diff --git a/.github/workflows/ubuntu-cpu-wheels.yml b/.github/workflows/ubuntu-cpu-wheels.yml
index 635bb35..69e1b5f 100644
--- a/.github/workflows/ubuntu-cpu-wheels.yml
+++ b/.github/workflows/ubuntu-cpu-wheels.yml
@@ -2,8 +2,6 @@ name: build-wheels-cpu-ubuntu
 
 on:
   push:
-    branches:
-      - wheels
     tags:
       - '*'
   workflow_dispatch:
diff --git a/.github/workflows/ubuntu-cuda-wheels.yml b/.github/workflows/ubuntu-cuda-wheels.yml
new file mode 100644
index 0000000..24af57f
--- /dev/null
+++ b/.github/workflows/ubuntu-cuda-wheels.yml
@@ -0,0 +1,138 @@
+name: build-wheels-cuda-ubuntu
+
+on:
+  push:
+    branches:
+      - wheels
+    tags:
+      - '*'
+  workflow_dispatch:
+
+concurrency:
+  group: build-wheels-cuda-ubuntu-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  generate_build_matrix:
+    # see https://github.com/pytorch/pytorch/pull/50633
+    runs-on: ubuntu-latest
+    outputs:
+      matrix: ${{ steps.set-matrix.outputs.matrix }}
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+      - name: Generating build matrix
+        id: set-matrix
+        run: |
+          # outputting for debugging purposes
+          python ./scripts/github_actions/generate_build_matrix.py --enable-cuda
+          MATRIX=$(python ./scripts/github_actions/generate_build_matrix.py --enable-cuda)
+          echo "::set-output name=matrix::${MATRIX}"
+
+  build-manylinux-wheels:
+    needs: generate_build_matrix
+    name: ${{ matrix.torch }} ${{ matrix.python-version }}
+    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        ${{ fromJson(needs.generate_build_matrix.outputs.matrix) }}
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: display
+        shell: bash
+        run: |
+          function retry() {
+            $* || (sleep 1 && $*) || (sleep 2 && $*) || (sleep 4 && $*) || (sleep 8 && $*)
+          }
+
+          retry ls -lh
+
+      # see https://github.com/pytorch/test-infra/blob/9e3d392690719fac85bad0c9b67f530e48375ca1/tools/scripts/generate_binary_build_matrix.py
+      # https://github.com/pytorch/builder/tree/main/manywheel
+      # https://github.com/pytorch/builder/pull/476
+      # https://github.com/k2-fsa/k2/issues/733
+      # https://github.com/pytorch/pytorch/pull/50633  (generate build matrix)
+      - name: Run the build process with Docker
+        uses: addnab/docker-run-action@v3
+        with:
+            image: ${{ matrix.image }}
+            options: -v ${{ github.workspace }}:/var/www -e PYTHON_VERSION=${{ matrix.python-version }} -e TORCH_VERSION=${{ matrix.torch }} -e CUDA_VERSION=${{ matrix.cuda }}
+            run: |
+              echo "pwd: $PWD"
+              uname -a
+              id
+              cat /etc/*release
+              gcc --version
+              python3 --version
+              which python3
+
+              pushd /usr/local
+              rm cuda
+              ln -s cuda-$CUDA_VERSION cuda
+              popd
+              which nvcc
+              nvcc --version
+
+              /var/www/scripts/github_actions/build-ubuntu-cuda.sh
+
+      - name: Display wheels
+        shell: bash
+        run: |
+
+          function retry() {
+            $* || (sleep 1 && $*) || (sleep 2 && $*) || (sleep 4 && $*) || (sleep 8 && $*)
+          }
+
+          ls -lh ./wheelhouse/
+
+          ls -lh ./wheelhouse/*.whl
+
+      - name: Upload Wheel
+        uses: actions/upload-artifact@v2
+        with:
+          name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cuda
+          path: wheelhouse/*.whl
+
+      # https://huggingface.co/docs/hub/spaces-github-actions
+      - name: Publish to huggingface
+        if: github.repository_owner == 'csukuangfj'
+        shell: bash
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+        run: |
+          git config --global user.email "csukuangfj@gmail.com"
+          git config --global user.name "Fangjun Kuang"
+
+          export GIT_LFS_SKIP_SMUDGE=1
+          git clone https://huggingface.co/csukuangfj/kaldifeat huggingface
+          cur_path=$PWD
+
+          function push_to_huggingface() {
+            cd $cur_path
+            rm -rf huggingface
+            git clone https://huggingface.co/csukuangfj/kaldifeat huggingface
+            cd huggingface
+            git pull
+
+            mkdir -p ubuntu-cuda
+            cp -v ../wheelhouse/*.whl ./ubuntu-cuda
+            git status
+            git lfs track "*.whl"
+            git add .
+            git commit -m "upload ubuntu-cuda wheel for torch ${{ matrix.torch }} python ${{ matrix.python-version }}"
+            git push https://csukuangfj:$HF_TOKEN@huggingface.co/csukuangfj/kaldifeat main
+
+            cd ..
+            rm -rf huggingface
+          }
+          function retry() {
+            $* || (sleep 1 && $*) || (sleep 2 && $*) || (sleep 4 && $*) || (sleep 8 && $*)
+          }
+
+          retry push_to_huggingface
diff --git a/scripts/github_actions/build-ubuntu-cuda.sh b/scripts/github_actions/build-ubuntu-cuda.sh
new file mode 100755
index 0000000..e812dd0
--- /dev/null
+++ b/scripts/github_actions/build-ubuntu-cuda.sh
@@ -0,0 +1,101 @@
+#!/usr/bin/env bash
+#
+set -ex
+
+if [ -z $PYTHON_VERSION ]; then
+  echo "Please set the environment variable PYTHON_VERSION"
+  echo "Example: export PYTHON_VERSION=3.8"
+  # Valid values: 3.6, 3.7, 3.8, 3.9, 3.10, 3.11
+  exit 1
+fi
+
+if [ -z $TORCH_VERSION ]; then
+  echo "Please set the environment variable TORCH_VERSION"
+  echo "Example: export TORCH_VERSION=1.10.0"
+  exit 1
+fi
+
+if [ -z $CUDA_VERSION ]; then
+  echo "Please set the environment variable CUDA_VERSION"
+  echo "Example: export CUDA_VERSION=10.2"
+  # valid values: 10.2, 11.1, 11.3, 11.6, 11.7, 11.8
+  exit 1
+fi
+
+
+yum -y install openssl-devel bzip2-devel libffi-devel xz-devel wget redhat-lsb-core
+
+
+echo "Installing ${PYTHON_VERSION}.3"
+curl -O https://www.python.org/ftp/python/${PYTHON_VERSION}.3/Python-${PYTHON_VERSION}.3.tgz
+tar xf Python-${PYTHON_VERSION}.3.tgz
+pushd Python-${PYTHON_VERSION}.3
+
+PYTHON_INSTALL_DIR=$PWD/py-${PYTHON_VERSION}
+
+if [[ $PYTHON_VERSION =~ 3.1. ]]; then
+  yum install -y openssl11-devel
+  sed -i 's/PKG_CONFIG openssl /PKG_CONFIG openssl11 /g' configure
+fi
+
+./configure --enable-shared --prefix=$PYTHON_INSTALL_DIR
+make install
+
+popd
+
+export PATH=$PYTHON_INSTALL_DIR/bin:$PATH
+export LD_LIBRARY_PATH=$PYTHON_INSTALL_DIR/lib:$LD_LIBRARY_PATH
+ls -lh $PYTHON_INSTALL_DIR/lib/
+
+python3 --version
+which python3
+
+if [[ $PYTHON_VERSION != 3.6 ]]; then
+  curl -O https://bootstrap.pypa.io/get-pip.py
+  python3 get-pip.py
+fi
+
+python3 -m pip install scikit-build
+python3 -m pip install -U pip cmake
+python3 -m pip install wheel twine typing_extensions
+python3 -m pip install bs4 requests tqdm auditwheel
+
+echo "Installing torch"
+./install_torch.sh
+
+rm -rf ~/.cache/pip
+yum clean all
+
+cd /var/www
+
+export CMAKE_CUDA_COMPILER_LAUNCHER=
+export KALDIFEAT_CMAKE_ARGS=" -DPYTHON_EXECUTABLE=$PYTHON_INSTALL_DIR/bin/python3 "
+export KALDIFEAT_MAKE_ARGS=" -j2 "
+
+
+python3 setup.py bdist_wheel
+
+auditwheel --verbose repair \
+  --exclude libc10.so \
+  --exclude libc10_cuda.so \
+  --exclude libcuda.so.1 \
+  --exclude libcudart.so.${CUDA_VERSION} \
+  --exclude libnvToolsExt.so.1 \
+  --exclude libnvrtc.so.${CUDA_VERSION} \
+  --exclude libtorch.so \
+  --exclude libtorch_cpu.so \
+  --exclude libtorch_cuda.so \
+  --exclude libtorch_python.so \
+  \
+  --exclude libcudnn.so.8 \
+  --exclude libcublas.so.11 \
+  --exclude libcublasLt.so.11 \
+  --exclude libcudart.so.11.0 \
+  --exclude libnvrtc.so.11.2 \
+  --exclude libtorch_cuda_cu.so \
+  --exclude libtorch_cuda_cpp.so \
+  --plat manylinux_2_17_x86_64 \
+  -w /var/www/wheelhouse \
+  dist/*.whl
+
+ls -lh  /var/www
diff --git a/scripts/github_actions/generate_build_matrix.py b/scripts/github_actions/generate_build_matrix.py
index 354b984..e135c57 100755
--- a/scripts/github_actions/generate_build_matrix.py
+++ b/scripts/github_actions/generate_build_matrix.py
@@ -160,11 +160,19 @@ def generate_build_matrix(enable_cuda, for_windows, for_macos, test_only_latest_
         if enable_cuda:
             for p in python_versions:
                 for c in cuda_versions:
-                    ans.append({"torch": torch, "python-version": p, "cuda": c})
+                    if c == "10.1":
+                        # no docker image for cuda 10.1
+                        continue
+                    ans.append(
+                        {
+                            "torch": torch,
+                            "python-version": p,
+                            "cuda": c,
+                            "image": f"pytorch/manylinux-builder:cuda{c}",
+                        }
+                    )
         else:
             for p in python_versions:
-                if p != "3.6":
-                    continue
                 if for_windows or for_macos:
                     p = "cp" + "".join(p.split("."))
                     ans.append({"torch": torch, "python-version": p})
diff --git a/scripts/github_actions/install_torch.sh b/scripts/github_actions/install_torch.sh
index 9f0be7c..7ba7485 100755
--- a/scripts/github_actions/install_torch.sh
+++ b/scripts/github_actions/install_torch.sh
@@ -14,6 +14,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+torch=$TORCH_VERSION
+cuda=$CUDA_VERSION
 case ${torch} in
   1.5.*)
     case ${cuda} in
@@ -155,6 +157,18 @@ case ${torch} in
         ;;
     esac
     ;;
+  2.0.*)
+    case ${cuda} in
+      11.7)
+        package="torch==${torch}+cu117"
+        url=https://download.pytorch.org/whl/torch_stable.html
+        ;;
+      11.8)
+        package="torch==${torch}+cu118"
+        url=https://download.pytorch.org/whl/torch_stable.html
+        ;;
+    esac
+    ;;
   *)
     echo "Unsupported PyTorch version: ${torch}"
     exit 1
@@ -170,3 +184,5 @@ if [ x"${url}" == "x" ]; then
 else
   retry python3 -m pip install -q $package -f $url
 fi
+
+rm -rfv ~/.cache/pip