
       with:
         files: release/*
diff --git a/cmake/install.cmake b/cmake/install.cmake
index d24c330b..e1dff8dc 100644
--- a/cmake/install.cmake
+++ b/cmake/install.cmake
@@ -23,16 +23,11 @@ endif()
 
 if (BUILD_PYTHON)
 if (PYTHON_LIB_DIRS)
-    if(NOT ENABLE_NV_STATIC_LIB)
-        install(DIRECTORY ${CMAKE_INSTALL_PREFIX}/${CMAKE_INSTALL_LIBDIR} DESTINATION ${PYTHON_LIB_DIRS} FILES_MATCHING PATTERN "*.so" PATTERN "libnccl.*" EXCLUDE)
-    else()
-        install(DIRECTORY ${CMAKE_INSTALL_PREFIX}/${CMAKE_INSTALL_LIBDIR} DESTINATION ${PYTHON_LIB_DIRS} FILES_MATCHING PATTERN "*.so")
-    endif()
+    install(DIRECTORY ${CMAKE_INSTALL_PREFIX}/${CMAKE_INSTALL_LIBDIR} DESTINATION ${PYTHON_LIB_DIRS} FILES_MATCHING PATTERN "libflash-attn.so")
 
     if (ENABLE_MULTINUMA)
-        install(DIRECTORY ${CMAKE_INSTALL_PREFIX}/${CMAKE_INSTALL_BINDIR} DESTINATION ${PYTHON_LIB_DIRS} USE_SOURCE_PERMISSIONS FILES_MATCHING PATTERN "*")
         install(TARGETS _allspark_client DESTINATION ${PYTHON_LIB_DIRS})
     endif()
-        install(TARGETS _allspark DESTINATION ${PYTHON_LIB_DIRS})
+    install(TARGETS _allspark DESTINATION ${PYTHON_LIB_DIRS})
 endif()
 endif()
diff --git a/scripts/release/python_manylinux_build_cuda.sh b/scripts/release/python_manylinux_build_cuda.sh
index ee65d190..661f226b 100755
--- a/scripts/release/python_manylinux_build_cuda.sh
+++ b/scripts/release/python_manylinux_build_cuda.sh
@@ -3,7 +3,6 @@ set -e -x
 
 ALL_VERSION="3.8 3.9 3.10 3.11"
 BUILD_VERSION=${@:-$ALL_VERSION}
-CUDA_VERSION=$(nvcc --version | grep -oP 'release \K[\d.]+')
 
 echo " going to build python wheels with version: ${BUILD_VERSION}"
 
@@ -19,6 +18,8 @@ architecture=$(arch)
 export PLAT=manylinux2014_x86_64
 export AS_PLATFORM=cuda
 
+CUDA_VERSION=$(nvcc --version | grep -oP 'release \K[\d.]+')
+
 mkdir -p local_cuda_libs
 ln -sf /usr/local/cuda-${CUDA_VERSION}/targets/x86_64-linux/lib/stubs/libnvidia-ml.so local_cuda_libs/libnvidia-ml.so.1
 ln -sf /usr/local/cuda-${CUDA_VERSION}/compat/libcuda.so.1 local_cuda_libs/libcuda.so.1
@@ -36,7 +37,8 @@ function repair_wheel {
         echo "Skipping non-platform wheel $wheel"
     else
         # TODO: add lib path to build lib path
-        auditwheel repair "$wheel" --plat "$PLAT" -w ${REPO_ROOT}/python/wheelhouse/
+        auditwheel repair "$wheel" --plat "$PLAT" -w ${REPO_ROOT}/python/wheelhouse/ --exclude libcublas.so.12 --exclude libcublasLt.so.12 --exclude libcudart.so.12 --exclude libcusparse.so.12 --exclude libnvJitLink.so.12 --exclude libcuda.so.1 --exclude libnccl.so.2 --exclude libnvidia-ml.so.1 --exclude libcudart.so.12
+
     fi
 }