diff --git a/.github/workflows/release_packages.yml b/.github/workflows/release_packages.yml
index fb77cd8a..54f7d856 100644
--- a/.github/workflows/release_packages.yml
+++ b/.github/workflows/release_packages.yml
@@ -56,10 +56,14 @@ jobs:
         VERSION_NUMBER=$(echo "$TAG_NAME" | sed 's/^v//')
         source /root/.bashrc
 
+        export AS_RELEASE_VERSION=$VERSION_NUMBER
+        export AS_BUILD_PACKAGE=ON
+
         if command -v nvcc &> /dev/null
         then
             export AS_PLATFORM="cuda"
             export AS_CUDA_SM="'70;75;80;86;89;90a'"
+            bash scripts/release/cpp_build_cuda.sh
         else
             # export ENABLE_MULTINUMA="ON"
             if [[ "${{ matrix.arch }}" == "ARM64" ]]; then
@@ -67,11 +71,8 @@ jobs:
             else
                 export AS_PLATFORM="x86"
             fi
+            bash build.sh
         fi
-
-        export AS_RELEASE_VERSION=$VERSION_NUMBER
-        export AS_BUILD_PACKAGE=ON
-        bash build.sh
   
     - name: Upload tgz package
       uses: actions/upload-artifact@v3
diff --git a/scripts/release/cpp_build_cuda.sh b/scripts/release/cpp_build_cuda.sh
new file mode 100644
index 00000000..3b9232d9
--- /dev/null
+++ b/scripts/release/cpp_build_cuda.sh
@@ -0,0 +1,11 @@
+#!/bin/bash
+set -e -x
+
+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-${CUDA_VERSION}/compat/libcuda.so.1 local_cuda_libs/libcuda.so.1
+export LD_LIBRARY_PATH=${PWD}/local_cuda_libs:${LD_LIBRARY_PATH}
+
+bash build.sh
diff --git a/scripts/release/python_manylinux_build_cuda.sh b/scripts/release/python_manylinux_build_cuda.sh
index cd01b52d..32d3ee1e 100755
--- a/scripts/release/python_manylinux_build_cuda.sh
+++ b/scripts/release/python_manylinux_build_cuda.sh
@@ -4,7 +4,7 @@ set -e -x
 # ALL_VERSION="3.8 3.9 3.10 3.11"
 ALL_VERSION="3.8"
 BUILD_VERSION=${@:-$ALL_VERSION}
-CUDA_VERSION=12.4
+CUDA_VERSION=$(nvcc --version | grep -oP 'release \K[\d.]+')
 
 echo " going to build python wheels with version: ${BUILD_VERSION}"