Support torch 2.8.0 (#120)

See also https://github.com/pytorch/pytorch/issues/120020

.github/workflows/build-doc.yml

@@ -54,6 +54,7 @@ jobs:
             ls -lh $KALDIFEAT_DIR
 
             export GIT_LFS_SKIP_SMUDGE=1
+            export GIT_CLONE_PROTECTION_ACTIVE=false
             git clone https://huggingface.co/csukuangfj/kaldifeat huggingface
             cd huggingface
 
@@ -68,6 +69,8 @@ jobs:
           make html
           cp source/cpu.html build/html/
           cp source/cuda.html build/html/
+          cp source/cpu-cn.html build/html/
+          cp source/cuda-cn.html build/html/
           touch build/html/.nojekyll
 
       - name: Deploy

.github/workflows/macos-cpu-wheels.yml

@@ -3,7 +3,8 @@ name: build-wheels-cpu-macos
 on:
   push:
     branches:
-      - fix-wheel
+      # - wheel
+      - torch-2.8.0
     tags:
       - '*'
   workflow_dispatch:
@@ -26,14 +27,17 @@ jobs:
         id: set-matrix
         run: |
           # outputting for debugging purposes
-          python ./scripts/github_actions/generate_build_matrix.py --for-macos
+          # python ./scripts/github_actions/generate_build_matrix.py --for-macos
-          MATRIX=$(python ./scripts/github_actions/generate_build_matrix.py --for-macos)
+          # MATRIX=$(python ./scripts/github_actions/generate_build_matrix.py --for-macos)
+
+          python ./scripts/github_actions/generate_build_matrix.py --for-macos --test-only-latest-torch
+          MATRIX=$(python ./scripts/github_actions/generate_build_matrix.py --for-macos --test-only-latest-torch)
           echo "::set-output name=matrix::${MATRIX}"
 
   build_wheels_macos_cpu:
     needs: generate_build_matrix
     name: ${{ matrix.torch }} ${{ matrix.python-version }}
-    runs-on: macos-latest
+    runs-on: macos-14
     strategy:
       fail-fast: false
       matrix:
@@ -44,23 +48,44 @@ jobs:
         with:
           fetch-depth: 0
 
-      # see https://cibuildwheel.readthedocs.io/en/stable/changelog/
+      - name: Setup Python
-      # for a list of versions
+        uses: actions/setup-python@v5
-      - name: Build wheels
+        with:
-        uses: pypa/cibuildwheel@v2.11.4
+          python-version: ${{ matrix.python-version }}
-        env:
-          CIBW_BEFORE_BUILD: pip install torch==${{ matrix.torch}} cmake numpy
-          CIBW_BUILD: ${{ matrix.python-version }}-*
-          CIBW_REPAIR_WHEEL_COMMAND_MACOS: ""
-          CIBW_BUILD_VERBOSITY: 3
 
-      - name: Display wheels
+      - name: Install dependencies
+        shell: bash
+        run: |
+          pip install -q torch==${{ matrix.torch}} cmake numpy wheel>=0.40.0 twine setuptools
+
+      - name: Build wheel
+        shell: bash
+        run: |
+          python3 setup.py bdist_wheel
+          mkdir wheelhouse
+          cp -v dist/* wheelhouse
+
+      - name: Display wheels (before fix)
+        shell: bash
+        run: |
+          ls -lh ./wheelhouse/
+
+      - name: Fix wheel platform tag
+        run: |
+          # See https://github.com/glencoesoftware/zeroc-ice-py-macos-x86_64/pull/3/files
+          # See:
+          #  * https://github.com/pypa/wheel/issues/406
+          python -m wheel tags \
+            --platform-tag=macosx_11_0_arm64 \
+            --remove wheelhouse/*.whl
+
+      - name: Display wheels (after fix)
         shell: bash
         run: |
           ls -lh ./wheelhouse/
 
       - name: Upload Wheel
-        uses: actions/upload-artifact@v2
+        uses: actions/upload-artifact@v4
         with:
           name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-macos-latest-cpu
           path: wheelhouse/*.whl
@@ -86,8 +111,9 @@ jobs:
             cd huggingface
             git pull
 
-            mkdir -p macos
+            d=cpu/1.25.5.dev20241029/macos
-            cp -v ../wheelhouse/*.whl ./macos
+            mkdir -p $d
+            cp -v ../wheelhouse/*.whl ./$d
             git status
             git lfs track "*.whl"
             git add .

.github/workflows/run-tests-macos-cpu.yml

@@ -66,7 +66,7 @@ jobs:
         run: |
           python3 -m pip install -qq --upgrade pip
           python3 -m pip install -qq wheel twine typing_extensions soundfile numpy
-          python3 -m pip install -qq torch==${{ matrix.torch }} -f https://download.pytorch.org/whl/torch_stable.html
+          python3 -m pip install -qq torch==${{ matrix.torch }} -f https://download.pytorch.org/whl/torch_stable.html || python3 -m pip install -qq torch==${{ matrix.torch }} -f https://download.pytorch.org/whl/torch/
 
           python3 -c "import torch; print('torch version:', torch.__version__)"
 

.github/workflows/run-tests-ubuntu-cpu.yml

@@ -69,7 +69,7 @@ jobs:
           python3 -m pip install --upgrade pip
           python3 -m pip install wheel twine typing_extensions soundfile
           python3 -m pip install bs4 requests tqdm numpy
-          python3 -m pip install -qq torch==${{ matrix.torch }}+cpu -f https://download.pytorch.org/whl/torch_stable.html
+          python3 -m pip install -qq torch==${{ matrix.torch }}+cpu -f https://download.pytorch.org/whl/torch_stable.html || python3 -m pip install -qq torch==${{ matrix.torch }}+cpu -f https://download.pytorch.org/whl/torch/
 
           python3 -c "import torch; print('torch version:', torch.__version__)"
 

.github/workflows/run-tests-windows-cpu.yml

@@ -71,7 +71,7 @@ jobs:
 
       - name: Install PyTorch ${{ matrix.torch }}
         run: |
-          pip3 install -qq torch==${{ matrix.torch }}+cpu -f https://download.pytorch.org/whl/torch_stable.html
+          pip3 install -qq torch==${{ matrix.torch }}+cpu -f https://download.pytorch.org/whl/torch_stable.html || pip3 install -qq torch==${{ matrix.torch }}+cpu -f https://download.pytorch.org/whl/torch/
           pip3 install -qq wheel twine dataclasses numpy typing_extensions soundfile
 
       - name: Display CMake version
@@ -107,7 +107,7 @@ jobs:
           python3 -c "import kaldifeat; print(kaldifeat.__version__)"
 
       - name: Upload Wheel
-        uses: actions/upload-artifact@v2
+        uses: actions/upload-artifact@v4
         with:
           name: python-${{ matrix.python-version }}-${{ matrix.os }}-cpu
           path: dist/*.whl

.github/workflows/run-tests-windows-cuda.yml

@@ -113,7 +113,7 @@ jobs:
           python3 -m pip install -qq --upgrade pip
           python3 -m pip install -qq wheel twine numpy typing_extensions
           python3 -m pip install -qq dataclasses soundfile numpy
-          python3 -m pip install -qq torch==${{ matrix.torch }}${v} -f https://download.pytorch.org/whl/torch_stable.html numpy
+          python3 -m pip install -qq torch==${{ matrix.torch }}${v} -f https://download.pytorch.org/whl/torch_stable.html numpy || python3 -m pip install -qq torch==${{ matrix.torch }}${v} -f https://download.pytorch.org/whl/torch/ numpy
 
           python3 -c "import torch; print('torch version:', torch.__version__)"
 
@@ -159,7 +159,7 @@ jobs:
           python3 -c "import kaldifeat; print(kaldifeat.__version__)"
 
       - name: Upload Wheel
-        uses: actions/upload-artifact@v2
+        uses: actions/upload-artifact@v4
         with:
           name: python-${{ matrix.python-version }}-${{ matrix.os }}-cuda-${{ matrix.cuda }}
           path: dist/*.whl

.github/workflows/test-wheels.yml

@@ -47,7 +47,7 @@ jobs:
         if: startsWith(matrix.os, 'ubuntu') || startsWith(matrix.os, 'windows')
         shell: bash
         run: |
-          pip install torch==${{ github.event.inputs.torch_version }}+cpu -f https://download.pytorch.org/whl/torch_stable.html
+          pip install torch==${{ github.event.inputs.torch_version }}+cpu -f https://download.pytorch.org/whl/torch_stable.html || pip install torch==${{ github.event.inputs.torch_version }}+cpu -f https://download.pytorch.org/whl/torch/
 
       - name: Install kaldifeat
         shell: bash

.github/workflows/ubuntu-arm64-cpu-wheels.yml

@@ -0,0 +1,168 @@
+name: build-wheels-cpu-arm64-ubuntu
+
+on:
+  push:
+    branches:
+      # - wheel
+      - torch-2.8.0
+    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

.github/workflows/ubuntu-cpu-wheels.yml

@@ -3,7 +3,8 @@ name: build-wheels-cpu-ubuntu
 on:
   push:
     branches:
-      - fix-wheel
+      # - wheel
+      - torch-2.8.0
     tags:
       - '*'
   workflow_dispatch:
@@ -26,8 +27,11 @@ jobs:
         id: set-matrix
         run: |
           # outputting for debugging purposes
-          python ./scripts/github_actions/generate_build_matrix.py
+          # python ./scripts/github_actions/generate_build_matrix.py
-          MATRIX=$(python ./scripts/github_actions/generate_build_matrix.py)
+          # MATRIX=$(python ./scripts/github_actions/generate_build_matrix.py)
+
+          python ./scripts/github_actions/generate_build_matrix.py --test-only-latest-torch
+          MATRIX=$(python ./scripts/github_actions/generate_build_matrix.py --test-only-latest-torch)
           echo "::set-output name=matrix::${MATRIX}"
 
   build-manylinux-wheels:
@@ -53,7 +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
@@ -62,6 +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
@@ -69,12 +136,6 @@ jobs:
         run: |
           ls -lh ./wheelhouse/
 
-      - name: Upload Wheel
-        uses: actions/upload-artifact@v2
-        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'
@@ -97,8 +158,9 @@ jobs:
             cd huggingface
             git pull
 
-            mkdir -p ubuntu-cpu
+            d=cpu/1.25.5.dev20250307/linux-x64
-            cp -v ../wheelhouse/*.whl ./ubuntu-cpu
+            mkdir -p $d
+            cp -v ../wheelhouse/*.whl ./$d
             git status
             git lfs track "*.whl"
             git add .

.github/workflows/ubuntu-cuda-wheels.yml

@@ -3,7 +3,8 @@ name: build-wheels-cuda-ubuntu
 on:
   push:
     branches:
-      - fix-wheel
+      - wheel
+      # - torch-2.7.1
     tags:
       - '*'
   workflow_dispatch:
@@ -26,8 +27,11 @@ jobs:
         id: set-matrix
         run: |
           # outputting for debugging purposes
-          python ./scripts/github_actions/generate_build_matrix.py --enable-cuda
+          # python ./scripts/github_actions/generate_build_matrix.py --enable-cuda
-          MATRIX=$(python ./scripts/github_actions/generate_build_matrix.py --enable-cuda)
+          # MATRIX=$(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=$(python ./scripts/github_actions/generate_build_matrix.py --enable-cuda --test-only-latest-torch)
           echo "::set-output name=matrix::${MATRIX}"
 
   build-manylinux-wheels:
@@ -44,6 +48,15 @@ jobs:
         with:
           fetch-depth: 0
 
+      - name: Free space
+        shell: bash
+        run: |
+          df -h
+          rm -rf /opt/hostedtoolcache
+          df -h
+          echo "pwd: $PWD"
+          echo "github.workspace ${{ github.workspace }}"
+
       # 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
@@ -53,7 +66,7 @@ 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
@@ -63,6 +76,69 @@ 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
@@ -81,9 +157,10 @@ 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
@@ -107,8 +184,9 @@ jobs:
             cd huggingface
             git pull
 
-            mkdir -p ubuntu-cuda
+            d=cuda/1.25.5.dev20241029/linux
-            cp -v ../wheelhouse/*.whl ./ubuntu-cuda
+            mkdir -p $d
+            cp -v ../wheelhouse/*.whl ./$d
             git status
             git lfs track "*.whl"
             git add .

.github/workflows/windows-x64-cpu-wheels.yml

@@ -3,7 +3,8 @@ name: build-wheels-cpu-win64
 on:
   push:
     branches:
-      - fix-wheel
+      # - wheel
+      - torch-2.8.0
     tags:
       - '*'
   workflow_dispatch:
@@ -26,8 +27,11 @@ jobs:
         id: set-matrix
         run: |
           # outputting for debugging purposes
-          python ./scripts/github_actions/generate_build_matrix.py --for-windows
+          # python ./scripts/github_actions/generate_build_matrix.py --for-windows
-          MATRIX=$(python ./scripts/github_actions/generate_build_matrix.py --for-windows)
+          # MATRIX=$(python ./scripts/github_actions/generate_build_matrix.py --for-windows)
+
+          python ./scripts/github_actions/generate_build_matrix.py --for-windows --test-only-latest-torch
+          MATRIX=$(python ./scripts/github_actions/generate_build_matrix.py --for-windows --test-only-latest-torch)
           echo "::set-output name=matrix::${MATRIX}"
 
   build_wheels_win64_cpu:
@@ -44,15 +48,23 @@ jobs:
         with:
           fetch-depth: 0
 
-      # see https://cibuildwheel.readthedocs.io/en/stable/changelog/
+      - name: Setup Python
-      # for a list of versions
+        uses: actions/setup-python@v5
-      - name: Build wheels
+        with:
-        uses: pypa/cibuildwheel@v2.11.4
+          python-version: ${{ matrix.python-version }}
-        env:
+
-          CIBW_BEFORE_BUILD: pip install torch==${{ matrix.torch}}+cpu -f https://download.pytorch.org/whl/torch_stable.html cmake numpy
+      - name: Install dependencies
-          CIBW_BUILD: ${{ matrix.python-version }}-win_amd64
+        shell: bash
-          CIBW_REPAIR_WHEEL_COMMAND_WINDOWS: ""
+        run: |
-          CIBW_BUILD_VERBOSITY: 3
+          pip install -q torch==${{ matrix.torch}} cmake numpy wheel>=0.40.0 twine setuptools
+          pip install torch==${{ matrix.torch}}+cpu -f https://download.pytorch.org/whl/torch_stable.html cmake numpy || pip install torch==${{ matrix.torch}}+cpu -f https://download.pytorch.org/whl/torch/ cmake numpy
+
+      - name: Build wheel
+        shell: bash
+        run: |
+          python3 setup.py bdist_wheel
+          mkdir wheelhouse
+          cp -v dist/* wheelhouse
 
       - name: Display wheels
         shell: bash
@@ -60,7 +72,7 @@ jobs:
           ls -lh ./wheelhouse/
 
       - name: Upload Wheel
-        uses: actions/upload-artifact@v2
+        uses: actions/upload-artifact@v4
         with:
           name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-windows-latest-cpu
           path: wheelhouse/*.whl
@@ -86,8 +98,9 @@ jobs:
             cd huggingface
             git pull
 
-            mkdir -p windows-cpu
+            d=cpu/1.25.5.dev20241029/windows
-            cp -v ../wheelhouse/*.whl ./windows-cpu
+            mkdir -p $d
+            cp -v ../wheelhouse/*.whl ./$d
             git status
             git lfs track "*.whl"
             git add .

.gitignore

@@ -6,3 +6,5 @@ __pycache__/
 test-1hour.wav
 path.sh
 torch_version.py
+cpu*.html
+cuda*.html

CMakeLists.txt

@@ -1,4 +1,7 @@
 # Copyright (c)  2021  Xiaomi Corporation (author: Fangjun Kuang)
+if (CMAKE_VERSION VERSION_GREATER_EQUAL "4.0.0")
+  set(CMAKE_POLICY_VERSION_MINIMUM 3.5)
+endif()
 
 cmake_minimum_required(VERSION 3.8 FATAL_ERROR)
 
@@ -7,7 +10,7 @@ project(kaldifeat)
 # remember to change the version in
 # scripts/conda/kaldifeat/meta.yaml
 # scripts/conda-cpu/kaldifeat/meta.yaml
-set(kaldifeat_VERSION "1.25.4")
+set(kaldifeat_VERSION "1.25.5")
 
 set(CMAKE_ARCHIVE_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/lib")
 set(CMAKE_LIBRARY_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/lib")
@@ -17,6 +20,7 @@ set(CMAKE_SKIP_BUILD_RPATH FALSE)
 set(BUILD_RPATH_USE_ORIGIN TRUE)
 set(CMAKE_INSTALL_RPATH_USE_LINK_PATH TRUE)
 
+
 if(NOT APPLE)
   set(kaldifeat_rpath_origin "$ORIGIN")
 else()
@@ -33,7 +37,7 @@ if(NOT CMAKE_BUILD_TYPE)
 endif()
 
 if (NOT CMAKE_CXX_STANDARD)
-  set(CMAKE_CXX_STANDARD 14 CACHE STRING "The C++ version to be used.")
+  set(CMAKE_CXX_STANDARD 17 CACHE STRING "The C++ version to be used.")
 endif()
 message(STATUS "C++ Standard version: ${CMAKE_CXX_STANDARD}")
 set(CMAKE_CXX_EXTENSIONS OFF)
@@ -42,10 +46,9 @@ 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}")
 
 if(BUILD_SHARED_LIBS AND MSVC)

README.md

@@ -14,6 +14,20 @@
 **Note**: If you are looking for a version that does not depend on PyTorch,
 please see <https://github.com/csukuangfj/kaldi-native-fbank>
 
+# Installation
+
+Refer to
+<https://csukuangfj.github.io/kaldifeat/installation/from_wheels.html>
+for installation.
+
+> Never use `pip install kaldifeat`
+
+> Never use `pip install kaldifeat`
+
+> Never use `pip install kaldifeat`
+
+
+
 <sub>
 <table>
 <tr>
@@ -312,8 +326,4 @@ extraction.
 
 See <https://github.com/k2-fsa/sherpa/blob/master/sherpa/bin/pruned_stateless_emformer_rnnt2/decode.py>
 
-# Installation
 
-Refer to
-<https://csukuangfj.github.io/kaldifeat>
-for installation.

cmake/pybind11.cmake

@@ -8,19 +8,18 @@ function(download_pybind11)
 
   include(FetchContent)
 
-  # latest commit as of 2022.10.31 that supports python 3.11
+  set(pybind11_URL  "https://github.com/pybind/pybind11/archive/refs/tags/v2.12.0.tar.gz")
-  set(pybind11_URL  "https://github.com/pybind/pybind11/archive/5bc0943ed96836f46489f53961f6c438d2935357.zip")
+  set(pybind11_URL2 "https://hf-mirror.com/csukuangfj/sherpa-onnx-cmake-deps/resolve/main/pybind11-2.12.0.tar.gz")
-  set(pybind11_URL2 "https://huggingface.co/csukuangfj/k2-cmake-deps/resolve/main/pybind11-5bc0943ed96836f46489f53961f6c438d2935357.zip")
+  set(pybind11_HASH "SHA256=bf8f242abd1abcd375d516a7067490fb71abd79519a282d22b6e4d19282185a7")
-  set(pybind11_HASH "SHA256=ff65a1a8c9e6ceec11e7ed9d296f2e22a63e9ff0c4264b3af29c72b4f18f25a0")
 
   # If you don't have access to the Internet,
   # please pre-download pybind11
   set(possible_file_locations
-    $ENV{HOME}/Downloads/pybind11-5bc0943ed96836f46489f53961f6c438d2935357.zip
+    $ENV{HOME}/Downloads/pybind11-2.12.0.tar.gz
-    ${PROJECT_SOURCE_DIR}/pybind11-5bc0943ed96836f46489f53961f6c438d2935357.zip
+    ${CMAKE_SOURCE_DIR}/pybind11-2.12.0.tar.gz
-    ${PROJECT_BINARY_DIR}/pybind11-5bc0943ed96836f46489f53961f6c438d2935357.zip
+    ${CMAKE_BINARY_DIR}/pybind11-2.12.0.tar.gz
-    /tmp/pybind11-5bc0943ed96836f46489f53961f6c438d2935357.zip
+    /tmp/pybind11-2.12.0.tar.gz
-    /star-fj/fangjun/download/github/pybind11-5bc0943ed96836f46489f53961f6c438d2935357.zip
+    /star-fj/fangjun/download/github/pybind11-2.12.0.tar.gz
   )
 
   foreach(f IN LISTS possible_file_locations)

doc/source/installation/from_wheels.rst

@@ -9,6 +9,11 @@ You can find pre-compiled wheels at
 We give a few examples below to show you how to install `kaldifeat`_ from
 pre-compiled wheels.
 
+.. hint::
+
+   The following lists only some examples. We suggest that you always select the
+   latest version of ``kaldifeat``.
+
 Linux (CPU)
 -----------
 
@@ -16,22 +21,29 @@ Suppose you want to install the following wheel:
 
 .. code-block:: bash
 
-   https://huggingface.co/csukuangfj/kaldifeat/resolve/main/ubuntu-cpu/kaldifeat-1.25.3.dev20231221+cpu.torch2.1.2-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
+   https://huggingface.co/csukuangfj/kaldifeat/resolve/main/ubuntu-cpu/kaldifeat-1.25.4.dev20240725+cpu.torch2.4.0-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
 
 you can use one of the following methods:
 
 .. code-block:: bash
 
    # method 1
-   pip install torch==2.1.2+cpu -f https://download.pytorch.org/whl/torch_stable.html
+   pip install torch==2.4.0+cpu -f https://download.pytorch.org/whl/torch/
-   pip install kaldifeat==1.25.3.dev20231221+cpu.torch2.1.2 -f https://csukuangfj.github.io/kaldifeat/cpu.html
+   pip install kaldifeat==1.25.4.dev20240725+cpu.torch2.4.0 -f https://csukuangfj.github.io/kaldifeat/cpu.html
 
+   # For users from China
+   # 中国国内用户，如果访问不了 huggingface, 请使用
+   # pip install kaldifeat==1.25.4.dev20240725+cpu.torch2.4.0 -f https://csukuangfj.github.io/kaldifeat/cpu-cn.html
 
    # method 2
-   pip install torch==2.1.2+cpu -f https://download.pytorch.org/whl/torch_stable.html
+   pip install torch==2.4.0+cpu -f https://download.pytorch.org/whl/torch/
-   wget https://huggingface.co/csukuangfj/kaldifeat/resolve/main/ubuntu-cpu/kaldifeat-1.25.3.dev20231221+cpu.torch2.1.2-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
+   wget https://huggingface.co/csukuangfj/kaldifeat/resolve/main/ubuntu-cpu/kaldifeat-1.25.4.dev20240725+cpu.torch2.4.0-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
-   pip install ./kaldifeat-1.25.3.dev20231221+cpu.torch2.1.2-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
 
+   # For users from China
+   # 中国国内用户，如果访问不了 huggingface, 请使用
+   # wget https://hf-mirror.com/csukuangfj/kaldifeat/resolve/main/ubuntu-cpu/kaldifeat-1.25.4.dev20240725+cpu.torch2.4.0-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
+
+   pip install ./kaldifeat-1.25.4.dev20240725+cpu.torch2.4.0-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
 
 Windows (CPU)
 --------------
@@ -40,20 +52,29 @@ Suppose you want to install the following wheel:
 
 .. code-block:: bash
 
-   https://huggingface.co/csukuangfj/kaldifeat/resolve/main/windows-cpu/kaldifeat-1.25.3.dev20231221+cpu.torch2.1.2-cp311-cp311-win_amd64.whl
+   https://huggingface.co/csukuangfj/kaldifeat/resolve/main/windows-cpu/kaldifeat-1.25.4.dev20240725+cpu.torch2.4.0-cp312-cp312-win_amd64.whl
 
 you can use one of the following methods:
 
 .. code-block:: bash
 
    # method 1
-   pip install torch==2.1.2+cpu -f https://download.pytorch.org/whl/torch_stable.html
+   pip install torch==2.4.0+cpu -f https://download.pytorch.org/whl/torch/
-   pip install kaldifeat==1.25.3.dev20231221+cpu.torch2.1.2 -f https://csukuangfj.github.io/kaldifeat/cpu.html
+   pip install kaldifeat==1.25.4.dev20240725+cpu.torch2.4.0 -f https://csukuangfj.github.io/kaldifeat/cpu.html
+
+   # For users from China
+   # 中国国内用户，如果访问不了 huggingface, 请使用
+   # pip install kaldifeat==1.25.4.dev20240725+cpu.torch2.4.0 -f https://csukuangfj.github.io/kaldifeat/cpu-cn.html
 
    # method 2
-   pip install torch==2.1.2+cpu -f https://download.pytorch.org/whl/torch_stable.html
+   pip install torch==2.4.0+cpu -f https://download.pytorch.org/whl/torch/
-   wget https://huggingface.co/csukuangfj/kaldifeat/resolve/main/windows-cpu/kaldifeat-1.25.3.dev20231221+cpu.torch2.1.2-cp311-cp311-win_amd64.whl
+   wget https://huggingface.co/csukuangfj/kaldifeat/resolve/main/windows-cpu/kaldifeat-1.25.4.dev20240725+cpu.torch2.4.0-cp312-cp312-win_amd64.whl
-   pip install ./kaldifeat-1.25.3.dev20231221+cpu.torch2.1.2-cp311-cp311-win_amd64.whl
+
+   # For users from China
+   # 中国国内用户，如果访问不了 huggingface, 请使用
+   # wget https://hf-mirror.com/csukuangfj/kaldifeat/resolve/main/windows-cpu/kaldifeat-1.25.4.dev20240725+cpu.torch2.4.0-cp312-cp312-win_amd64.whl
+
+   pip install ./kaldifeat-1.25.4.dev20240725+cpu.torch2.4.0-cp312-cp312-win_amd64.whl
 
 macOS (CPU)
 -----------
@@ -62,20 +83,29 @@ Suppose you want to install the following wheel:
 
 .. code-block:: bash
 
-   https://huggingface.co/csukuangfj/kaldifeat/resolve/main/macos/kaldifeat-1.25.3.dev20231221+cpu.torch2.1.2-cp311-cp311-macosx_10_9_x86_64.whl
+   https://huggingface.co/csukuangfj/kaldifeat/resolve/main/macos/kaldifeat-1.25.4.dev20240725+cpu.torch2.4.0-cp311-cp311-macosx_11_0_arm64.whl
 
 you can use one of the following methods:
 
 .. code-block:: bash
 
    # method 1
-   pip install torch==2.1.2
+   pip install torch==2.4.0
-   pip install kaldifeat==1.25.3.dev20231221+cpu.torch2.1.2 -f https://csukuangfj.github.io/kaldifeat/cpu.html
+   pip install kaldifeat==1.25.4.dev20240725+cpu.torch2.4.0 -f https://csukuangfj.github.io/kaldifeat/cpu.html
+
+   # For users from China
+   # 中国国内用户，如果访问不了 huggingface, 请使用
+   # pip install kaldifeat==1.25.4.dev20240725+cpu.torch2.4.0 -f https://csukuangfj.github.io/kaldifeat/cpu-cn.html
 
    # method 2
-   pip install torch==2.1.2 -f https://download.pytorch.org/whl/torch_stable.html
+   pip install torch==2.4.0 -f https://download.pytorch.org/whl/torch/
-   wget https://huggingface.co/csukuangfj/kaldifeat/resolve/main/macos/kaldifeat-1.25.3.dev20231221+cpu.torch2.1.2-cp311-cp311-macosx_10_9_x86_64.whl
+   wget https://huggingface.co/csukuangfj/kaldifeat/resolve/main/macos/kaldifeat-1.25.4.dev20240725+cpu.torch2.4.0-cp311-cp311-macosx_11_0_arm64.whl
-   pip install ./kaldifeat-1.25.3.dev20231221+cpu.torch2.1.2-cp311-cp311-macosx_10_9_x86_64.whl
+
+   # For users from China
+   # 中国国内用户，如果访问不了 huggingface, 请使用
+   # wget https://hf-mirror.com/csukuangfj/kaldifeat/resolve/main/macos/kaldifeat-1.25.4.dev20240725+cpu.torch2.4.0-cp311-cp311-macosx_11_0_arm64.whl
+
+   pip install ./kaldifeat-1.25.4.dev20240725+cpu.torch2.4.0-cp311-cp311-macosx_11_0_arm64.whl
 
 Linux (CUDA)
 ------------
@@ -84,17 +114,26 @@ Suppose you want to install the following wheel:
 
 .. code-block:: bash
 
-   https://huggingface.co/csukuangfj/kaldifeat/resolve/main/ubuntu-cuda/kaldifeat-1.25.3.dev20231221+cuda12.1.torch2.1.2-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
+   https://huggingface.co/csukuangfj/kaldifeat/resolve/main/ubuntu-cuda/kaldifeat-1.25.4.dev20240725+cuda12.4.torch2.4.0-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
 
 you can use one of the following methods:
 
 .. code-block:: bash
 
    # method 1
-   pip install torch==2.1.2+cu121 -f https://download.pytorch.org/whl/torch_stable.html
+   pip install torch==2.4.0+cu124 -f https://download.pytorch.org/whl/torch/
-   pip install kaldifeat==1.25.3.dev20231221+cuda12.1.torch2.1.2 -f https://csukuangfj.github.io/kaldifeat/cuda.html
+   pip install kaldifeat==1.25.4.dev20240725+cuda12.4.torch2.4.0 -f https://csukuangfj.github.io/kaldifeat/cuda.html
+
+   # For users from China
+   # 中国国内用户，如果访问不了 huggingface, 请使用
+   # pip install kaldifeat==1.25.4.dev20240725+cuda12.4.torch2.4.0 -f https://csukuangfj.github.io/kaldifeat/cuda-cn.html
 
    # method 2
-   pip install torch==2.1.2+cu121 -f https://download.pytorch.org/whl/torch_stable.html
+   pip install torch==2.4.0+cu124 -f https://download.pytorch.org/whl/torch/
-   wget https://huggingface.co/csukuangfj/kaldifeat/resolve/main/ubuntu-cuda/kaldifeat-1.25.3.dev20231221+cuda12.1.torch2.1.2-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
+   wget https://huggingface.co/csukuangfj/kaldifeat/resolve/main/ubuntu-cuda/kaldifeat-1.25.4.dev20240725+cuda12.4.torch2.4.0-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
-   pip install ./kaldifeat-1.25.3.dev20231221+cuda12.1.torch2.1.2-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
+
+   # For users from China
+   # 中国国内用户，如果访问不了 huggingface, 请使用
+   # wget https://hf-mirror.com/csukuangfj/kaldifeat/resolve/main/ubuntu-cuda/kaldifeat-1.25.4.dev20240725+cuda12.4.torch2.4.0-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
+
+   pip install ./kaldifeat-1.25.4.dev20240725+cuda12.4.torch2.4.0-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl

kaldifeat/csrc/mel-computations.cc

@@ -138,7 +138,7 @@ MelBanks::MelBanks(const MelBanksOptions &opts,
                   << " and vtln-high " << vtln_high << ", versus "
                   << "low-freq " << low_freq << " and high-freq " << high_freq;
 
-  // we will transpose bins_mat_ at the end of this funciton
+  // we will transpose bins_mat_ at the end of this function
   bins_mat_ = torch::zeros({num_bins, num_fft_bins}, torch::kFloat);
 
   int32_t stride = bins_mat_.strides()[0];
@@ -179,12 +179,14 @@ MelBanks::MelBanks(const MelBanksOptions &opts,
         last_index = i;
       }
     }
-    KALDIFEAT_ASSERT(first_index != -1 && last_index >= first_index &&
+
-                     "You may have set num_mel_bins too large.");
+    // Note: It is possible that first_index == last_index == -1 at this line.
 
     // Replicate a bug in HTK, for testing purposes.
-    if (opts.htk_mode && bin == 0 && mel_low_freq != 0.0f)
+    if (opts.htk_mode && bin == 0 && mel_low_freq != 0.0f &&
+        first_index != -1) {
       this_bin[first_index] = 0.0f;
+    }
   }
 
   if (debug_) KALDIFEAT_LOG << bins_mat_;

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

scripts/conda-cpu/kaldifeat/meta.yaml

@@ -1,6 +1,6 @@
 package:
   name: kaldifeat
-  version: "1.25.4"
+  version: "1.25.5"
 
 source:
   path: "{{ environ.get('KALDIFEAT_ROOT_DIR') }}"

scripts/conda/kaldifeat/meta.yaml

@@ -1,6 +1,6 @@
 package:
   name: kaldifeat
-  version: "1.25.4"
+  version: "1.25.5"
 
 source:
   path: "{{ environ.get('KALDIFEAT_ROOT_DIR') }}"

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

scripts/github_actions/build-ubuntu-cpu.sh

@@ -15,49 +15,21 @@ if [ -z $TORCH_VERSION ]; then
   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 >/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
 python3 -m pip install bs4 requests tqdm auditwheel
 
 echo "Installing torch"
-python3 -m pip install -qq torch==$TORCH_VERSION+cpu -f https://download.pytorch.org/whl/torch_stable.html
+python3 -m pip install -qq torch==$TORCH_VERSION+cpu -f https://download.pytorch.org/whl/torch_stable.html || \
+python3 -m pip install -qq torch==$TORCH_VERSION+cpu -f https://download.pytorch.org/whl/torch/
 
 rm -rf ~/.cache/pip
 yum clean all
@@ -68,8 +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 \
@@ -90,8 +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 \
+  --plat $plat \
-  -w /var/www/wheelhouse \
   dist/*.whl
 
 ls -lh  /var/www

scripts/github_actions/build-ubuntu-cuda.sh

@@ -22,44 +22,12 @@ if [ -z $CUDA_VERSION ]; then
   exit 1
 fi
 
-
+if [[ $TORCH_VERSION =~ 2.2.* && $CUDA_VERSION =~ 12.* ]]; then
-yum -y install openssl-devel bzip2-devel libffi-devel xz-devel wget redhat-lsb-core
+  # see https://github.com/pytorch/pytorch/issues/113948
-
+  export TORCH_CUDA_ARCH_LIST="8.0 8.6 8.9 9.0"
-
-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 >/dev/null 2>&1
+# python3 -m pip install scikit-build
-make install >/dev/null 2>&1
-
-popd
-
-echo "pwd: $PWD"
-# rm -rf Python-${PYTHON_VERSION}.3
-
-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 -U pip cmake
 python3 -m pip install wheel twine typing_extensions
 python3 -m pip install bs4 requests tqdm auditwheel
@@ -67,6 +35,56 @@ 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
+# Call Stack (most recent call first):
+#   /Python-3.8.2/py-3.8/lib/python3.8/site-packages/torch/share/cmake/Caffe2/public/utils.cmake:401 (cuda_select_nvcc_arch_flags)
+#   /Python-3.8.2/py-3.8/lib/python3.8/site-packages/torch/share/cmake/Caffe2/public/cuda.cmake:342 (torch_cuda_get_nvcc_gencode_flag)
+#   /Python-3.8.2/py-3.8/lib/python3.8/site-packages/torch/share/cmake/Caffe2/Caffe2Config.cmake:87 (include)
+#   /Python-3.8.2/py-3.8/lib/python3.8/site-packages/torch/share/cmake/Torch/TorchConfig.cmake:68 (find_package)
+#   cmake/torch.cmake:14 (find_package)
+#   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
 
@@ -76,9 +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 \
@@ -91,14 +120,34 @@ auditwheel --verbose repair \
   --exclude libtorch_cuda.so \
   --exclude libtorch_python.so \
   \
-  --exclude libcudnn.so.8 \
+  --exclude libcublas.so \
   --exclude libcublas.so.11 \
+  --exclude libcublas.so.12 \
+  --exclude libcublasLt.so \
   --exclude libcublasLt.so.11 \
+  --exclude libcublasLt.so.12 \
   --exclude libcudart.so.11.0 \
+  --exclude libcudart.so.12 \
+  --exclude libcudnn.so.8 \
+  --exclude libcufft.so \
+  --exclude libcufft.so.11 \
+  --exclude libcupti.so \
+  --exclude libcupti.so.12 \
+  --exclude libcurand.so \
+  --exclude libcurand.so.10 \
+  --exclude libcusparse.so \
+  --exclude libcusparse.so.12 \
+  --exclude libnccl.so \
+  --exclude libnccl.so.2 \
+  --exclude libnvJitLink.so \
+  --exclude libnvJitLink.so.12 \
+  --exclude libnvrtc.so \
   --exclude libnvrtc.so.11.2 \
-  --exclude libtorch_cuda_cu.so \
+  --exclude libnvrtc.so.12 \
+  --exclude libshm.so \
   --exclude libtorch_cuda_cpp.so \
-  --plat manylinux_2_17_x86_64 \
+  --exclude libtorch_cuda_cu.so \
+  --plat $plat \
   -w /var/www/wheelhouse \
   dist/*.whl
 

scripts/github_actions/generate_build_matrix.py

@@ -1,10 +1,31 @@
 #!/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
 
 
+def version_ge(a, b):
+    a_major, a_minor = list(map(int, a.split(".")))[:2]
+    b_major, b_minor = list(map(int, b.split(".")))[:2]
+    if a_major > b_major:
+        return True
+
+    if a_major == b_major and a_minor >= b_minor:
+        return True
+
+    return False
+
+
 def get_args():
     parser = argparse.ArgumentParser()
     parser.add_argument(
@@ -28,6 +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",
@@ -38,7 +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
@@ -53,77 +83,91 @@ 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"],
         #  },
-        "1.6.0": {
+        #  "1.6.0": {
-            "python-version": ["3.6", "3.7", "3.8"],
+        #      "python-version": ["3.6", "3.7", "3.8"],
-            "cuda": ["10.1", "10.2"] if not for_windows else ["10.1.243", "10.2.89"],
+        #      "cuda": ["10.1", "10.2"] if not for_windows else ["10.1.243", "10.2.89"],
-        },
+        #  },
-        "1.7.0": {
+        #  "1.7.0": {
-            "python-version": ["3.6", "3.7", "3.8"],
+        #      "python-version": ["3.6", "3.7", "3.8"],
-            "cuda": ["10.1", "10.2", "11.0"]
+        #      "cuda": (
-            if not for_windows
+        #          ["10.1", "10.2", "11.0"]
-            else ["10.1.243", "10.2.89", "11.0.3"],
+        #          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"]
+        #  "1.7.1": {
-            if not for_windows
+        #      "python-version": ["3.6", "3.7", "3.8", "3.9"],
-            else ["10.1.243", "10.2.89", "11.0.3"],
+        #      "cuda": (
-        },
+        #          ["10.1", "10.2", "11.0"]
-        "1.8.0": {
+        #          if not for_windows
-            "python-version": ["3.6", "3.7", "3.8", "3.9"],
+        #          else ["10.1.243", "10.2.89", "11.0.3"]
-            "cuda": ["10.1", "10.2", "11.1"]
+        #      ),
-            if not for_windows
+        #  },
-            else ["10.1.243", "10.2.89", "11.1.1"],
+        #  "1.8.0": {
-        },
+        #      "python-version": ["3.6", "3.7", "3.8", "3.9"],
-        "1.8.1": {
+        #      "cuda": (
-            "python-version": ["3.6", "3.7", "3.8", "3.9"],
+        #          ["10.1", "10.2", "11.1"]
-            "cuda": ["10.1", "10.2", "11.1"]
+        #          if not for_windows
-            if not for_windows
+        #          else ["10.1.243", "10.2.89", "11.1.1"]
-            else ["10.1.243", "10.2.89", "11.1.1"],
+        #      ),
-        },
+        #  },
-        "1.9.0": {
+        #  "1.8.1": {
-            "python-version": ["3.6", "3.7", "3.8", "3.9"],
+        #      "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"],
+        #      "cuda": (
-        },
+        #          ["10.1", "10.2", "11.1"]
-        "1.9.1": {
+        #          if not for_windows
-            "python-version": ["3.6", "3.7", "3.8", "3.9"],
+        #          else ["10.1.243", "10.2.89", "11.1.1"]
-            "cuda": ["10.2", "11.1"] if not for_windows else ["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": ["10.2", "11.1", "11.3"]
+            "cuda": (
-            if not for_windows
+                ["10.2", "11.1", "11.3"]
-            else ["10.2.89", "11.1.1", "11.3.1"],
+                if not for_windows
+                else ["10.2.89", "11.1.1", "11.3.1"]
+            ),
         },
         "1.10.1": {
             "python-version": ["3.6", "3.7", "3.8", "3.9"],
-            "cuda": ["10.2", "11.1", "11.3"]
+            "cuda": (
-            if not for_windows
+                ["10.2", "11.1", "11.3"]
-            else ["10.2.89", "11.1.1", "11.3.1"],
+                if not for_windows
+                else ["10.2.89", "11.1.1", "11.3.1"]
+            ),
         },
         "1.10.2": {
             "python-version": ["3.6", "3.7", "3.8", "3.9"],
-            "cuda": ["10.2", "11.1", "11.3"]
+            "cuda": (
-            if not for_windows
+                ["10.2", "11.1", "11.3"]
-            else ["10.2.89", "11.1.1", "11.3.1"],
+                if not for_windows
+                else ["10.2.89", "11.1.1", "11.3.1"]
+            ),
         },
         "1.11.0": {
             "python-version": ["3.7", "3.8", "3.9", "3.10"],
-            "cuda": ["10.2", "11.3", "11.5"]
+            "cuda": (
-            if not for_windows
+                ["10.2", "11.3", "11.5"] if not for_windows else ["11.3.1", "11.5.2"]
-            else ["11.3.1", "11.5.2"],
+            ),
         },
         "1.12.0": {
             "python-version": ["3.7", "3.8", "3.9", "3.10"],
-            "cuda": ["10.2", "11.3", "11.6"]
+            "cuda": (
-            if not for_windows
+                ["10.2", "11.3", "11.6"] if not for_windows else ["11.3.1", "11.6.2"]
-            else ["11.3.1", "11.6.2"],
+            ),
         },
         "1.12.1": {
             "python-version": ["3.7", "3.8", "3.9", "3.10"],
-            "cuda": ["10.2", "11.3", "11.6"]
+            "cuda": (
-            if not for_windows
+                ["10.2", "11.3", "11.6"] if not for_windows else ["11.3.1", "11.6.2"]
-            else ["11.3.1", "11.6.2"],
+            ),
         },
         "1.13.0": {
             "python-version": ["3.7", "3.8", "3.9", "3.10", "3.11"],
@@ -131,47 +175,171 @@ def generate_build_matrix(enable_cuda, for_windows, for_macos, test_only_latest_
         },
         "1.13.1": {
             "python-version": ["3.7", "3.8", "3.9", "3.10", "3.11"],
-            "cuda": ["11.6", "11.7"]  # default 11.7
+            "cuda": (
-            if not for_windows
+                ["11.6", "11.7"]  # default 11.7
-            else ["11.6.2", "11.7.1"],
+                if not for_windows
+                else ["11.6.2", "11.7.1"]
+            ),
         },
         "2.0.0": {
             "python-version": ["3.8", "3.9", "3.10", "3.11"],
-            "cuda": ["11.7", "11.8"]  # default 11.7
+            "cuda": (
-            if not for_windows
+                ["11.7", "11.8"]  # default 11.7
-            else ["11.7.1", "11.8.0"],
+                if not for_windows
+                else ["11.7.1", "11.8.0"]
+            ),
         },
         "2.0.1": {
             "python-version": ["3.8", "3.9", "3.10", "3.11"],
-            "cuda": ["11.7", "11.8"]  # default 11.7
+            "cuda": (
-            if not for_windows
+                ["11.7", "11.8"]  # default 11.7
-            else ["11.7.1", "11.8.0"],
+                if not for_windows
+                else ["11.7.1", "11.8.0"]
+            ),
         },
         "2.1.0": {
             "python-version": ["3.8", "3.9", "3.10", "3.11"],
-            "cuda": ["11.8", "12.1"]  # default 12.1
+            "cuda": (
-            if not for_windows
+                ["11.8", "12.1"]  # default 12.1
-            else ["11.8.0", "12.1.0"],
+                if not for_windows
+                else ["11.8.0", "12.1.0"]
+            ),
         },
         "2.1.1": {
             "python-version": ["3.8", "3.9", "3.10", "3.11"],
-            "cuda": ["11.8", "12.1"]  # default 12.1
+            "cuda": (
-            if not for_windows
+                ["11.8", "12.1"]  # default 12.1
-            else ["11.8.0", "12.1.0"],
+                if not for_windows
+                else ["11.8.0", "12.1.0"]
+            ),
         },
         "2.1.2": {
             "python-version": ["3.8", "3.9", "3.10", "3.11"],
-            "cuda": ["11.8", "12.1"]  # default 12.1
+            "cuda": (
-            if not for_windows
+                ["11.8", "12.1"]  # default 12.1
-            else ["11.8.0", "12.1.0"],
+                if not for_windows
+                else ["11.8.0", "12.1.0"]
+            ),
+        },
+        "2.2.0": {
+            "python-version": ["3.8", "3.9", "3.10", "3.11", "3.12"],
+            "cuda": (
+                ["11.8", "12.1"]  # default 12.1
+                if not for_windows
+                else ["11.8.0", "12.1.0"]
+            ),
+        },
+        "2.2.1": {
+            "python-version": ["3.8", "3.9", "3.10", "3.11", "3.12"],
+            "cuda": (
+                ["11.8", "12.1"]  # default 12.1
+                if not for_windows
+                else ["11.8.0", "12.1.0"]
+            ),
+        },
+        "2.2.2": {
+            "python-version": ["3.8", "3.9", "3.10", "3.11", "3.12"],
+            "cuda": (
+                ["11.8", "12.1"]  # default 12.1
+                if not for_windows
+                else ["11.8.0", "12.1.0"]
+            ),
+        },
+        "2.3.0": {
+            "python-version": ["3.8", "3.9", "3.10", "3.11", "3.12"],
+            "cuda": (
+                ["11.8", "12.1"]  # default 12.1
+                if not for_windows
+                else ["11.8.0", "12.1.0"]
+            ),
+        },
+        "2.3.1": {
+            "python-version": ["3.8", "3.9", "3.10", "3.11", "3.12"],
+            "cuda": (
+                ["11.8", "12.1"]  # default 12.1
+                if not for_windows
+                else ["11.8.0", "12.1.0"]
+            ),
+        },
+        "2.4.0": {
+            "python-version": ["3.8", "3.9", "3.10", "3.11", "3.12"],
+            "cuda": (
+                ["11.8", "12.1", "12.4"]  # default 12.1
+                if not for_windows
+                else ["11.8.0", "12.1.0", "12.4.0"]
+            ),
+        },
+        "2.4.1": {
+            "python-version": ["3.8", "3.9", "3.10", "3.11", "3.12"],
+            "cuda": (
+                ["11.8", "12.1", "12.4"]  # default 12.1
+                if not for_windows
+                else ["11.8.0", "12.1.0", "12.4.0"]
+            ),
+        },
+        "2.5.0": {
+            # Only Linux supports python 3.13
+            "python-version": ["3.9", "3.10", "3.11", "3.12", "3.13"],
+            "cuda": (
+                ["11.8", "12.1", "12.4"]  # default 12.4
+                if not for_windows
+                else ["11.8.0", "12.1.0", "12.4.0"]
+            ),
+        },
+        "2.5.1": {
+            # Only Linux supports python 3.13
+            "python-version": ["3.9", "3.10", "3.11", "3.12", "3.13"],
+            "cuda": (
+                ["11.8", "12.1", "12.4"]  # default 12.4
+                if not for_windows
+                else ["11.8.0", "12.1.0", "12.4.0"]
+            ),
+        },
+        "2.6.0": {
+            "python-version": ["3.9", "3.10", "3.11", "3.12", "3.13"],
+            "cuda": (
+                ["11.8", "12.4", "12.6"]  # default 12.4
+                if not for_windows
+                else ["11.8.0", "12.4.0", "12.6.0"]
+            ),
+        },
+        "2.7.0": {
+            "python-version": ["3.9", "3.10", "3.11", "3.12", "3.13"],
+            "cuda": (
+                ["11.8", "12.6", "12.8"]
+                if not for_windows
+                else ["11.8.0", "12.6.2", "12.8.1"]
+            ),
+        },
+        "2.7.1": {
+            "python-version": ["3.9", "3.10", "3.11", "3.12", "3.13"],
+            "cuda": (
+                ["11.8", "12.6", "12.8"]  # default 12.6
+                if not for_windows
+                else ["11.8.0", "12.6.2", "12.8.1"]
+            ),
+        },
+        "2.8.0": {
+            "python-version": ["3.9", "3.10", "3.11", "3.12", "3.13"],
+            "cuda": (
+                ["12.6", "12.8", "12.9"]  # default 12.8
+                if not for_windows
+                else ["12.6.2", "12.8.1", "12.9.1"]
+            ),
         },
         # https://github.com/Jimver/cuda-toolkit/blob/master/src/links/windows-links.ts
     }
     if test_only_latest_torch:
-        latest = "2.1.2"
+        latest = "2.8.0"
         matrix = {latest: matrix[latest]}
 
     if for_windows or for_macos:
+        if "2.5.1" in matrix:
+            matrix["2.5.1"]["python-version"].remove("3.13")
+
+        if "2.5.0" in matrix:
+            matrix["2.5.0"]["python-version"].remove("3.13")
+
         if "1.13.0" in matrix:
             matrix["1.13.0"]["python-version"].remove("3.11")
 
@@ -179,9 +347,18 @@ 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:
@@ -192,12 +369,30 @@ 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:
@@ -205,15 +400,53 @@ def generate_build_matrix(enable_cuda, for_windows, for_macos, test_only_latest_
                 if p in excluded_python_versions:
                     continue
 
-                if for_windows or for_macos:
+                if for_windows:
-                    p = "cp" + "".join(p.split("."))
                     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"
+                            if not for_arm64
+                            else "pytorch/manylinuxaarch64-builder:cpu-aarch64-195148266541a9789074265141cb7dc19dc14c54",
+                            "is_2_28": "0",
+                        }
+                    )
+                elif version_ge(torch, "2.2.0"):
+                    ans.append(
+                        {
+                            "torch": torch,
+                            "python-version": p,
+                            "image": "pytorch/manylinux-builder:cpu-2.2"
+                            if not for_arm64
+                            else "pytorch/manylinuxaarch64-builder:cpu-aarch64-195148266541a9789074265141cb7dc19dc14c54",
+                            "is_2_28": "0",
+                        }
+                    )
                 else:
                     ans.append(
                         {
                             "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",
                         }
                     )
 
@@ -226,6 +459,7 @@ 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,
     )
 

scripts/github_actions/install_cuda.sh

@@ -55,6 +55,9 @@ case "$cuda" in
   12.1)
     url=https://developer.download.nvidia.com/compute/cuda/12.1.0/local_installers/cuda_12.1.0_530.30.02_linux.run
     ;;
+  12.4)
+    url=https://developer.download.nvidia.com/compute/cuda/12.4.0/local_installers/cuda_12.4.0_550.54.14_linux.run
+    ;;
   *)
     echo "Unknown cuda version: $cuda"
     exit 1

scripts/github_actions/install_cudnn.sh

@@ -48,6 +48,9 @@ case $cuda in
   12.1)
     filename=cudnn-linux-x86_64-8.9.5.29_cuda12-archive.tar.xz
     ;;
+  12.4)
+    filename=cudnn-linux-x86_64-8.9.7.29_cuda12-archive.tar.xz
+    ;;
   *)
     echo "Unsupported cuda version: $cuda"
     exit 1

scripts/github_actions/install_torch.sh

@@ -14,6 +14,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+set -x
+
 if [ x"$TORCH_VERSION" != x"" ] && [ x"$CUDA_VERSION" != x"" ]; then
     torch=$TORCH_VERSION
     cuda=$CUDA_VERSION
@@ -185,6 +187,117 @@ case ${torch} in
         ;;
     esac
     ;;
+  2.2.*)
+    case ${cuda} in
+      11.8)
+        package="torch==${torch}+cu118"
+        url=https://download.pytorch.org/whl/torch_stable.html
+        ;;
+      12.1)
+        package="torch==${torch}"
+        # Leave it empty to use PyPI.
+        url=
+        ;;
+    esac
+    ;;
+  2.3.*)
+    case ${cuda} in
+      11.8)
+        package="torch==${torch}+cu118"
+        url=https://download.pytorch.org/whl/torch_stable.html
+        ;;
+      12.1)
+        package="torch==${torch}"
+        # Leave it empty to use PyPI.
+        url=
+        ;;
+    esac
+    ;;
+  2.4.*)
+    case ${cuda} in
+      11.8)
+        package="torch==${torch}+cu118"
+        # https://download.pytorch.org/whl/nightly/torch/
+        url=https://download.pytorch.org/whl/torch/
+        ;;
+      12.1)
+        package="torch==${torch}+cu121"
+        url=https://download.pytorch.org/whl/torch/
+        ;;
+      12.4)
+        package="torch==${torch}+cu124"
+        url=https://download.pytorch.org/whl/torch/
+        ;;
+    esac
+    ;;
+  2.5.*)
+    case ${cuda} in
+      11.8)
+        package="torch==${torch}+cu118"
+        # https://download.pytorch.org/whl/nightly/torch/
+        url=https://download.pytorch.org/whl/torch/
+        ;;
+      12.1)
+        package="torch==${torch}+cu121"
+        url=https://download.pytorch.org/whl/torch/
+        ;;
+      12.4)
+        package="torch==${torch}+cu124"
+        url=https://download.pytorch.org/whl/torch/
+        ;;
+    esac
+    ;;
+  2.6.*)
+    case ${cuda} in
+      11.8)
+        package="torch==${torch}+cu118"
+        # https://download.pytorch.org/whl/nightly/torch/
+        url=https://download.pytorch.org/whl/torch/
+        ;;
+      12.4)
+        package="torch==${torch}+cu124"
+        url=https://download.pytorch.org/whl/torch/
+        ;;
+      12.6)
+        package="torch==${torch}+cu126"
+        url=https://download.pytorch.org/whl/torch/
+        ;;
+    esac
+    ;;
+  2.7.*)
+    case ${cuda} in
+      11.8)
+        package="torch==${torch}+cu118"
+        # https://download.pytorch.org/whl/nightly/torch/
+        url=https://download.pytorch.org/whl/torch/
+        ;;
+      12.6)
+        package="torch==${torch}+cu126"
+        url=https://download.pytorch.org/whl/torch/
+        ;;
+      12.8)
+        package="torch==${torch}+cu128"
+        url=https://download.pytorch.org/whl/torch/
+        ;;
+    esac
+    ;;
+  2.8.*)
+    case ${cuda} in
+      12.6)
+        package="torch==${torch}+cu126"
+        # https://download.pytorch.org/whl/nightly/torch/
+        url=https://download.pytorch.org/whl/torch/
+        ;;
+      12.8)
+        package="torch==${torch}+cu128"
+        url=https://download.pytorch.org/whl/torch/
+        ;;
+      12.9)
+        package="torch==${torch}+cu129"
+        url=https://download.pytorch.org/whl/torch/
+        ;;
+    esac
+    ;;
   *)
     echo "Unsupported PyTorch version: ${torch}"
     exit 1
@@ -201,4 +314,6 @@ 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