diff --git a/.github/workflows/macos-cpu-wheels.yml b/.github/workflows/macos-cpu-wheels.yml
index 5cd97a6..50d8e7c 100644
--- a/.github/workflows/macos-cpu-wheels.yml
+++ b/.github/workflows/macos-cpu-wheels.yml
@@ -84,7 +84,7 @@ jobs:
           ls -lh ./wheelhouse/
 
       - name: Upload Wheel
-        uses: actions/upload-artifact@v3
+        uses: actions/upload-artifact@v4
         with:
           name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-macos-latest-cpu
           path: wheelhouse/*.whl
diff --git a/.github/workflows/run-tests-windows-cpu.yml b/.github/workflows/run-tests-windows-cpu.yml
index 8d33247..d02568c 100644
--- a/.github/workflows/run-tests-windows-cpu.yml
+++ b/.github/workflows/run-tests-windows-cpu.yml
@@ -107,7 +107,7 @@ jobs:
           python3 -c "import kaldifeat; print(kaldifeat.__version__)"
 
       - name: Upload Wheel
-        uses: actions/upload-artifact@v3
+        uses: actions/upload-artifact@v4
         with:
           name: python-${{ matrix.python-version }}-${{ matrix.os }}-cpu
           path: dist/*.whl
diff --git a/.github/workflows/run-tests-windows-cuda.yml b/.github/workflows/run-tests-windows-cuda.yml
index 9845b56..21cd46e 100644
--- a/.github/workflows/run-tests-windows-cuda.yml
+++ b/.github/workflows/run-tests-windows-cuda.yml
@@ -159,7 +159,7 @@ jobs:
           python3 -c "import kaldifeat; print(kaldifeat.__version__)"
 
       - name: Upload Wheel
-        uses: actions/upload-artifact@v3
+        uses: actions/upload-artifact@v4
         with:
           name: python-${{ matrix.python-version }}-${{ matrix.os }}-cuda-${{ matrix.cuda }}
           path: dist/*.whl
diff --git a/.github/workflows/ubuntu-arm64-cpu-wheels.yml b/.github/workflows/ubuntu-arm64-cpu-wheels.yml
new file mode 100644
index 0000000..48b8d83
--- /dev/null
+++ b/.github/workflows/ubuntu-arm64-cpu-wheels.yml
@@ -0,0 +1,168 @@
+name: build-wheels-cpu-arm64-ubuntu
+
+on:
+  push:
+    branches:
+      # - wheel
+      - fix-cuda-12.6
+    tags:
+      - '*'
+  workflow_dispatch:
+
+concurrency:
+  group: build-wheels-cpu-arm64-ubuntu-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  generate_build_matrix:
+    # see https://github.com/pytorch/pytorch/pull/50633
+    runs-on: ubuntu-latest
+    outputs:
+      matrix: ${{ steps.set-matrix.outputs.matrix }}
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+      - name: Generating build matrix
+        id: set-matrix
+        run: |
+          # outputting for debugging purposes
+          python ./scripts/github_actions/generate_build_matrix.py --for-arm64
+          MATRIX=$(python ./scripts/github_actions/generate_build_matrix.py --for-arm64)
+
+          # python ./scripts/github_actions/generate_build_matrix.py --test-only-latest-torch --for-arm64
+          # MATRIX=$(python ./scripts/github_actions/generate_build_matrix.py --test-only-latest-torch --for-arm64)
+          echo "::set-output name=matrix::${MATRIX}"
+
+  build-manylinux-wheels:
+    needs: generate_build_matrix
+    name: ${{ matrix.torch }} ${{ matrix.python-version }}
+    runs-on: ubuntu-22.04-arm
+    strategy:
+      fail-fast: false
+      matrix:
+        ${{ fromJson(needs.generate_build_matrix.outputs.matrix) }}
+
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      # 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 ${{ github.workspace }}:/var/www -e IS_2_28=${{ matrix.is_2_28 }} -e PYTHON_VERSION=${{ matrix.python-version }} -e TORCH_VERSION=${{ matrix.torch }}
+            run: |
+              echo "pwd: $PWD"
+              uname -a
+              id
+              cat /etc/*release
+              gcc --version
+              python3 --version
+              which python3
+
+              ls -lh /opt/python/
+
+              echo "---"
+
+              ls -lh /opt/python/cp*
+
+              ls -lh /opt/python/*/bin
+
+              echo "---"
+              find /opt/python/cp* -name "libpython*"
+              echo "-----"
+              find /opt/_internal/cp* -name "libpython*"
+
+              echo "-----"
+              find / -name "libpython*"
+              echo "----"
+              ls -lh /usr/lib64/libpython3.so
+
+              # cp36-cp36m
+              # cp37-cp37m
+              # cp38-cp38
+              # cp39-cp39
+              # cp310-cp310
+              # cp311-cp311
+              # cp312-cp312
+              # cp313-cp313
+              # cp313-cp313t  (no gil)
+              if [[ $PYTHON_VERSION == "3.6" ]]; then
+                python_dir=/opt/python/cp36-cp36m
+                export PYTHONPATH=/opt/python/cp36-cp36m/lib/python3.6/site-packages:$PYTHONPATH
+              elif [[ $PYTHON_VERSION == "3.7" ]]; then
+                python_dir=/opt/python/cp37-cp37m
+                export PYTHONPATH=/opt/python/cp37-cp37m/lib/python3.7/site-packages:$PYTHONPATH
+              elif [[ $PYTHON_VERSION == "3.8" ]]; then
+                python_dir=/opt/python/cp38-cp38
+                export PYTHONPATH=/opt/python/cp38-cp38/lib/python3.8/site-packages:$PYTHONPATH
+              elif [[ $PYTHON_VERSION == "3.9" ]]; then
+                python_dir=/opt/python/cp39-cp39
+                export PYTHONPATH=/opt/python/cp39-cp39/lib/python3.9/site-packages:$PYTHONPATH
+              elif [[ $PYTHON_VERSION == "3.10" ]]; then
+                python_dir=/opt/python/cp310-cp310
+                export PYTHONPATH=/opt/python/cp310-cp310/lib/python3.10/site-packages:$PYTHONPATH
+              elif [[ $PYTHON_VERSION == "3.11" ]]; then
+                python_dir=/opt/python/cp311-cp311
+                export PYTHONPATH=/opt/python/cp311-cp311/lib/python3.11/site-packages:$PYTHONPATH
+              elif [[ $PYTHON_VERSION == "3.12" ]]; then
+                python_dir=/opt/python/cp312-cp312
+                export PYTHONPATH=/opt/python/cp312-cp312/lib/python3.12/site-packages:$PYTHONPATH
+              elif [[ $PYTHON_VERSION == "3.13" ]]; then
+                python_dir=/opt/python/cp313-cp313
+                export PYTHONPATH=/opt/python/cp313-cp313/lib/python3.13/site-packages:$PYTHONPATH
+              else
+                echo "Unsupported Python version $PYTHON_VERSION"
+                exit 1
+              fi
+
+              export PYTHON_INSTALL_DIR=$python_dir
+              export PATH=$PYTHON_INSTALL_DIR/bin:$PATH
+
+              python3 --version
+              which python3
+
+              /var/www/scripts/github_actions/build-ubuntu-cpu-arm64.sh
+
+      - name: Display wheels
+        shell: bash
+        run: |
+          ls -lh ./wheelhouse/
+
+      # https://huggingface.co/docs/hub/spaces-github-actions
+      - name: Publish to huggingface
+        if: github.repository_owner == 'csukuangfj'
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+        uses: nick-fields/retry@v2
+        with:
+          max_attempts: 20
+          timeout_seconds: 200
+          shell: bash
+          command: |
+            git config --global user.email "csukuangfj@gmail.com"
+            git config --global user.name "Fangjun Kuang"
+
+            rm -rf huggingface
+
+            export GIT_LFS_SKIP_SMUDGE=1
+
+            git clone https://huggingface.co/csukuangfj/kaldifeat huggingface
+            cd huggingface
+            git pull
+
+            d=cpu/1.25.5.dev20250307/linux-arm64
+            mkdir -p $d
+            cp -v ../wheelhouse/*.whl ./$d
+            git status
+            git lfs track "*.whl"
+            git add .
+            git commit -m "upload ubuntu-arm64-cpu wheel for torch ${{ matrix.torch }} python ${{ matrix.python-version }}"
+            git push https://csukuangfj:$HF_TOKEN@huggingface.co/csukuangfj/kaldifeat main
diff --git a/.github/workflows/ubuntu-cpu-wheels.yml b/.github/workflows/ubuntu-cpu-wheels.yml
index b11c342..5fb9201 100644
--- a/.github/workflows/ubuntu-cpu-wheels.yml
+++ b/.github/workflows/ubuntu-cpu-wheels.yml
@@ -3,7 +3,8 @@ name: build-wheels-cpu-ubuntu
 on:
   push:
     branches:
-      - fix-wheel-2
+      # - fix-cuda-12.6
+      - wheel
     tags:
       - '*'
   workflow_dispatch:
@@ -56,7 +57,7 @@ jobs:
         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 }}
+            options: -v ${{ github.workspace }}:/var/www -e IS_2_28=${{ matrix.is_2_28 }} -e PYTHON_VERSION=${{ matrix.python-version }} -e TORCH_VERSION=${{ matrix.torch }}
             run: |
               echo "pwd: $PWD"
               uname -a
@@ -65,6 +66,69 @@ jobs:
               gcc --version
               python3 --version
               which python3
+
+              ls -lh /opt/python/
+
+              echo "---"
+
+              ls -lh /opt/python/cp*
+
+              ls -lh /opt/python/*/bin
+
+              echo "---"
+              find /opt/python/cp* -name "libpython*"
+              echo "-----"
+              find /opt/_internal/cp* -name "libpython*"
+
+              echo "-----"
+              find / -name "libpython*"
+              echo "----"
+              ls -lh /usr/lib64/libpython3.so || true
+
+              # cp36-cp36m
+              # cp37-cp37m
+              # cp38-cp38
+              # cp39-cp39
+              # cp310-cp310
+              # cp311-cp311
+              # cp312-cp312
+              # cp313-cp313
+              # cp313-cp313t  (no gil)
+              if [[ $PYTHON_VERSION == "3.6" ]]; then
+                python_dir=/opt/python/cp36-cp36m
+                export PYTHONPATH=/opt/python/cp36-cp36m/lib/python3.6/site-packages:$PYTHONPATH
+              elif [[ $PYTHON_VERSION == "3.7" ]]; then
+                python_dir=/opt/python/cp37-cp37m
+                export PYTHONPATH=/opt/python/cp37-cp37m/lib/python3.7/site-packages:$PYTHONPATH
+              elif [[ $PYTHON_VERSION == "3.8" ]]; then
+                python_dir=/opt/python/cp38-cp38
+                export PYTHONPATH=/opt/python/cp38-cp38/lib/python3.8/site-packages:$PYTHONPATH
+              elif [[ $PYTHON_VERSION == "3.9" ]]; then
+                python_dir=/opt/python/cp39-cp39
+                export PYTHONPATH=/opt/python/cp39-cp39/lib/python3.9/site-packages:$PYTHONPATH
+              elif [[ $PYTHON_VERSION == "3.10" ]]; then
+                python_dir=/opt/python/cp310-cp310
+                export PYTHONPATH=/opt/python/cp310-cp310/lib/python3.10/site-packages:$PYTHONPATH
+              elif [[ $PYTHON_VERSION == "3.11" ]]; then
+                python_dir=/opt/python/cp311-cp311
+                export PYTHONPATH=/opt/python/cp311-cp311/lib/python3.11/site-packages:$PYTHONPATH
+              elif [[ $PYTHON_VERSION == "3.12" ]]; then
+                python_dir=/opt/python/cp312-cp312
+                export PYTHONPATH=/opt/python/cp312-cp312/lib/python3.12/site-packages:$PYTHONPATH
+              elif [[ $PYTHON_VERSION == "3.13" ]]; then
+                python_dir=/opt/python/cp313-cp313
+                export PYTHONPATH=/opt/python/cp313-cp313/lib/python3.13/site-packages:$PYTHONPATH
+              else
+                echo "Unsupported Python version $PYTHON_VERSION"
+                exit 1
+              fi
+
+              export PYTHON_INSTALL_DIR=$python_dir
+              export PATH=$PYTHON_INSTALL_DIR/bin:$PATH
+
+              python3 --version
+              which python3
+
               /var/www/scripts/github_actions/build-ubuntu-cpu.sh
 
       - name: Display wheels
@@ -72,12 +136,6 @@ jobs:
         run: |
           ls -lh ./wheelhouse/
 
-      - name: Upload Wheel
-        uses: actions/upload-artifact@v3
-        with:
-          name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cpu
-          path: wheelhouse/*.whl
-
       # https://huggingface.co/docs/hub/spaces-github-actions
       - name: Publish to huggingface
         if: github.repository_owner == 'csukuangfj'
@@ -100,7 +158,7 @@ jobs:
             cd huggingface
             git pull
 
-            d=cpu/1.25.5.dev20241029/linux
+            d=cpu/1.25.5.dev20250307/linux-x64
             mkdir -p $d
             cp -v ../wheelhouse/*.whl ./$d
             git status
diff --git a/.github/workflows/ubuntu-cuda-wheels.yml b/.github/workflows/ubuntu-cuda-wheels.yml
index 710e37e..a7f750c 100644
--- a/.github/workflows/ubuntu-cuda-wheels.yml
+++ b/.github/workflows/ubuntu-cuda-wheels.yml
@@ -3,7 +3,7 @@ name: build-wheels-cuda-ubuntu
 on:
   push:
     branches:
-      - fix-wheel-2
+      - wheel
     tags:
       - '*'
   workflow_dispatch:
@@ -26,11 +26,11 @@ jobs:
         id: set-matrix
         run: |
           # outputting for debugging purposes
-          # python ./scripts/github_actions/generate_build_matrix.py --enable-cuda
-          # MATRIX=$(python ./scripts/github_actions/generate_build_matrix.py --enable-cuda)
+          python ./scripts/github_actions/generate_build_matrix.py --enable-cuda
+          MATRIX=$(python ./scripts/github_actions/generate_build_matrix.py --enable-cuda)
 
-          python ./scripts/github_actions/generate_build_matrix.py --enable-cuda --test-only-latest-torch
-          MATRIX=$(python ./scripts/github_actions/generate_build_matrix.py --enable-cuda --test-only-latest-torch)
+          # python ./scripts/github_actions/generate_build_matrix.py --enable-cuda --test-only-latest-torch
+          # MATRIX=$(python ./scripts/github_actions/generate_build_matrix.py --enable-cuda --test-only-latest-torch)
           echo "::set-output name=matrix::${MATRIX}"
 
   build-manylinux-wheels:
@@ -65,7 +65,7 @@ jobs:
         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 }}
+            options: -v ${{ github.workspace }}:/var/www -e IS_2_28=${{ matrix.is_2_28 }} -e PYTHON_VERSION=${{ matrix.python-version }} -e TORCH_VERSION=${{ matrix.torch }} -e CUDA_VERSION=${{ matrix.cuda }}
             run: |
               echo "pwd: $PWD"
               uname -a
@@ -75,6 +75,69 @@ jobs:
               python3 --version
               which python3
 
+              ls -lh /opt/python/
+
+              echo "---"
+
+              ls -lh /opt/python/cp*
+
+              ls -lh /opt/python/*/bin
+
+              echo "---"
+              find /opt/python/cp* -name "libpython*"
+              echo "-----"
+              find /opt/_internal/cp* -name "libpython*"
+
+              echo "-----"
+              find / -name "libpython*"
+
+              # cp36-cp36m
+              # cp37-cp37m
+              # cp38-cp38
+              # cp39-cp39
+              # cp310-cp310
+              # cp311-cp311
+              # cp312-cp312
+              # cp313-cp313
+              # cp313-cp313t  (no gil)
+              if [[ $PYTHON_VERSION == "3.6" ]]; then
+                python_dir=/opt/python/cp36-cp36m
+                export PYTHONPATH=/opt/python/cp36-cp36m/lib/python3.6/site-packages:$PYTHONPATH
+              elif [[ $PYTHON_VERSION == "3.7" ]]; then
+                python_dir=/opt/python/cp37-cp37m
+                export PYTHONPATH=/opt/python/cp37-cp37m/lib/python3.7/site-packages:$PYTHONPATH
+              elif [[ $PYTHON_VERSION == "3.8" ]]; then
+                python_dir=/opt/python/cp38-cp38
+                export PYTHONPATH=/opt/python/cp38-cp38/lib/python3.8/site-packages:$PYTHONPATH
+              elif [[ $PYTHON_VERSION == "3.9" ]]; then
+                python_dir=/opt/python/cp39-cp39
+                export PYTHONPATH=/opt/python/cp39-cp39/lib/python3.9/site-packages:$PYTHONPATH
+              elif [[ $PYTHON_VERSION == "3.10" ]]; then
+                python_dir=/opt/python/cp310-cp310
+                export PYTHONPATH=/opt/python/cp310-cp310/lib/python3.10/site-packages:$PYTHONPATH
+              elif [[ $PYTHON_VERSION == "3.11" ]]; then
+                python_dir=/opt/python/cp311-cp311
+                export PYTHONPATH=/opt/python/cp311-cp311/lib/python3.11/site-packages:$PYTHONPATH
+              elif [[ $PYTHON_VERSION == "3.12" ]]; then
+                python_dir=/opt/python/cp312-cp312
+                export PYTHONPATH=/opt/python/cp312-cp312/lib/python3.12/site-packages:$PYTHONPATH
+              elif [[ $PYTHON_VERSION == "3.13" ]]; then
+                python_dir=/opt/python/cp313-cp313
+                export PYTHONPATH=/opt/python/cp313-cp313/lib/python3.13/site-packages:$PYTHONPATH
+              else
+                echo "Unsupported Python version $PYTHON_VERSION"
+                exit 1
+              fi
+
+              export PYTHON_INSTALL_DIR=$python_dir
+              export PATH=$PYTHON_INSTALL_DIR/bin:$PATH
+
+              # There are no libpython.so inside $PYTHON_INSTALL_DIR
+              # since they are statically linked.
+
+              python3 --version
+              which python3
+
               pushd /usr/local
               rm cuda
               ln -s cuda-$CUDA_VERSION cuda
@@ -93,9 +156,10 @@ jobs:
           ls -lh ./wheelhouse/
 
       - name: Upload Wheel
-        uses: actions/upload-artifact@v3
+        if: false
+        uses: actions/upload-artifact@v4
         with:
-          name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cuda
+          name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-latest-cuda-is_2_28-${{ matrix.is_2_28 }}
           path: wheelhouse/*.whl
 
       # https://huggingface.co/docs/hub/spaces-github-actions
diff --git a/.github/workflows/windows-x64-cpu-wheels.yml b/.github/workflows/windows-x64-cpu-wheels.yml
index d2256c6..e9b12e1 100644
--- a/.github/workflows/windows-x64-cpu-wheels.yml
+++ b/.github/workflows/windows-x64-cpu-wheels.yml
@@ -63,7 +63,7 @@ jobs:
           ls -lh ./wheelhouse/
 
       - name: Upload Wheel
-        uses: actions/upload-artifact@v3
+        uses: actions/upload-artifact@v4
         with:
           name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-windows-latest-cpu
           path: wheelhouse/*.whl
diff --git a/CMakeLists.txt b/CMakeLists.txt
index 11f839a..d9acc3f 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -43,7 +43,7 @@ list(APPEND CMAKE_MODULE_PATH ${CMAKE_CURRENT_SOURCE_DIR}/cmake/Modules)
 list(APPEND CMAKE_MODULE_PATH ${CMAKE_CURRENT_SOURCE_DIR}/cmake)
 
 option(BUILD_SHARED_LIBS "Whether to build shared libraries" ON)
-option(kaldifeat_BUILD_TESTS "Whether to build tests or not" ON)
+option(kaldifeat_BUILD_TESTS "Whether to build tests or not" OFF)
 option(kaldifeat_BUILD_PYMODULE "Whether to build python module or not" ON)
 
 message(STATUS "BUILD_SHARED_LIBS ${BUILD_SHARED_LIBS}")
diff --git a/kaldifeat/python/csrc/CMakeLists.txt b/kaldifeat/python/csrc/CMakeLists.txt
index cf3bbae..10eea97 100644
--- a/kaldifeat/python/csrc/CMakeLists.txt
+++ b/kaldifeat/python/csrc/CMakeLists.txt
@@ -29,10 +29,10 @@ endif()
 target_link_libraries(_kaldifeat PRIVATE kaldifeat_core)
 if(UNIX AND NOT APPLE)
   target_link_libraries(_kaldifeat PUBLIC ${TORCH_DIR}/lib/libtorch_python.so)
-  target_link_libraries(_kaldifeat PUBLIC ${PYTHON_LIBRARY})
+  # target_link_libraries(_kaldifeat PUBLIC ${PYTHON_LIBRARY})
 elseif(WIN32)
   target_link_libraries(_kaldifeat PUBLIC ${TORCH_DIR}/lib/torch_python.lib)
-  target_link_libraries(_kaldifeat PUBLIC ${PYTHON_LIBRARIES})
+  # target_link_libraries(_kaldifeat PUBLIC ${PYTHON_LIBRARIES})
 endif()
 
 install(TARGETS _kaldifeat
diff --git a/scripts/github_actions/build-ubuntu-cpu-arm64.sh b/scripts/github_actions/build-ubuntu-cpu-arm64.sh
new file mode 100755
index 0000000..6ba3e01
--- /dev/null
+++ b/scripts/github_actions/build-ubuntu-cpu-arm64.sh
@@ -0,0 +1,78 @@
+#!/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: 3.8, 3.9, 3.10, 3.11
+  exit 1
+fi
+
+if [ -z $TORCH_VERSION ]; then
+  echo "Please set the environment variable TORCH_VERSION"
+  echo "Example: export TORCH_VERSION=1.10.0"
+  exit 1
+fi
+
+export PATH=$PYTHON_INSTALL_DIR/bin:$PATH
+export LD_LIBRARY_PATH=$PYTHON_INSTALL_DIR/lib:$LD_LIBRARY_PATH
+ls -lh $PYTHON_INSTALL_DIR/lib/
+
+python3 --version
+which python3
+
+python3 -m pip install scikit-build
+python3 -m pip install -U pip cmake
+python3 -m pip install wheel twine typing_extensions
+python3 -m pip install bs4 requests tqdm auditwheel
+
+echo "Installing torch"
+python3 -m pip install -qq torch==$TORCH_VERSION || \
+python3 -m pip install -qq torch==$TORCH_VERSION -f https://download.pytorch.org/whl/torch_stable.html || \
+python3 -m pip install -qq torch==$TORCH_VERSION -f https://download.pytorch.org/whl/torch/
+
+rm -rf ~/.cache/pip
+yum clean all
+
+cd /var/www
+
+export CMAKE_CUDA_COMPILER_LAUNCHER=
+export KALDIFEAT_CMAKE_ARGS=" -DPYTHON_EXECUTABLE=$PYTHON_INSTALL_DIR/bin/python3 "
+export KALDIFEAT_MAKE_ARGS=" -j "
+
+nvcc --version || true
+rm -rf /usr/local/cuda*
+nvcc --version || true
+
+if [[ x"$IS_2_28" == x"1" ]]; then
+  plat=manylinux_2_28_aarch64
+else
+  plat=manylinux_2_17_aarch64
+fi
+
+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 $plat \
+  dist/*.whl
+
+ls -lh  /var/www
diff --git a/scripts/github_actions/build-ubuntu-cpu.sh b/scripts/github_actions/build-ubuntu-cpu.sh
index a113d28..d90a13e 100755
--- a/scripts/github_actions/build-ubuntu-cpu.sh
+++ b/scripts/github_actions/build-ubuntu-cpu.sh
@@ -15,46 +15,13 @@ if [ -z $TORCH_VERSION ]; then
   exit 1
 fi
 
-yum -y install openssl-devel bzip2-devel libffi-devel xz-devel wget redhat-lsb-core
-
-INSTALLED_PYTHON_VERSION=${PYTHON_VERSION}.2
-if [[ $PYTHON_VERSION == "3.13" ]]; then
-  INSTALLED_PYTHON_VERSION=${PYTHON_VERSION}.0
-fi
-echo "Installing $INSTALLED_PYTHON_VERSION"
-
-curl -O https://www.python.org/ftp/python/$INSTALLED_PYTHON_VERSION/Python-$INSTALLED_PYTHON_VERSION.tgz
-tar xf Python-$INSTALLED_PYTHON_VERSION.tgz
-pushd Python-$INSTALLED_PYTHON_VERSION
-
-PYTHON_INSTALL_DIR=$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 >/dev/null 2>&1
-make install >/dev/null 2>&1
-
-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/
 
-nvcc --version || true
-rm -rf /usr/local/cuda*
-nvcc --version || true
-
 python3 --version
 which python3
 
-if [[ $PYTHON_VERSION != 3.6 ]]; then
-  curl -O https://bootstrap.pypa.io/get-pip.py
-  python3 get-pip.py
-fi
-
 python3 -m pip install scikit-build
 python3 -m pip install -U pip cmake
 python3 -m pip install wheel twine typing_extensions
@@ -73,8 +40,16 @@ export CMAKE_CUDA_COMPILER_LAUNCHER=
 export KALDIFEAT_CMAKE_ARGS=" -DPYTHON_EXECUTABLE=$PYTHON_INSTALL_DIR/bin/python3 "
 export KALDIFEAT_MAKE_ARGS=" -j "
 
+nvcc --version || true
+rm -rf /usr/local/cuda*
+nvcc --version || true
 
 python3 setup.py bdist_wheel
+if [[ x"$IS_2_28" == x"1" ]]; then
+  plat=manylinux_2_28_x86_64
+else
+  plat=manylinux_2_17_x86_64
+fi
 
 auditwheel --verbose repair \
   --exclude libc10.so \
@@ -95,8 +70,7 @@ auditwheel --verbose repair \
   --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 \
+  --plat $plat \
   dist/*.whl
 
 ls -lh  /var/www
diff --git a/scripts/github_actions/build-ubuntu-cuda.sh b/scripts/github_actions/build-ubuntu-cuda.sh
index 5d0ae06..dcabb24 100755
--- a/scripts/github_actions/build-ubuntu-cuda.sh
+++ b/scripts/github_actions/build-ubuntu-cuda.sh
@@ -27,49 +27,7 @@ if [[ $TORCH_VERSION =~ 2.2.* && $CUDA_VERSION =~ 12.* ]]; then
   export TORCH_CUDA_ARCH_LIST="8.0 8.6 8.9 9.0"
 fi
 
-
-yum -y install openssl-devel bzip2-devel libffi-devel xz-devel wget redhat-lsb-core
-
-
-INSTALLED_PYTHON_VERSION=${PYTHON_VERSION}.2
-if [[ $PYTHON_VERSION == "3.13" ]]; then
-  INSTALLED_PYTHON_VERSION=${PYTHON_VERSION}.0
-fi
-echo "Installing $INSTALLED_PYTHON_VERSION"
-
-curl -O https://www.python.org/ftp/python/$INSTALLED_PYTHON_VERSION/Python-$INSTALLED_PYTHON_VERSION.tgz
-tar xf Python-$INSTALLED_PYTHON_VERSION.tgz
-pushd Python-$INSTALLED_PYTHON_VERSION
-
-PYTHON_INSTALL_DIR=$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 >/dev/null 2>&1
-make install >/dev/null 2>&1
-
-popd
-
-echo "pwd: $PWD"
-# rm -rf Python-${PYTHON_VERSION}.2
-
-export PATH=$PYTHON_INSTALL_DIR/bin:$PATH
-export LD_LIBRARY_PATH=$PYTHON_INSTALL_DIR/lib:$LD_LIBRARY_PATH
-ls -lh $PYTHON_INSTALL_DIR
-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 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
@@ -77,6 +35,9 @@ python3 -m pip install bs4 requests tqdm auditwheel
 echo "Installing torch"
 ./install_torch.sh
 
+python3 -c "import torch; print(torch.__file__)"
+
+
 # -- Autodetected CUDA architecture(s): 5.0;8.0;8.6;8.9;9.0;9.0a
 # CMake Error at /Python-3.8.2/py-3.8/lib/python3.8/site-packages/torch/share/cmake/Caffe2/Modules_CUDA_fix/upstream/FindCUDA/select_compute_arch.cmake:227 (message):
 #   Unknown CUDA Architecture Name 9.0a in CUDA_SELECT_NVCC_ARCH_FLAGS
@@ -89,6 +50,41 @@ echo "Installing torch"
 #   CMakeLists.txt:62 (include)
 sed -i.bak /9.0a/d /Python-*/py-3.*/lib/python3.*/site-packages/torch/share/cmake/Caffe2/Modules_CUDA_fix/upstream/FindCUDA/select_compute_arch.cmake || true
 
+if [[ x"$IS_2_28" != x"1" ]]; then
+  yum -y install openssl-devel
+fi
+
+yum -y install zlib-devel bzip2-devel libffi-devel xz-devel wget redhat-lsb-core
+
+INSTALLED_PYTHON_VERSION=${PYTHON_VERSION}.2
+if [[ $PYTHON_VERSION == "3.13" ]]; then
+  INSTALLED_PYTHON_VERSION=${PYTHON_VERSION}.0
+fi
+echo "Installing $INSTALLED_PYTHON_VERSION"
+
+curl -O https://www.python.org/ftp/python/$INSTALLED_PYTHON_VERSION/Python-$INSTALLED_PYTHON_VERSION.tgz
+tar xf Python-$INSTALLED_PYTHON_VERSION.tgz
+pushd Python-$INSTALLED_PYTHON_VERSION
+
+PYTHON_INSTALL_DIR=$PWD/py-${PYTHON_VERSION}
+
+if [[ $PYTHON_VERSION =~ 3.1. && x"$IS_2_28" != x"1" ]]; then
+  yum install -y openssl11-devel
+  sed -i 's/PKG_CONFIG openssl /PKG_CONFIG openssl11 /g' configure
+fi
+
+./configure --enable-shared --prefix=$PYTHON_INSTALL_DIR >/dev/null
+make install >/dev/null
+
+popd
+
+export PATH=$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
+
 rm -rf ~/.cache/pip >/dev/null 2>&1
 yum clean all >/dev/null 2>&1
 
@@ -98,9 +94,20 @@ export CMAKE_CUDA_COMPILER_LAUNCHER=
 export KALDIFEAT_CMAKE_ARGS=" -DPYTHON_EXECUTABLE=$PYTHON_INSTALL_DIR/bin/python3 "
 export KALDIFEAT_MAKE_ARGS=" -j2 "
 
+echo "KALDIFEAT_CMAKE_ARGS: $KALDIFEAT_CMAKE_ARGS"
+
 
 python3 setup.py bdist_wheel
 
+if [[ x"$IS_2_28" == x"1" ]]; then
+  plat=manylinux_2_28_x86_64
+else
+  plat=manylinux_2_17_x86_64
+fi
+export PATH=$PYTHON_INSTALL_DIR/bin:$PATH
+python3 --version
+which python3
+
 auditwheel --verbose repair \
   --exclude libc10.so \
   --exclude libc10_cuda.so \
@@ -140,7 +147,7 @@ auditwheel --verbose repair \
   --exclude libshm.so \
   --exclude libtorch_cuda_cpp.so \
   --exclude libtorch_cuda_cu.so \
-  --plat manylinux_2_17_x86_64 \
+  --plat $plat \
   -w /var/www/wheelhouse \
   dist/*.whl
 
diff --git a/scripts/github_actions/generate_build_matrix.py b/scripts/github_actions/generate_build_matrix.py
index 6bfcbe6..3892043 100755
--- a/scripts/github_actions/generate_build_matrix.py
+++ b/scripts/github_actions/generate_build_matrix.py
@@ -1,6 +1,15 @@
 #!/usr/bin/env python3
 # Copyright    2022  Xiaomi Corp.        (authors: Fangjun Kuang)
 
+"""
+See also
+https://github.com/pytorch/test-infra/blob/main/.github/workflows/test_build_wheels_linux_with_cuda.yml
+https://github.com/pytorch/test-infra/blob/main/.github/workflows/test_build_wheels_linux_without_cuda.yml
+
+https://github.com/pytorch/test-infra/actions/workflows/test_build_wheels_linux_with_cuda.yml
+https://github.com/pytorch/test-infra/blob/main/tools/scripts/generate_binary_build_matrix.py
+"""
+
 import argparse
 import json
 
@@ -40,6 +49,13 @@ def get_args():
         help="True for macOS",
     )
 
+    parser.add_argument(
+        "--for-arm64",
+        action="store_true",
+        default=False,
+        help="True for arm64",
+    )
+
     parser.add_argument(
         "--test-only-latest-torch",
         action="store_true",
@@ -50,7 +66,9 @@ def get_args():
     return parser.parse_args()
 
 
-def generate_build_matrix(enable_cuda, for_windows, for_macos, test_only_latest_torch):
+def generate_build_matrix(
+    enable_cuda, for_windows, for_macos, test_only_latest_torch, for_arm64
+):
     matrix = {
         # 1.5.x is removed because there are compilation errors.
         #  See
@@ -69,46 +87,46 @@ def generate_build_matrix(enable_cuda, for_windows, for_macos, test_only_latest_
         #      "python-version": ["3.6", "3.7", "3.8"],
         #      "cuda": ["10.1", "10.2"] if not for_windows else ["10.1.243", "10.2.89"],
         #  },
-        "1.7.0": {
-            "python-version": ["3.6", "3.7", "3.8"],
-            "cuda": (
-                ["10.1", "10.2", "11.0"]
-                if not for_windows
-                else ["10.1.243", "10.2.89", "11.0.3"]
-            ),
-        },
-        "1.7.1": {
-            "python-version": ["3.6", "3.7", "3.8", "3.9"],
-            "cuda": (
-                ["10.1", "10.2", "11.0"]
-                if not for_windows
-                else ["10.1.243", "10.2.89", "11.0.3"]
-            ),
-        },
-        "1.8.0": {
-            "python-version": ["3.6", "3.7", "3.8", "3.9"],
-            "cuda": (
-                ["10.1", "10.2", "11.1"]
-                if not for_windows
-                else ["10.1.243", "10.2.89", "11.1.1"]
-            ),
-        },
-        "1.8.1": {
-            "python-version": ["3.6", "3.7", "3.8", "3.9"],
-            "cuda": (
-                ["10.1", "10.2", "11.1"]
-                if not for_windows
-                else ["10.1.243", "10.2.89", "11.1.1"]
-            ),
-        },
-        "1.9.0": {
-            "python-version": ["3.6", "3.7", "3.8", "3.9"],
-            "cuda": ["10.2", "11.1"] if not for_windows else ["10.2.89", "11.1.1"],
-        },
-        "1.9.1": {
-            "python-version": ["3.6", "3.7", "3.8", "3.9"],
-            "cuda": ["10.2", "11.1"] if not for_windows else ["10.2.89", "11.1.1"],
-        },
+        #  "1.7.0": {
+        #      "python-version": ["3.6", "3.7", "3.8"],
+        #      "cuda": (
+        #          ["10.1", "10.2", "11.0"]
+        #          if not for_windows
+        #          else ["10.1.243", "10.2.89", "11.0.3"]
+        #      ),
+        #  },
+        #  "1.7.1": {
+        #      "python-version": ["3.6", "3.7", "3.8", "3.9"],
+        #      "cuda": (
+        #          ["10.1", "10.2", "11.0"]
+        #          if not for_windows
+        #          else ["10.1.243", "10.2.89", "11.0.3"]
+        #      ),
+        #  },
+        #  "1.8.0": {
+        #      "python-version": ["3.6", "3.7", "3.8", "3.9"],
+        #      "cuda": (
+        #          ["10.1", "10.2", "11.1"]
+        #          if not for_windows
+        #          else ["10.1.243", "10.2.89", "11.1.1"]
+        #      ),
+        #  },
+        #  "1.8.1": {
+        #      "python-version": ["3.6", "3.7", "3.8", "3.9"],
+        #      "cuda": (
+        #          ["10.1", "10.2", "11.1"]
+        #          if not for_windows
+        #          else ["10.1.243", "10.2.89", "11.1.1"]
+        #      ),
+        #  },
+        #  "1.9.0": {
+        #      "python-version": ["3.6", "3.7", "3.8", "3.9"],
+        #      "cuda": ["10.2", "11.1"] if not for_windows else ["10.2.89", "11.1.1"],
+        #  },
+        #  "1.9.1": {
+        #      "python-version": ["3.6", "3.7", "3.8", "3.9"],
+        #      "cuda": ["10.2", "11.1"] if not for_windows else ["10.2.89", "11.1.1"],
+        #  },
         "1.10.0": {
             "python-version": ["3.6", "3.7", "3.8", "3.9"],
             "cuda": (
@@ -305,9 +323,18 @@ def generate_build_matrix(enable_cuda, for_windows, for_macos, test_only_latest_
             matrix["1.13.1"]["python-version"].remove("3.11")
 
     excluded_python_versions = ["3.6"]
+    enabled_torch_versions = ["1.10.0"]
+
+    enabled_torch_versions += ["1.13.0", "1.13.1"]
+
+    min_torch_version = "2.0.0"
 
     ans = []
     for torch, python_cuda in matrix.items():
+        if enabled_torch_versions and torch not in enabled_torch_versions:
+            if not version_ge(torch, min_torch_version):
+                continue
+
         python_versions = python_cuda["python-version"]
         cuda_versions = python_cuda["cuda"]
         if enable_cuda:
@@ -318,12 +345,30 @@ def generate_build_matrix(enable_cuda, for_windows, for_macos, test_only_latest_
                     if c in ["10.1", "11.0"]:
                         # no docker image for cuda 10.1 and 11.0
                         continue
+
+                    if version_ge(torch, "2.7.0") or (
+                        version_ge(torch, "2.6.0") and c == "12.6"
+                    ):
+                        # case 1: torch >= 2.7
+                        # case 2: torch == 2.6.0 && cuda == 12.6
+                        ans.append(
+                            {
+                                "torch": torch,
+                                "python-version": p,
+                                "cuda": c,
+                                "image": f"pytorch/manylinux2_28-builder:cuda{c}",
+                                "is_2_28": "1",
+                            }
+                        )
+                        continue
+
                     ans.append(
                         {
                             "torch": torch,
                             "python-version": p,
                             "cuda": c,
                             "image": f"pytorch/manylinux-builder:cuda{c}",
+                            "is_2_28": "0",
                         }
                     )
         else:
@@ -336,13 +381,27 @@ def generate_build_matrix(enable_cuda, for_windows, for_macos, test_only_latest_
                     ans.append({"torch": torch, "python-version": p})
                 elif for_macos:
                     ans.append({"torch": torch, "python-version": p})
+                elif version_ge(torch, "2.6.0"):
+                    ans.append(
+                        {
+                            "torch": torch,
+                            "python-version": p,
+                            "image": "pytorch/manylinux2_28-builder:cpu"
+                            if not for_arm64
+                            else "pytorch/manylinux2_28_aarch64-builder:cpu-aarch64",
+                            "is_2_28": "1",
+                        }
+                    )
                 elif version_ge(torch, "2.4.0"):
                     ans.append(
                         {
                             "torch": torch,
                             "python-version": p,
                             #  "image": "pytorch/manylinux-builder:cpu-2.4",
-                            "image": "pytorch/manylinux-builder:cpu-27677ead7c8293c299a885ae2c474bf445e653a5",
+                            "image": "pytorch/manylinux-builder:cpu-27677ead7c8293c299a885ae2c474bf445e653a5"
+                            if not for_arm64
+                            else "pytorch/manylinuxaarch64-builder:cpu-aarch64-195148266541a9789074265141cb7dc19dc14c54",
+                            "is_2_28": "0",
                         }
                     )
                 elif version_ge(torch, "2.2.0"):
@@ -350,7 +409,10 @@ def generate_build_matrix(enable_cuda, for_windows, for_macos, test_only_latest_
                         {
                             "torch": torch,
                             "python-version": p,
-                            "image": "pytorch/manylinux-builder:cpu-2.2",
+                            "image": "pytorch/manylinux-builder:cpu-2.2"
+                            if not for_arm64
+                            else "pytorch/manylinuxaarch64-builder:cpu-aarch64-195148266541a9789074265141cb7dc19dc14c54",
+                            "is_2_28": "0",
                         }
                     )
                 else:
@@ -358,7 +420,10 @@ def generate_build_matrix(enable_cuda, for_windows, for_macos, test_only_latest_
                         {
                             "torch": torch,
                             "python-version": p,
-                            "image": f"pytorch/manylinux-builder:cuda10.2",
+                            "image": "pytorch/manylinux-builder:cuda10.2"
+                            if not for_arm64
+                            else "pytorch/manylinuxaarch64-builder:cpu-aarch64-195148266541a9789074265141cb7dc19dc14c54",
+                            "is_2_28": "0",
                         }
                     )
 
@@ -371,6 +436,7 @@ def main():
         enable_cuda=args.enable_cuda,
         for_windows=args.for_windows,
         for_macos=args.for_macos,
+        for_arm64=args.for_arm64,
         test_only_latest_torch=args.test_only_latest_torch,
     )
 
diff --git a/scripts/github_actions/install_torch.sh b/scripts/github_actions/install_torch.sh
index 80de75a..b0cea34 100755
--- a/scripts/github_actions/install_torch.sh
+++ b/scripts/github_actions/install_torch.sh
@@ -280,4 +280,6 @@ else
   retry python3 -m pip install -q $package -f $url
 fi
 
-rm -rfv ~/.cache/pip
+python3 -m torch.utils.collect_env
+
+rm -rf ~/.cache/pip