From f9bc968de76a4ad673129862f795a7377253352c Mon Sep 17 00:00:00 2001
From: youge325 <A372707325@126.com>
Date: Fri, 10 Apr 2026 15:35:17 +0800
Subject: [PATCH 01/14] Publish ARM wheels to GitHub

---
 .github/workflows/paddle_arm_wheel.yaml | 219 ++++++++++++++++++++++++
 1 file changed, 219 insertions(+)
 create mode 100644 .github/workflows/paddle_arm_wheel.yaml

diff --git a/.github/workflows/paddle_arm_wheel.yaml b/.github/workflows/paddle_arm_wheel.yaml
new file mode 100644
index 000000000..a9f3f0805
--- /dev/null
+++ b/.github/workflows/paddle_arm_wheel.yaml
@@ -0,0 +1,219 @@
+name: Build ARM Wheels for Paddle
+
+on:
+  push:
+    branches: [paddle]
+    tags: ["v*"]
+  pull_request:
+  merge_group:
+  workflow_dispatch:
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref_name }}-${{ github.ref_type == 'branch' && github.sha }}-${{ github.event_name == 'workflow_dispatch' }}
+  cancel-in-progress: true
+
+permissions:
+  id-token: write
+  contents: write
+
+defaults:
+  run:
+    shell: bash -l -eo pipefail {0}
+
+jobs:
+  build-paddlecodec-wheel-arm:
+    # If your repo does not have GitHub-hosted ARM runners enabled,
+    # replace this with self-hosted labels, e.g. [self-hosted, Linux, ARM64].
+    runs-on: ubuntu-24.04-arm
+    container:
+      image: quay.io/pypa/manylinux_2_28_aarch64
+    strategy:
+      fail-fast: false
+      matrix:
+        python-version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
+    permissions:
+      id-token: write
+      contents: read
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v6
+
+      - name: Setup conda environment
+        uses: conda-incubator/setup-miniconda@v3
+        with:
+          auto-update-conda: true
+          miniforge-version: latest
+          activate-environment: build
+          python-version: ${{ matrix.python-version }}
+
+      - name: Install build dependencies
+        run: |
+          python -m pip install --upgrade pip
+          pip install build wheel setuptools
+
+      - name: Install PaddlePaddle nightly
+        run: |
+          pip install --pre paddlepaddle -i https://www.paddlepaddle.org.cn/packages/nightly/cpu/
+
+      - name: Run pre-build script
+        run: |
+          bash packaging/pre_build_script.sh
+
+      - name: Build wheel
+        run: |
+          # Use pre-built FFmpeg from PyTorch S3
+          export BUILD_AGAINST_ALL_FFMPEG_FROM_S3=1
+          export TORCHCODEC_CMAKE_BUILD_DIR=$(pwd)/build_cmake
+          python -m build --wheel -vvv --no-isolation
+
+      - name: Repair wheel
+        run: |
+          pip install auditwheel
+
+          # 1. Extract internal libraries from the wheel to a temporary directory
+          # This allows auditwheel to find them when checking dependencies
+          mkdir -p temp_libs
+          unzip -j dist/*.whl "torchcodec/*.so" -d temp_libs || true
+
+          # 2. Prepare LD_LIBRARY_PATH
+          # FFmpeg libraries
+          FFMPEG_LIB_PATHS=$(find $(pwd)/build_cmake/_deps -type d -name "lib" | tr '\n' ':')
+          # PaddlePaddle libraries
+          PADDLE_PATH=$(python -c "import paddle; print(paddle.__path__[0])")
+          PADDLE_LIB_PATHS="$PADDLE_PATH/base:$PADDLE_PATH/libs"
+          # Wheel internal libraries
+          INTERNAL_LIB_PATH=$(pwd)/temp_libs
+
+          export LD_LIBRARY_PATH=${FFMPEG_LIB_PATHS}${PADDLE_LIB_PATHS}:${INTERNAL_LIB_PATH}:${LD_LIBRARY_PATH}
+
+          # 3. Repair wheel with auditwheel
+          # We exclude all external libraries because we want to rely on system libraries (like FFmpeg)
+          # or libraries provided by other packages (like PaddlePaddle).
+          # auditwheel 6.1.0+ supports wildcards in --exclude.
+          auditwheel repair dist/*.whl --plat manylinux_2_28_aarch64 -w wheelhouse/ --exclude "*"
+
+          # Cleanup
+          rm -rf temp_libs
+          rm dist/*.whl
+          mv wheelhouse/*.whl dist/
+          rmdir wheelhouse
+
+      - name: Upload wheel artifact
+        uses: actions/upload-artifact@v5
+        with:
+          name: paddlecodec-wheel-linux-arm64-py${{ matrix.python-version }}
+          path: dist/*.whl
+
+      - name: Run post-build script
+        run: |
+          bash packaging/post_build_script.sh
+
+      - name: List wheel contents
+        run: |
+          wheel_path=$(find dist -type f -name "*.whl")
+          echo "Wheel path: $wheel_path"
+          unzip -l $wheel_path
+
+  test-paddlecodec-wheel-arm:
+    needs: build-paddlecodec-wheel-arm
+    runs-on: ubuntu-24.04-arm
+    strategy:
+      fail-fast: false
+      matrix:
+        python-version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
+        # FFmpeg 8.0 depends on libopenvino.so.2520, PaddlePaddle CPU depends on libopenvino.so.2500
+        # There has some conflict causing test failures, but it works with PaddlePaddle GPU.
+        # We skip FFmpeg 8.0 tests for PaddlePaddle CPU builds for now.
+        ffmpeg-version: ["4.4.2", "5.1.2", "6.1.1", "7.0.1"]
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v6
+
+      - name: Set up Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python-version }}
+
+      - name: Download wheel artifact
+        uses: actions/download-artifact@v4
+        with:
+          name: paddlecodec-wheel-linux-arm64-py${{ matrix.python-version }}
+          path: dist/
+
+      - name: Install FFmpeg via conda
+        uses: conda-incubator/setup-miniconda@v3
+        with:
+          auto-update-conda: true
+          miniforge-version: latest
+          activate-environment: test
+          python-version: ${{ matrix.python-version }}
+
+      - name: Install FFmpeg from conda-forge
+        run: |
+          conda install "ffmpeg=${{ matrix.ffmpeg-version }}" -c conda-forge -y
+          ffmpeg -version
+
+      - name: Install PaddlePaddle nightly in conda env
+        run: |
+          pip install --pre paddlepaddle -i https://www.paddlepaddle.org.cn/packages/nightly/cpu/
+
+      - name: Install paddlecodec from wheel
+        run: |
+          wheel_path=$(find dist -type f -name "*.whl")
+          echo "Installing $wheel_path"
+          pip install $wheel_path -vvv
+
+      - name: Install test dependencies
+        run: |
+          pip install numpy pytest pillow
+
+      - name: Delete src folder
+        run: |
+          # Delete src/ to ensure we're testing the installed wheel, not source code
+          rm -rf src/
+          ls -la
+
+      - name: Run tests
+        run: |
+          pytest --override-ini="addopts=-v" -s test_paddle
+
+  publish-pypi-arm:
+    runs-on: ubuntu-latest
+    name: Publish ARM Wheels to PyPI
+    if: "startsWith(github.ref, 'refs/tags/')"
+    needs:
+      - test-paddlecodec-wheel-arm
+    permissions:
+      id-token: write
+
+    steps:
+      - name: Retrieve ARM release distributions
+        uses: actions/download-artifact@v6
+        with:
+          pattern: paddlecodec-wheel-linux-arm64-*
+          path: dist/
+          merge-multiple: true
+
+      - name: Publish ARM release distributions to PyPI
+        uses: pypa/gh-action-pypi-publish@release/v1
+
+  publish-release-arm:
+    runs-on: ubuntu-latest
+    name: Publish ARM Wheels to GitHub
+    if: "startsWith(github.ref, 'refs/tags/')"
+    needs:
+      - test-paddlecodec-wheel-arm
+    permissions:
+      contents: write
+    steps:
+      - uses: actions/download-artifact@v6
+        with:
+          pattern: paddlecodec-wheel-linux-arm64-*
+          path: dist/
+          merge-multiple: true
+      - name: Publish ARM wheels to GitHub
+        uses: softprops/action-gh-release@v2
+        with:
+          draft: true
+          files: dist/*
+          tag_name: ${{ github.ref_name }}
\ No newline at end of file

From 17cdfffa354cbd2743f0d8cbf1fec671e6013e5c Mon Sep 17 00:00:00 2001
From: youge325 <A372707325@126.com>
Date: Fri, 10 Apr 2026 16:18:21 +0800
Subject: [PATCH 02/14] fix

---
 .github/workflows/paddle_arm_wheel.yaml | 29 ++++++++++++++++---------
 1 file changed, 19 insertions(+), 10 deletions(-)

diff --git a/.github/workflows/paddle_arm_wheel.yaml b/.github/workflows/paddle_arm_wheel.yaml
index a9f3f0805..3779c6511 100644
--- a/.github/workflows/paddle_arm_wheel.yaml
+++ b/.github/workflows/paddle_arm_wheel.yaml
@@ -20,6 +20,9 @@ defaults:
   run:
     shell: bash -l -eo pipefail {0}
 
+env:
+  ARM_BUILD_FFMPEG_VERSION: "7.0.1"
+
 jobs:
   build-paddlecodec-wheel-arm:
     # If your repo does not have GitHub-hosted ARM runners enabled,
@@ -59,10 +62,17 @@ jobs:
         run: |
           bash packaging/pre_build_script.sh
 
+      - name: Install build-time FFmpeg
+        run: |
+          conda install "ffmpeg=${{ env.ARM_BUILD_FFMPEG_VERSION }}" pkg-config -c conda-forge -y
+          ffmpeg -version
+          pkg-config --modversion libavcodec
+
       - name: Build wheel
         run: |
-          # Use pre-built FFmpeg from PyTorch S3
-          export BUILD_AGAINST_ALL_FFMPEG_FROM_S3=1
+          # Linux ARM does not have pre-built multi-version FFmpeg tarballs in the
+          # upstream S3 bucket, so build against the native conda FFmpeg install.
+          export I_CONFIRM_THIS_IS_NOT_A_LICENSE_VIOLATION=1
           export TORCHCODEC_CMAKE_BUILD_DIR=$(pwd)/build_cmake
           python -m build --wheel -vvv --no-isolation
 
@@ -76,15 +86,15 @@ jobs:
           unzip -j dist/*.whl "torchcodec/*.so" -d temp_libs || true
 
           # 2. Prepare LD_LIBRARY_PATH
-          # FFmpeg libraries
-          FFMPEG_LIB_PATHS=$(find $(pwd)/build_cmake/_deps -type d -name "lib" | tr '\n' ':')
+          # FFmpeg libraries from the active conda environment
+          FFMPEG_LIB_PATHS="${CONDA_PREFIX}/lib"
           # PaddlePaddle libraries
           PADDLE_PATH=$(python -c "import paddle; print(paddle.__path__[0])")
           PADDLE_LIB_PATHS="$PADDLE_PATH/base:$PADDLE_PATH/libs"
           # Wheel internal libraries
           INTERNAL_LIB_PATH=$(pwd)/temp_libs
 
-          export LD_LIBRARY_PATH=${FFMPEG_LIB_PATHS}${PADDLE_LIB_PATHS}:${INTERNAL_LIB_PATH}:${LD_LIBRARY_PATH}
+          export LD_LIBRARY_PATH=${FFMPEG_LIB_PATHS}:${PADDLE_LIB_PATHS}:${INTERNAL_LIB_PATH}:${LD_LIBRARY_PATH}
 
           # 3. Repair wheel with auditwheel
           # We exclude all external libraries because we want to rely on system libraries (like FFmpeg)
@@ -121,10 +131,9 @@ jobs:
       fail-fast: false
       matrix:
         python-version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
-        # FFmpeg 8.0 depends on libopenvino.so.2520, PaddlePaddle CPU depends on libopenvino.so.2500
-        # There has some conflict causing test failures, but it works with PaddlePaddle GPU.
-        # We skip FFmpeg 8.0 tests for PaddlePaddle CPU builds for now.
-        ffmpeg-version: ["4.4.2", "5.1.2", "6.1.1", "7.0.1"]
+        # ARM wheels currently build against a single native FFmpeg major because
+        # the pre-built multi-version Linux FFmpeg bundle is only published for x86_64.
+        ffmpeg-version: ["7.0.1"]
     steps:
       - name: Checkout repository
         uses: actions/checkout@v6
@@ -216,4 +225,4 @@ jobs:
         with:
           draft: true
           files: dist/*
-          tag_name: ${{ github.ref_name }}
\ No newline at end of file
+          tag_name: ${{ github.ref_name }}

From 27b73c4337bf02eeaa696c7f371900f2f68c551a Mon Sep 17 00:00:00 2001
From: youge325 <A372707325@126.com>
Date: Fri, 10 Apr 2026 17:30:57 +0800
Subject: [PATCH 03/14] try to use pytorch image

---
 .github/workflows/paddle_arm_wheel.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/.github/workflows/paddle_arm_wheel.yaml b/.github/workflows/paddle_arm_wheel.yaml
index 3779c6511..b22dd181d 100644
--- a/.github/workflows/paddle_arm_wheel.yaml
+++ b/.github/workflows/paddle_arm_wheel.yaml
@@ -29,7 +29,7 @@ jobs:
     # replace this with self-hosted labels, e.g. [self-hosted, Linux, ARM64].
     runs-on: ubuntu-24.04-arm
     container:
-      image: quay.io/pypa/manylinux_2_28_aarch64
+      image: pytorch/manylinux2_28_aarch64-builder:cpu-aarch64
     strategy:
       fail-fast: false
       matrix:

From c9cbfb350842c12f06e482716e7f63cb13fffeac Mon Sep 17 00:00:00 2001
From: youge325 <A372707325@126.com>
Date: Fri, 10 Apr 2026 20:11:15 +0800
Subject: [PATCH 04/14] try to fix network problem

---
 .github/workflows/paddle_arm_wheel.yaml |  3 ++
 test_paddle/test_video_decode.py        | 56 ++++++++++++++++++++-----
 2 files changed, 49 insertions(+), 10 deletions(-)

diff --git a/.github/workflows/paddle_arm_wheel.yaml b/.github/workflows/paddle_arm_wheel.yaml
index b22dd181d..dcd29157b 100644
--- a/.github/workflows/paddle_arm_wheel.yaml
+++ b/.github/workflows/paddle_arm_wheel.yaml
@@ -127,6 +127,9 @@ jobs:
   test-paddlecodec-wheel-arm:
     needs: build-paddlecodec-wheel-arm
     runs-on: ubuntu-24.04-arm
+    env:
+      PADDLECODEC_TEST_VIDEO: test/resources/testsrc2.mp4
+      PADDLECODEC_TEST_NUM_FRAMES: "8"
     strategy:
       fail-fast: false
       matrix:
diff --git a/test_paddle/test_video_decode.py b/test_paddle/test_video_decode.py
index 4e41cb19e..62e8ff593 100644
--- a/test_paddle/test_video_decode.py
+++ b/test_paddle/test_video_decode.py
@@ -10,6 +10,8 @@
 import httpx
 import numpy as np
 
+DEFAULT_TEST_VIDEO = "https://paddlenlp.bj.bcebos.com/datasets/paddlemix/demo_video/example_video.mp4"
+
 
 @dataclass
 class VideoMetadata(Mapping):
@@ -168,14 +170,48 @@ def sample_indices_fn_func(metadata, **fn_kwargs):
     return video, metadata
 
 
+def _get_optional_int_env(name: str) -> Optional[int]:
+    value = os.getenv(name)
+    return int(value) if value else None
+
+
+def _assert_video_decode_smoke(
+    video,
+    metadata: VideoMetadata,
+    expected_num_frames: Optional[int] = None,
+):
+    assert video.ndim == 4
+    assert len(metadata.frames_indices) == video.shape[0]
+    if expected_num_frames is not None:
+        assert video.shape[0] == expected_num_frames
+    assert metadata.total_num_frames >= video.shape[0] >= 1
+    assert metadata.fps is not None and metadata.fps > 0
+    assert metadata.width is not None and metadata.width > 0
+    assert metadata.height is not None and metadata.height > 0
+    assert metadata.duration is not None and metadata.duration > 0
+    assert tuple(video.shape[-2:]) == (metadata.height, metadata.width)
+    assert video.shape[-3] == 3
+    assert video.to(paddle.int64).sum().item() > 0
+
+
 def test_video_decode():
-    url = "https://paddlenlp.bj.bcebos.com/datasets/paddlemix/demo_video/example_video.mp4"
-    video, metadata = load_video(url, backend="torchcodec")
-    assert video.to(paddle.int64).sum().item() == 247759890390
-    assert metadata.total_num_frames == 263
-    assert metadata.fps == pytest.approx(29.99418249715141)
-    assert metadata.width == 1920
-    assert metadata.height == 1080
-    assert metadata.duration == pytest.approx(8.768367)
-    for i, idx in enumerate(metadata.frames_indices):
-        assert idx == i
+    video_path = os.getenv("PADDLECODEC_TEST_VIDEO", DEFAULT_TEST_VIDEO)
+    num_frames = _get_optional_int_env("PADDLECODEC_TEST_NUM_FRAMES")
+
+    video, metadata = load_video(
+        video_path,
+        backend="torchcodec",
+        num_frames=num_frames,
+    )
+
+    if video_path == DEFAULT_TEST_VIDEO and num_frames is None:
+        assert video.to(paddle.int64).sum().item() == 247759890390
+        assert metadata.total_num_frames == 263
+        assert metadata.fps == pytest.approx(29.99418249715141)
+        assert metadata.width == 1920
+        assert metadata.height == 1080
+        assert metadata.duration == pytest.approx(8.768367)
+        for i, idx in enumerate(metadata.frames_indices):
+            assert idx == i
+    else:
+        _assert_video_decode_smoke(video, metadata, expected_num_frames=num_frames)

From 516c690aaacc6a1e8ea8c0d10fe85fa730450b33 Mon Sep 17 00:00:00 2001
From: youge325 <A372707325@126.com>
Date: Sat, 11 Apr 2026 11:27:46 +0800
Subject: [PATCH 05/14] update ffmpeg and wheel build workflow with upstream
 main branch

---
 .github/workflows/build_ffmpeg.yaml           | 27 +++++++
 .github/workflows/paddle_arm_wheel.yaml       | 62 ++++++++-------
 ...fetch_and_expose_non_gpl_ffmpeg_libs.cmake | 76 +++++++++++++------
 3 files changed, 113 insertions(+), 52 deletions(-)

diff --git a/.github/workflows/build_ffmpeg.yaml b/.github/workflows/build_ffmpeg.yaml
index 9b9317e3b..5494bf0b1 100644
--- a/.github/workflows/build_ffmpeg.yaml
+++ b/.github/workflows/build_ffmpeg.yaml
@@ -48,6 +48,33 @@ jobs:
         mkdir -p "${artifact_dir}"
         mv ffmpeg.tar.gz "${artifact_dir}/${FFMPEG_VERSION}.tar.gz"
 
+  LGPL-Linux-aarch64:
+    strategy:
+      fail-fast: false
+      matrix:
+        ffmpeg-version: ["4.4.4", "5.1.4", "6.1.1", "7.0.1", "8.0"]
+    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
+    permissions:
+      id-token: write
+      contents: read
+    with:
+      job-name: Build
+      upload-artifact: ffmpeg-lgpl-linux_aarch64-${{ matrix.ffmpeg-version }}
+      repository: meta-pytorch/torchcodec
+      runner: linux.arm64.2xlarge
+      docker-image: pytorch/manylinux2_28_aarch64-builder:cpu-aarch64
+      script: |
+        export FFMPEG_VERSION="${{ matrix.ffmpeg-version }}"
+        export FFMPEG_ROOT="${PWD}/ffmpeg"
+
+        packaging/build_ffmpeg.sh
+
+        tar -cf ffmpeg.tar.gz ffmpeg/include ffmpeg/lib
+
+        artifact_dir="${RUNNER_ARTIFACT_DIR}/$(date +%Y-%m-%d)/linux_aarch64"
+        mkdir -p "${artifact_dir}"
+        mv ffmpeg.tar.gz "${artifact_dir}/${FFMPEG_VERSION}.tar.gz"
+
   LGPL-macOS:
     strategy:
       fail-fast: false
diff --git a/.github/workflows/paddle_arm_wheel.yaml b/.github/workflows/paddle_arm_wheel.yaml
index dcd29157b..319823357 100644
--- a/.github/workflows/paddle_arm_wheel.yaml
+++ b/.github/workflows/paddle_arm_wheel.yaml
@@ -21,19 +21,36 @@ defaults:
     shell: bash -l -eo pipefail {0}
 
 env:
-  ARM_BUILD_FFMPEG_VERSION: "7.0.1"
+  ARM_TEST_PYTHON_VERSION: "3.10"
 
 jobs:
-  build-paddlecodec-wheel-arm:
+  generate-matrix:
+    runs-on: ubuntu-latest
+    outputs:
+      build-matrix: ${{ steps.set-matrix.outputs.build-matrix }}
+      test-matrix: ${{ steps.set-matrix.outputs.test-matrix }}
+    steps:
+      - name: Generate ARM build and test matrix
+        id: set-matrix
+        run: |
+          printf '%s\n' \
+            'build-matrix={"python-version":["3.9","3.10","3.11","3.12","3.13"]}' \
+            >> "$GITHUB_OUTPUT"
+          printf '%s\n' \
+            'test-matrix={"include":[{"python-version":"'"${ARM_TEST_PYTHON_VERSION}"'","ffmpeg-version":"4.4.2"},{"python-version":"'"${ARM_TEST_PYTHON_VERSION}"'","ffmpeg-version":"5.1.2"},{"python-version":"'"${ARM_TEST_PYTHON_VERSION}"'","ffmpeg-version":"6.1.1"},{"python-version":"'"${ARM_TEST_PYTHON_VERSION}"'","ffmpeg-version":"7.0.1"}]}' \
+            >> "$GITHUB_OUTPUT"
+
+  build:
+    needs: generate-matrix
     # If your repo does not have GitHub-hosted ARM runners enabled,
     # replace this with self-hosted labels, e.g. [self-hosted, Linux, ARM64].
+    name: Build and Upload ARM wheel
     runs-on: ubuntu-24.04-arm
     container:
       image: pytorch/manylinux2_28_aarch64-builder:cpu-aarch64
     strategy:
       fail-fast: false
-      matrix:
-        python-version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
+      matrix: ${{ fromJson(needs.generate-matrix.outputs.build-matrix) }}
     permissions:
       id-token: write
       contents: read
@@ -62,16 +79,11 @@ jobs:
         run: |
           bash packaging/pre_build_script.sh
 
-      - name: Install build-time FFmpeg
-        run: |
-          conda install "ffmpeg=${{ env.ARM_BUILD_FFMPEG_VERSION }}" pkg-config -c conda-forge -y
-          ffmpeg -version
-          pkg-config --modversion libavcodec
-
       - name: Build wheel
         run: |
-          # Linux ARM does not have pre-built multi-version FFmpeg tarballs in the
-          # upstream S3 bucket, so build against the native conda FFmpeg install.
+          # Match upstream's Linux workflow and build the ARM wheel against the
+          # pre-built non-GPL FFmpeg bundles published on S3.
+          export BUILD_AGAINST_ALL_FFMPEG_FROM_S3=1
           export I_CONFIRM_THIS_IS_NOT_A_LICENSE_VIOLATION=1
           export TORCHCODEC_CMAKE_BUILD_DIR=$(pwd)/build_cmake
           python -m build --wheel -vvv --no-isolation
@@ -86,8 +98,8 @@ jobs:
           unzip -j dist/*.whl "torchcodec/*.so" -d temp_libs || true
 
           # 2. Prepare LD_LIBRARY_PATH
-          # FFmpeg libraries from the active conda environment
-          FFMPEG_LIB_PATHS="${CONDA_PREFIX}/lib"
+          # FFmpeg libraries fetched from the build-time S3 bundles
+          FFMPEG_LIB_PATHS=$(find "$(pwd)/build_cmake/_deps" -type d -name "lib" -print | paste -sd: -)
           # PaddlePaddle libraries
           PADDLE_PATH=$(python -c "import paddle; print(paddle.__path__[0])")
           PADDLE_LIB_PATHS="$PADDLE_PATH/base:$PADDLE_PATH/libs"
@@ -124,28 +136,22 @@ jobs:
           echo "Wheel path: $wheel_path"
           unzip -l $wheel_path
 
-  test-paddlecodec-wheel-arm:
-    needs: build-paddlecodec-wheel-arm
+  install-and-test:
+    needs: [generate-matrix, build]
+    name: Install and test ARM wheel
     runs-on: ubuntu-24.04-arm
     env:
       PADDLECODEC_TEST_VIDEO: test/resources/testsrc2.mp4
       PADDLECODEC_TEST_NUM_FRAMES: "8"
     strategy:
       fail-fast: false
-      matrix:
-        python-version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
-        # ARM wheels currently build against a single native FFmpeg major because
-        # the pre-built multi-version Linux FFmpeg bundle is only published for x86_64.
-        ffmpeg-version: ["7.0.1"]
+      # Match upstream's staged Linux wheel flow: build all Python variants, then
+      # smoke-test one representative ARM wheel across multiple runtime FFmpeg majors.
+      matrix: ${{ fromJson(needs.generate-matrix.outputs.test-matrix) }}
     steps:
       - name: Checkout repository
         uses: actions/checkout@v6
 
-      - name: Set up Python ${{ matrix.python-version }}
-        uses: actions/setup-python@v5
-        with:
-          python-version: ${{ matrix.python-version }}
-
       - name: Download wheel artifact
         uses: actions/download-artifact@v4
         with:
@@ -194,7 +200,7 @@ jobs:
     name: Publish ARM Wheels to PyPI
     if: "startsWith(github.ref, 'refs/tags/')"
     needs:
-      - test-paddlecodec-wheel-arm
+      - install-and-test
     permissions:
       id-token: write
 
@@ -214,7 +220,7 @@ jobs:
     name: Publish ARM Wheels to GitHub
     if: "startsWith(github.ref, 'refs/tags/')"
     needs:
-      - test-paddlecodec-wheel-arm
+      - install-and-test
     permissions:
       contents: write
     steps:
diff --git a/src/torchcodec/_core/fetch_and_expose_non_gpl_ffmpeg_libs.cmake b/src/torchcodec/_core/fetch_and_expose_non_gpl_ffmpeg_libs.cmake
index 07abd2e87..781fce5a0 100644
--- a/src/torchcodec/_core/fetch_and_expose_non_gpl_ffmpeg_libs.cmake
+++ b/src/torchcodec/_core/fetch_and_expose_non_gpl_ffmpeg_libs.cmake
@@ -23,31 +23,59 @@ set(
 if (LINUX)
     set(lib_dir "lib")
 
-    set(
-        platform_url
-        ${base_url}/linux_x86_64
-    )
+    if (CMAKE_SYSTEM_PROCESSOR MATCHES "aarch64|arm64|ARM64")
+        set(
+            platform_url
+            ${base_url}/linux_aarch64
+        )
 
-    set(
-        f4_sha256
-        1a083f1922443bedb5243d04896383b8c606778a7ddb9d886c8303e55339fe0c
-    )
-    set(
-        f5_sha256
-        65d6ad54082d94dcb3f801d73df2265e0e1bb303c7afbce7723e3b77ccd0e207
-    )
-    set(
-        f6_sha256
-        8bd5939c2f4a4b072e837e7870c13fe7d13824e5ff087ab534e4db4e90b7be9c
-    )
-    set(
-        f7_sha256
-        1cb946d8b7c6393c2c3ebe1f900b8de7a2885fe614c45d4ec32c9833084f2f26
-    )
-    set(
-        f8_sha256
-        c55b3c1a4b5e4d5fdd7c632bea3ab6f45b4e37cc8e0999dda3f84a8ed8defad8
-    )
+        set(
+            f4_sha256
+            a310a2ed9ffe555fd3278dae15065541098dd35e124564671dcda6a6620ac842
+        )
+        set(
+            f5_sha256
+            89ca7996bccbc2db49adaa401d20fdbabffe0e1b4e07a0f81d6b143e858b7c8d
+        )
+        set(
+            f6_sha256
+            ae44c67b4587d061b8e9cc8990ca891ee013fe52ad79e5016ba29871562621da
+        )
+        set(
+            f7_sha256
+            948e2cac66ca6f68ff526d5e84138e94bce0f1a7c83f502d15d85d0bd3ddc112
+        )
+        set(
+            f8_sha256
+            b9cfd99ae75a14e58300854967d4dc49de0b3daa551df51ea1f52a3f08d2c8af
+        )
+    else()
+        set(
+            platform_url
+            ${base_url}/linux_x86_64
+        )
+
+        set(
+            f4_sha256
+            1a083f1922443bedb5243d04896383b8c606778a7ddb9d886c8303e55339fe0c
+        )
+        set(
+            f5_sha256
+            65d6ad54082d94dcb3f801d73df2265e0e1bb303c7afbce7723e3b77ccd0e207
+        )
+        set(
+            f6_sha256
+            8bd5939c2f4a4b072e837e7870c13fe7d13824e5ff087ab534e4db4e90b7be9c
+        )
+        set(
+            f7_sha256
+            1cb946d8b7c6393c2c3ebe1f900b8de7a2885fe614c45d4ec32c9833084f2f26
+        )
+        set(
+            f8_sha256
+            c55b3c1a4b5e4d5fdd7c632bea3ab6f45b4e37cc8e0999dda3f84a8ed8defad8
+        )
+    endif()
     set(
        f4_library_file_names
        libavutil.so.56

From b9052e703aead49d59b9e51e95c237fa852631ef Mon Sep 17 00:00:00 2001
From: youge325 <A372707325@126.com>
Date: Sat, 11 Apr 2026 12:04:01 +0800
Subject: [PATCH 06/14] run full tests

---
 .github/workflows/paddle_arm_wheel.yaml |  3 --
 test_paddle/test_video_decode.py        | 58 +++++--------------------
 2 files changed, 10 insertions(+), 51 deletions(-)

diff --git a/.github/workflows/paddle_arm_wheel.yaml b/.github/workflows/paddle_arm_wheel.yaml
index 319823357..3a3c532f3 100644
--- a/.github/workflows/paddle_arm_wheel.yaml
+++ b/.github/workflows/paddle_arm_wheel.yaml
@@ -140,9 +140,6 @@ jobs:
     needs: [generate-matrix, build]
     name: Install and test ARM wheel
     runs-on: ubuntu-24.04-arm
-    env:
-      PADDLECODEC_TEST_VIDEO: test/resources/testsrc2.mp4
-      PADDLECODEC_TEST_NUM_FRAMES: "8"
     strategy:
       fail-fast: false
       # Match upstream's staged Linux wheel flow: build all Python variants, then
diff --git a/test_paddle/test_video_decode.py b/test_paddle/test_video_decode.py
index 62e8ff593..f708b464e 100644
--- a/test_paddle/test_video_decode.py
+++ b/test_paddle/test_video_decode.py
@@ -10,8 +10,6 @@
 import httpx
 import numpy as np
 
-DEFAULT_TEST_VIDEO = "https://paddlenlp.bj.bcebos.com/datasets/paddlemix/demo_video/example_video.mp4"
-
 
 @dataclass
 class VideoMetadata(Mapping):
@@ -168,50 +166,14 @@ def sample_indices_fn_func(metadata, **fn_kwargs):
     video_decoder = VIDEO_DECODERS[backend]
     video, metadata = video_decoder(file_obj, sample_indices_fn, **kwargs)
     return video, metadata
-
-
-def _get_optional_int_env(name: str) -> Optional[int]:
-    value = os.getenv(name)
-    return int(value) if value else None
-
-
-def _assert_video_decode_smoke(
-    video,
-    metadata: VideoMetadata,
-    expected_num_frames: Optional[int] = None,
-):
-    assert video.ndim == 4
-    assert len(metadata.frames_indices) == video.shape[0]
-    if expected_num_frames is not None:
-        assert video.shape[0] == expected_num_frames
-    assert metadata.total_num_frames >= video.shape[0] >= 1
-    assert metadata.fps is not None and metadata.fps > 0
-    assert metadata.width is not None and metadata.width > 0
-    assert metadata.height is not None and metadata.height > 0
-    assert metadata.duration is not None and metadata.duration > 0
-    assert tuple(video.shape[-2:]) == (metadata.height, metadata.width)
-    assert video.shape[-3] == 3
-    assert video.to(paddle.int64).sum().item() > 0
-
-
 def test_video_decode():
-    video_path = os.getenv("PADDLECODEC_TEST_VIDEO", DEFAULT_TEST_VIDEO)
-    num_frames = _get_optional_int_env("PADDLECODEC_TEST_NUM_FRAMES")
-
-    video, metadata = load_video(
-        video_path,
-        backend="torchcodec",
-        num_frames=num_frames,
-    )
-
-    if video_path == DEFAULT_TEST_VIDEO and num_frames is None:
-        assert video.to(paddle.int64).sum().item() == 247759890390
-        assert metadata.total_num_frames == 263
-        assert metadata.fps == pytest.approx(29.99418249715141)
-        assert metadata.width == 1920
-        assert metadata.height == 1080
-        assert metadata.duration == pytest.approx(8.768367)
-        for i, idx in enumerate(metadata.frames_indices):
-            assert idx == i
-    else:
-        _assert_video_decode_smoke(video, metadata, expected_num_frames=num_frames)
+    url = "https://paddlenlp.bj.bcebos.com/datasets/paddlemix/demo_video/example_video.mp4"
+    video, metadata = load_video(url, backend="torchcodec")
+    assert video.to(paddle.int64).sum().item() == 247759890390
+    assert metadata.total_num_frames == 263
+    assert metadata.fps == pytest.approx(29.99418249715141)
+    assert metadata.width == 1920
+    assert metadata.height == 1080
+    assert metadata.duration == pytest.approx(8.768367)
+    for i, idx in enumerate(metadata.frames_indices):
+        assert idx == i

From cb68f641c72edb053683e4ac700b0c5d126f0173 Mon Sep 17 00:00:00 2001
From: youge325 <A372707325@126.com>
Date: Sat, 11 Apr 2026 13:21:42 +0800
Subject: [PATCH 07/14] Revert "run full tests"

This reverts commit b9052e703aead49d59b9e51e95c237fa852631ef.
---
 .github/workflows/paddle_arm_wheel.yaml |  3 ++
 test_paddle/test_video_decode.py        | 58 ++++++++++++++++++++-----
 2 files changed, 51 insertions(+), 10 deletions(-)

diff --git a/.github/workflows/paddle_arm_wheel.yaml b/.github/workflows/paddle_arm_wheel.yaml
index 3a3c532f3..319823357 100644
--- a/.github/workflows/paddle_arm_wheel.yaml
+++ b/.github/workflows/paddle_arm_wheel.yaml
@@ -140,6 +140,9 @@ jobs:
     needs: [generate-matrix, build]
     name: Install and test ARM wheel
     runs-on: ubuntu-24.04-arm
+    env:
+      PADDLECODEC_TEST_VIDEO: test/resources/testsrc2.mp4
+      PADDLECODEC_TEST_NUM_FRAMES: "8"
     strategy:
       fail-fast: false
       # Match upstream's staged Linux wheel flow: build all Python variants, then
diff --git a/test_paddle/test_video_decode.py b/test_paddle/test_video_decode.py
index f708b464e..62e8ff593 100644
--- a/test_paddle/test_video_decode.py
+++ b/test_paddle/test_video_decode.py
@@ -10,6 +10,8 @@
 import httpx
 import numpy as np
 
+DEFAULT_TEST_VIDEO = "https://paddlenlp.bj.bcebos.com/datasets/paddlemix/demo_video/example_video.mp4"
+
 
 @dataclass
 class VideoMetadata(Mapping):
@@ -166,14 +168,50 @@ def sample_indices_fn_func(metadata, **fn_kwargs):
     video_decoder = VIDEO_DECODERS[backend]
     video, metadata = video_decoder(file_obj, sample_indices_fn, **kwargs)
     return video, metadata
+
+
+def _get_optional_int_env(name: str) -> Optional[int]:
+    value = os.getenv(name)
+    return int(value) if value else None
+
+
+def _assert_video_decode_smoke(
+    video,
+    metadata: VideoMetadata,
+    expected_num_frames: Optional[int] = None,
+):
+    assert video.ndim == 4
+    assert len(metadata.frames_indices) == video.shape[0]
+    if expected_num_frames is not None:
+        assert video.shape[0] == expected_num_frames
+    assert metadata.total_num_frames >= video.shape[0] >= 1
+    assert metadata.fps is not None and metadata.fps > 0
+    assert metadata.width is not None and metadata.width > 0
+    assert metadata.height is not None and metadata.height > 0
+    assert metadata.duration is not None and metadata.duration > 0
+    assert tuple(video.shape[-2:]) == (metadata.height, metadata.width)
+    assert video.shape[-3] == 3
+    assert video.to(paddle.int64).sum().item() > 0
+
+
 def test_video_decode():
-    url = "https://paddlenlp.bj.bcebos.com/datasets/paddlemix/demo_video/example_video.mp4"
-    video, metadata = load_video(url, backend="torchcodec")
-    assert video.to(paddle.int64).sum().item() == 247759890390
-    assert metadata.total_num_frames == 263
-    assert metadata.fps == pytest.approx(29.99418249715141)
-    assert metadata.width == 1920
-    assert metadata.height == 1080
-    assert metadata.duration == pytest.approx(8.768367)
-    for i, idx in enumerate(metadata.frames_indices):
-        assert idx == i
+    video_path = os.getenv("PADDLECODEC_TEST_VIDEO", DEFAULT_TEST_VIDEO)
+    num_frames = _get_optional_int_env("PADDLECODEC_TEST_NUM_FRAMES")
+
+    video, metadata = load_video(
+        video_path,
+        backend="torchcodec",
+        num_frames=num_frames,
+    )
+
+    if video_path == DEFAULT_TEST_VIDEO and num_frames is None:
+        assert video.to(paddle.int64).sum().item() == 247759890390
+        assert metadata.total_num_frames == 263
+        assert metadata.fps == pytest.approx(29.99418249715141)
+        assert metadata.width == 1920
+        assert metadata.height == 1080
+        assert metadata.duration == pytest.approx(8.768367)
+        for i, idx in enumerate(metadata.frames_indices):
+            assert idx == i
+    else:
+        _assert_video_decode_smoke(video, metadata, expected_num_frames=num_frames)

From 1d3ea2f40070cb04b3545496bfe2acc64f13dee6 Mon Sep 17 00:00:00 2001
From: youge325 <A372707325@126.com>
Date: Sat, 11 Apr 2026 17:16:19 +0800
Subject: [PATCH 08/14] add multi-python multi-ffmpeg build workflow

---
 .github/workflows/paddle_arm_wheel.yaml | 40 ++++++++++++++++++-------
 1 file changed, 29 insertions(+), 11 deletions(-)

diff --git a/.github/workflows/paddle_arm_wheel.yaml b/.github/workflows/paddle_arm_wheel.yaml
index 319823357..849eed992 100644
--- a/.github/workflows/paddle_arm_wheel.yaml
+++ b/.github/workflows/paddle_arm_wheel.yaml
@@ -20,9 +20,6 @@ defaults:
   run:
     shell: bash -l -eo pipefail {0}
 
-env:
-  ARM_TEST_PYTHON_VERSION: "3.10"
-
 jobs:
   generate-matrix:
     runs-on: ubuntu-latest
@@ -33,12 +30,33 @@ jobs:
       - name: Generate ARM build and test matrix
         id: set-matrix
         run: |
-          printf '%s\n' \
-            'build-matrix={"python-version":["3.9","3.10","3.11","3.12","3.13"]}' \
-            >> "$GITHUB_OUTPUT"
-          printf '%s\n' \
-            'test-matrix={"include":[{"python-version":"'"${ARM_TEST_PYTHON_VERSION}"'","ffmpeg-version":"4.4.2"},{"python-version":"'"${ARM_TEST_PYTHON_VERSION}"'","ffmpeg-version":"5.1.2"},{"python-version":"'"${ARM_TEST_PYTHON_VERSION}"'","ffmpeg-version":"6.1.1"},{"python-version":"'"${ARM_TEST_PYTHON_VERSION}"'","ffmpeg-version":"7.0.1"}]}' \
-            >> "$GITHUB_OUTPUT"
+          python - <<'PY'
+          import itertools
+          import json
+          import os
+
+          python_versions = ["3.9", "3.10", "3.11", "3.12", "3.13"]
+          # FFmpeg 8.0 is intentionally excluded for Paddle CPU builds to match
+          # the existing x86 workflow and avoid the current OpenVINO conflict.
+          ffmpeg_versions = ["4.4.2", "5.1.2", "6.1.1", "7.0.1"]
+
+          build_matrix = {"python-version": python_versions}
+          test_matrix = {
+              "include": [
+                  {
+                      "python-version": python_version,
+                      "ffmpeg-version": ffmpeg_version,
+                  }
+                  for python_version, ffmpeg_version in itertools.product(
+                      python_versions, ffmpeg_versions
+                  )
+              ]
+          }
+
+          with open(os.environ["GITHUB_OUTPUT"], "a", encoding="utf-8") as f:
+              f.write(f"build-matrix={json.dumps(build_matrix)}\n")
+              f.write(f"test-matrix={json.dumps(test_matrix)}\n")
+          PY
 
   build:
     needs: generate-matrix
@@ -145,8 +163,8 @@ jobs:
       PADDLECODEC_TEST_NUM_FRAMES: "8"
     strategy:
       fail-fast: false
-      # Match upstream's staged Linux wheel flow: build all Python variants, then
-      # smoke-test one representative ARM wheel across multiple runtime FFmpeg majors.
+      # Match the existing Paddle x86 flow while keeping the upstream-style staged
+      # build/test split: exercise every Python and FFmpeg combination post-build.
       matrix: ${{ fromJson(needs.generate-matrix.outputs.test-matrix) }}
     steps:
       - name: Checkout repository

From 3cb2a0fc4f4dc76d7a9bd822b14b607bd523ef58 Mon Sep 17 00:00:00 2001
From: youge325 <A372707325@126.com>
Date: Sat, 11 Apr 2026 19:57:49 +0800
Subject: [PATCH 09/14] try to use actions/cache@v4

---
 .github/workflows/paddle_arm_wheel.yaml | 42 ++++++++++++++++--
 .github/workflows/paddle_wheel.yaml     | 41 +++++++++++++++++-
 test_paddle/test_video_decode.py        | 57 ++++++-------------------
 3 files changed, 91 insertions(+), 49 deletions(-)

diff --git a/.github/workflows/paddle_arm_wheel.yaml b/.github/workflows/paddle_arm_wheel.yaml
index 849eed992..dd3356a97 100644
--- a/.github/workflows/paddle_arm_wheel.yaml
+++ b/.github/workflows/paddle_arm_wheel.yaml
@@ -20,6 +20,11 @@ defaults:
   run:
     shell: bash -l -eo pipefail {0}
 
+env:
+  PADDLECODEC_TEST_VIDEO_URL: https://paddlenlp.bj.bcebos.com/datasets/paddlemix/demo_video/example_video.mp4
+  PADDLECODEC_TEST_VIDEO_CACHE_KEY: paddlecodec-test-video-v1-example-video
+  PADDLECODEC_TEST_VIDEO_PATH: .github/test-assets/example_video.mp4
+
 jobs:
   generate-matrix:
     runs-on: ubuntu-latest
@@ -58,6 +63,32 @@ jobs:
               f.write(f"test-matrix={json.dumps(test_matrix)}\n")
           PY
 
+  prepare-test-video:
+    runs-on: ubuntu-latest
+    name: Prepare cached Paddle test video
+    steps:
+      - name: Restore cached test video
+        id: cache-test-video
+        uses: actions/cache@v4
+        with:
+          path: ${{ env.PADDLECODEC_TEST_VIDEO_PATH }}
+          key: ${{ env.PADDLECODEC_TEST_VIDEO_CACHE_KEY }}
+
+      - name: Download test video
+        if: steps.cache-test-video.outputs.cache-hit != 'true'
+        run: |
+          mkdir -p "$(dirname "${PADDLECODEC_TEST_VIDEO_PATH}")"
+          curl --fail --location --retry 5 --retry-all-errors \
+            --output "${PADDLECODEC_TEST_VIDEO_PATH}" \
+            "${PADDLECODEC_TEST_VIDEO_URL}"
+
+      - name: Upload cached test video artifact
+        uses: actions/upload-artifact@v5
+        with:
+          name: paddlecodec-test-video
+          path: ${{ env.PADDLECODEC_TEST_VIDEO_PATH }}
+          if-no-files-found: error
+
   build:
     needs: generate-matrix
     # If your repo does not have GitHub-hosted ARM runners enabled,
@@ -155,12 +186,11 @@ jobs:
           unzip -l $wheel_path
 
   install-and-test:
-    needs: [generate-matrix, build]
+    needs: [generate-matrix, prepare-test-video, build]
     name: Install and test ARM wheel
     runs-on: ubuntu-24.04-arm
     env:
-      PADDLECODEC_TEST_VIDEO: test/resources/testsrc2.mp4
-      PADDLECODEC_TEST_NUM_FRAMES: "8"
+      PADDLECODEC_TEST_VIDEO: .github/test-assets/example_video.mp4
     strategy:
       fail-fast: false
       # Match the existing Paddle x86 flow while keeping the upstream-style staged
@@ -176,6 +206,12 @@ jobs:
           name: paddlecodec-wheel-linux-arm64-py${{ matrix.python-version }}
           path: dist/
 
+      - name: Download cached test video artifact
+        uses: actions/download-artifact@v4
+        with:
+          name: paddlecodec-test-video
+          path: .github/test-assets/
+
       - name: Install FFmpeg via conda
         uses: conda-incubator/setup-miniconda@v3
         with:
diff --git a/.github/workflows/paddle_wheel.yaml b/.github/workflows/paddle_wheel.yaml
index 4976e20c9..86dd9c0e0 100644
--- a/.github/workflows/paddle_wheel.yaml
+++ b/.github/workflows/paddle_wheel.yaml
@@ -20,7 +20,38 @@ defaults:
   run:
     shell: bash -l -eo pipefail {0}
 
+env:
+  PADDLECODEC_TEST_VIDEO_URL: https://paddlenlp.bj.bcebos.com/datasets/paddlemix/demo_video/example_video.mp4
+  PADDLECODEC_TEST_VIDEO_CACHE_KEY: paddlecodec-test-video-v1-example-video
+  PADDLECODEC_TEST_VIDEO_PATH: .github/test-assets/example_video.mp4
+
 jobs:
+  prepare-test-video:
+    runs-on: ubuntu-latest
+    name: Prepare cached Paddle test video
+    steps:
+      - name: Restore cached test video
+        id: cache-test-video
+        uses: actions/cache@v4
+        with:
+          path: ${{ env.PADDLECODEC_TEST_VIDEO_PATH }}
+          key: ${{ env.PADDLECODEC_TEST_VIDEO_CACHE_KEY }}
+
+      - name: Download test video
+        if: steps.cache-test-video.outputs.cache-hit != 'true'
+        run: |
+          mkdir -p "$(dirname "${PADDLECODEC_TEST_VIDEO_PATH}")"
+          curl --fail --location --retry 5 --retry-all-errors \
+            --output "${PADDLECODEC_TEST_VIDEO_PATH}" \
+            "${PADDLECODEC_TEST_VIDEO_URL}"
+
+      - name: Upload cached test video artifact
+        uses: actions/upload-artifact@v5
+        with:
+          name: paddlecodec-test-video
+          path: ${{ env.PADDLECODEC_TEST_VIDEO_PATH }}
+          if-no-files-found: error
+
   build-paddlecodec-wheel:
     runs-on: ubuntu-latest
     container:
@@ -113,8 +144,10 @@ jobs:
           unzip -l $wheel_path
 
   test-paddlecodec-wheel:
-    needs: build-paddlecodec-wheel
+    needs: [prepare-test-video, build-paddlecodec-wheel]
     runs-on: ubuntu-latest
+    env:
+      PADDLECODEC_TEST_VIDEO: .github/test-assets/example_video.mp4
     strategy:
       fail-fast: false
       matrix:
@@ -138,6 +171,12 @@ jobs:
           name: paddlecodec-wheel-linux-py${{ matrix.python-version }}
           path: dist/
 
+      - name: Download cached test video artifact
+        uses: actions/download-artifact@v4
+        with:
+          name: paddlecodec-test-video
+          path: .github/test-assets/
+
       - name: Install FFmpeg via conda
         uses: conda-incubator/setup-miniconda@v3
         with:
diff --git a/test_paddle/test_video_decode.py b/test_paddle/test_video_decode.py
index 62e8ff593..1be61bd3e 100644
--- a/test_paddle/test_video_decode.py
+++ b/test_paddle/test_video_decode.py
@@ -10,8 +10,6 @@
 import httpx
 import numpy as np
 
-DEFAULT_TEST_VIDEO = "https://paddlenlp.bj.bcebos.com/datasets/paddlemix/demo_video/example_video.mp4"
-
 
 @dataclass
 class VideoMetadata(Mapping):
@@ -170,48 +168,17 @@ def sample_indices_fn_func(metadata, **fn_kwargs):
     return video, metadata
 
 
-def _get_optional_int_env(name: str) -> Optional[int]:
-    value = os.getenv(name)
-    return int(value) if value else None
-
-
-def _assert_video_decode_smoke(
-    video,
-    metadata: VideoMetadata,
-    expected_num_frames: Optional[int] = None,
-):
-    assert video.ndim == 4
-    assert len(metadata.frames_indices) == video.shape[0]
-    if expected_num_frames is not None:
-        assert video.shape[0] == expected_num_frames
-    assert metadata.total_num_frames >= video.shape[0] >= 1
-    assert metadata.fps is not None and metadata.fps > 0
-    assert metadata.width is not None and metadata.width > 0
-    assert metadata.height is not None and metadata.height > 0
-    assert metadata.duration is not None and metadata.duration > 0
-    assert tuple(video.shape[-2:]) == (metadata.height, metadata.width)
-    assert video.shape[-3] == 3
-    assert video.to(paddle.int64).sum().item() > 0
-
-
 def test_video_decode():
-    video_path = os.getenv("PADDLECODEC_TEST_VIDEO", DEFAULT_TEST_VIDEO)
-    num_frames = _get_optional_int_env("PADDLECODEC_TEST_NUM_FRAMES")
-
-    video, metadata = load_video(
-        video_path,
-        backend="torchcodec",
-        num_frames=num_frames,
+    video_path = os.getenv(
+        "PADDLECODEC_TEST_VIDEO",
+        "https://paddlenlp.bj.bcebos.com/datasets/paddlemix/demo_video/example_video.mp4",
     )
-
-    if video_path == DEFAULT_TEST_VIDEO and num_frames is None:
-        assert video.to(paddle.int64).sum().item() == 247759890390
-        assert metadata.total_num_frames == 263
-        assert metadata.fps == pytest.approx(29.99418249715141)
-        assert metadata.width == 1920
-        assert metadata.height == 1080
-        assert metadata.duration == pytest.approx(8.768367)
-        for i, idx in enumerate(metadata.frames_indices):
-            assert idx == i
-    else:
-        _assert_video_decode_smoke(video, metadata, expected_num_frames=num_frames)
+    video, metadata = load_video(video_path, backend="torchcodec")
+    assert video.to(paddle.int64).sum().item() == 247759890390
+    assert metadata.total_num_frames == 263
+    assert metadata.fps == pytest.approx(29.99418249715141)
+    assert metadata.width == 1920
+    assert metadata.height == 1080
+    assert metadata.duration == pytest.approx(8.768367)
+    for i, idx in enumerate(metadata.frames_indices):
+        assert idx == i

From 35f7341aa9a76d6a916c2330bffe0865dc2d1cd9 Mon Sep 17 00:00:00 2001
From: youge325 <A372707325@126.com>
Date: Sat, 11 Apr 2026 20:01:57 +0800
Subject: [PATCH 10/14] no need to build wheel for arm ourselves

---
 .github/workflows/build_ffmpeg.yaml | 27 ---------------------------
 1 file changed, 27 deletions(-)

diff --git a/.github/workflows/build_ffmpeg.yaml b/.github/workflows/build_ffmpeg.yaml
index 5494bf0b1..9b9317e3b 100644
--- a/.github/workflows/build_ffmpeg.yaml
+++ b/.github/workflows/build_ffmpeg.yaml
@@ -48,33 +48,6 @@ jobs:
         mkdir -p "${artifact_dir}"
         mv ffmpeg.tar.gz "${artifact_dir}/${FFMPEG_VERSION}.tar.gz"
 
-  LGPL-Linux-aarch64:
-    strategy:
-      fail-fast: false
-      matrix:
-        ffmpeg-version: ["4.4.4", "5.1.4", "6.1.1", "7.0.1", "8.0"]
-    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
-    permissions:
-      id-token: write
-      contents: read
-    with:
-      job-name: Build
-      upload-artifact: ffmpeg-lgpl-linux_aarch64-${{ matrix.ffmpeg-version }}
-      repository: meta-pytorch/torchcodec
-      runner: linux.arm64.2xlarge
-      docker-image: pytorch/manylinux2_28_aarch64-builder:cpu-aarch64
-      script: |
-        export FFMPEG_VERSION="${{ matrix.ffmpeg-version }}"
-        export FFMPEG_ROOT="${PWD}/ffmpeg"
-
-        packaging/build_ffmpeg.sh
-
-        tar -cf ffmpeg.tar.gz ffmpeg/include ffmpeg/lib
-
-        artifact_dir="${RUNNER_ARTIFACT_DIR}/$(date +%Y-%m-%d)/linux_aarch64"
-        mkdir -p "${artifact_dir}"
-        mv ffmpeg.tar.gz "${artifact_dir}/${FFMPEG_VERSION}.tar.gz"
-
   LGPL-macOS:
     strategy:
       fail-fast: false

From b6de0fe4d30cefaa5071c60930b0f51ed03dae07 Mon Sep 17 00:00:00 2001
From: youge325 <A372707325@126.com>
Date: Sat, 11 Apr 2026 21:06:25 +0800
Subject: [PATCH 11/14] use faster runner

---
 .github/workflows/paddle_arm_wheel.yaml | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/.github/workflows/paddle_arm_wheel.yaml b/.github/workflows/paddle_arm_wheel.yaml
index dd3356a97..b381196b3 100644
--- a/.github/workflows/paddle_arm_wheel.yaml
+++ b/.github/workflows/paddle_arm_wheel.yaml
@@ -91,10 +91,8 @@ jobs:
 
   build:
     needs: generate-matrix
-    # If your repo does not have GitHub-hosted ARM runners enabled,
-    # replace this with self-hosted labels, e.g. [self-hosted, Linux, ARM64].
     name: Build and Upload ARM wheel
-    runs-on: ubuntu-24.04-arm
+    runs-on: linux.arm64.2xlarge
     container:
       image: pytorch/manylinux2_28_aarch64-builder:cpu-aarch64
     strategy:
@@ -188,7 +186,9 @@ jobs:
   install-and-test:
     needs: [generate-matrix, prepare-test-video, build]
     name: Install and test ARM wheel
-    runs-on: ubuntu-24.04-arm
+    runs-on: linux.arm64.2xlarge
+    container:
+      image: pytorch/manylinux2_28_aarch64-builder:cpu-aarch64
     env:
       PADDLECODEC_TEST_VIDEO: .github/test-assets/example_video.mp4
     strategy:

From 89eae288a64874b67e263d62f755ee3d3ac69f5e Mon Sep 17 00:00:00 2001
From: youge325 <A372707325@126.com>
Date: Sat, 11 Apr 2026 22:08:45 +0800
Subject: [PATCH 12/14] use public runner

---
 .github/workflows/paddle_arm_wheel.yaml | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/.github/workflows/paddle_arm_wheel.yaml b/.github/workflows/paddle_arm_wheel.yaml
index b381196b3..07ff8ca2f 100644
--- a/.github/workflows/paddle_arm_wheel.yaml
+++ b/.github/workflows/paddle_arm_wheel.yaml
@@ -92,7 +92,7 @@ jobs:
   build:
     needs: generate-matrix
     name: Build and Upload ARM wheel
-    runs-on: linux.arm64.2xlarge
+    runs-on: ubuntu-24.04-arm
     container:
       image: pytorch/manylinux2_28_aarch64-builder:cpu-aarch64
     strategy:
@@ -186,7 +186,7 @@ jobs:
   install-and-test:
     needs: [generate-matrix, prepare-test-video, build]
     name: Install and test ARM wheel
-    runs-on: linux.arm64.2xlarge
+    runs-on: ubuntu-24.04-arm
     container:
       image: pytorch/manylinux2_28_aarch64-builder:cpu-aarch64
     env:

From 56feb21de7b404a9e049afce830078b73e19e7b3 Mon Sep 17 00:00:00 2001
From: youge325 <A372707325@126.com>
Date: Sat, 11 Apr 2026 22:31:40 +0800
Subject: [PATCH 13/14] merge two workflow files into one

---
 .github/workflows/paddle_arm_wheel.yaml | 291 ------------------------
 .github/workflows/paddle_wheel.yaml     |  74 +++---
 2 files changed, 43 insertions(+), 322 deletions(-)
 delete mode 100644 .github/workflows/paddle_arm_wheel.yaml

diff --git a/.github/workflows/paddle_arm_wheel.yaml b/.github/workflows/paddle_arm_wheel.yaml
deleted file mode 100644
index 07ff8ca2f..000000000
--- a/.github/workflows/paddle_arm_wheel.yaml
+++ /dev/null
@@ -1,291 +0,0 @@
-name: Build ARM Wheels for Paddle
-
-on:
-  push:
-    branches: [paddle]
-    tags: ["v*"]
-  pull_request:
-  merge_group:
-  workflow_dispatch:
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref_name }}-${{ github.ref_type == 'branch' && github.sha }}-${{ github.event_name == 'workflow_dispatch' }}
-  cancel-in-progress: true
-
-permissions:
-  id-token: write
-  contents: write
-
-defaults:
-  run:
-    shell: bash -l -eo pipefail {0}
-
-env:
-  PADDLECODEC_TEST_VIDEO_URL: https://paddlenlp.bj.bcebos.com/datasets/paddlemix/demo_video/example_video.mp4
-  PADDLECODEC_TEST_VIDEO_CACHE_KEY: paddlecodec-test-video-v1-example-video
-  PADDLECODEC_TEST_VIDEO_PATH: .github/test-assets/example_video.mp4
-
-jobs:
-  generate-matrix:
-    runs-on: ubuntu-latest
-    outputs:
-      build-matrix: ${{ steps.set-matrix.outputs.build-matrix }}
-      test-matrix: ${{ steps.set-matrix.outputs.test-matrix }}
-    steps:
-      - name: Generate ARM build and test matrix
-        id: set-matrix
-        run: |
-          python - <<'PY'
-          import itertools
-          import json
-          import os
-
-          python_versions = ["3.9", "3.10", "3.11", "3.12", "3.13"]
-          # FFmpeg 8.0 is intentionally excluded for Paddle CPU builds to match
-          # the existing x86 workflow and avoid the current OpenVINO conflict.
-          ffmpeg_versions = ["4.4.2", "5.1.2", "6.1.1", "7.0.1"]
-
-          build_matrix = {"python-version": python_versions}
-          test_matrix = {
-              "include": [
-                  {
-                      "python-version": python_version,
-                      "ffmpeg-version": ffmpeg_version,
-                  }
-                  for python_version, ffmpeg_version in itertools.product(
-                      python_versions, ffmpeg_versions
-                  )
-              ]
-          }
-
-          with open(os.environ["GITHUB_OUTPUT"], "a", encoding="utf-8") as f:
-              f.write(f"build-matrix={json.dumps(build_matrix)}\n")
-              f.write(f"test-matrix={json.dumps(test_matrix)}\n")
-          PY
-
-  prepare-test-video:
-    runs-on: ubuntu-latest
-    name: Prepare cached Paddle test video
-    steps:
-      - name: Restore cached test video
-        id: cache-test-video
-        uses: actions/cache@v4
-        with:
-          path: ${{ env.PADDLECODEC_TEST_VIDEO_PATH }}
-          key: ${{ env.PADDLECODEC_TEST_VIDEO_CACHE_KEY }}
-
-      - name: Download test video
-        if: steps.cache-test-video.outputs.cache-hit != 'true'
-        run: |
-          mkdir -p "$(dirname "${PADDLECODEC_TEST_VIDEO_PATH}")"
-          curl --fail --location --retry 5 --retry-all-errors \
-            --output "${PADDLECODEC_TEST_VIDEO_PATH}" \
-            "${PADDLECODEC_TEST_VIDEO_URL}"
-
-      - name: Upload cached test video artifact
-        uses: actions/upload-artifact@v5
-        with:
-          name: paddlecodec-test-video
-          path: ${{ env.PADDLECODEC_TEST_VIDEO_PATH }}
-          if-no-files-found: error
-
-  build:
-    needs: generate-matrix
-    name: Build and Upload ARM wheel
-    runs-on: ubuntu-24.04-arm
-    container:
-      image: pytorch/manylinux2_28_aarch64-builder:cpu-aarch64
-    strategy:
-      fail-fast: false
-      matrix: ${{ fromJson(needs.generate-matrix.outputs.build-matrix) }}
-    permissions:
-      id-token: write
-      contents: read
-    steps:
-      - name: Checkout repository
-        uses: actions/checkout@v6
-
-      - name: Setup conda environment
-        uses: conda-incubator/setup-miniconda@v3
-        with:
-          auto-update-conda: true
-          miniforge-version: latest
-          activate-environment: build
-          python-version: ${{ matrix.python-version }}
-
-      - name: Install build dependencies
-        run: |
-          python -m pip install --upgrade pip
-          pip install build wheel setuptools
-
-      - name: Install PaddlePaddle nightly
-        run: |
-          pip install --pre paddlepaddle -i https://www.paddlepaddle.org.cn/packages/nightly/cpu/
-
-      - name: Run pre-build script
-        run: |
-          bash packaging/pre_build_script.sh
-
-      - name: Build wheel
-        run: |
-          # Match upstream's Linux workflow and build the ARM wheel against the
-          # pre-built non-GPL FFmpeg bundles published on S3.
-          export BUILD_AGAINST_ALL_FFMPEG_FROM_S3=1
-          export I_CONFIRM_THIS_IS_NOT_A_LICENSE_VIOLATION=1
-          export TORCHCODEC_CMAKE_BUILD_DIR=$(pwd)/build_cmake
-          python -m build --wheel -vvv --no-isolation
-
-      - name: Repair wheel
-        run: |
-          pip install auditwheel
-
-          # 1. Extract internal libraries from the wheel to a temporary directory
-          # This allows auditwheel to find them when checking dependencies
-          mkdir -p temp_libs
-          unzip -j dist/*.whl "torchcodec/*.so" -d temp_libs || true
-
-          # 2. Prepare LD_LIBRARY_PATH
-          # FFmpeg libraries fetched from the build-time S3 bundles
-          FFMPEG_LIB_PATHS=$(find "$(pwd)/build_cmake/_deps" -type d -name "lib" -print | paste -sd: -)
-          # PaddlePaddle libraries
-          PADDLE_PATH=$(python -c "import paddle; print(paddle.__path__[0])")
-          PADDLE_LIB_PATHS="$PADDLE_PATH/base:$PADDLE_PATH/libs"
-          # Wheel internal libraries
-          INTERNAL_LIB_PATH=$(pwd)/temp_libs
-
-          export LD_LIBRARY_PATH=${FFMPEG_LIB_PATHS}:${PADDLE_LIB_PATHS}:${INTERNAL_LIB_PATH}:${LD_LIBRARY_PATH}
-
-          # 3. Repair wheel with auditwheel
-          # We exclude all external libraries because we want to rely on system libraries (like FFmpeg)
-          # or libraries provided by other packages (like PaddlePaddle).
-          # auditwheel 6.1.0+ supports wildcards in --exclude.
-          auditwheel repair dist/*.whl --plat manylinux_2_28_aarch64 -w wheelhouse/ --exclude "*"
-
-          # Cleanup
-          rm -rf temp_libs
-          rm dist/*.whl
-          mv wheelhouse/*.whl dist/
-          rmdir wheelhouse
-
-      - name: Upload wheel artifact
-        uses: actions/upload-artifact@v5
-        with:
-          name: paddlecodec-wheel-linux-arm64-py${{ matrix.python-version }}
-          path: dist/*.whl
-
-      - name: Run post-build script
-        run: |
-          bash packaging/post_build_script.sh
-
-      - name: List wheel contents
-        run: |
-          wheel_path=$(find dist -type f -name "*.whl")
-          echo "Wheel path: $wheel_path"
-          unzip -l $wheel_path
-
-  install-and-test:
-    needs: [generate-matrix, prepare-test-video, build]
-    name: Install and test ARM wheel
-    runs-on: ubuntu-24.04-arm
-    container:
-      image: pytorch/manylinux2_28_aarch64-builder:cpu-aarch64
-    env:
-      PADDLECODEC_TEST_VIDEO: .github/test-assets/example_video.mp4
-    strategy:
-      fail-fast: false
-      # Match the existing Paddle x86 flow while keeping the upstream-style staged
-      # build/test split: exercise every Python and FFmpeg combination post-build.
-      matrix: ${{ fromJson(needs.generate-matrix.outputs.test-matrix) }}
-    steps:
-      - name: Checkout repository
-        uses: actions/checkout@v6
-
-      - name: Download wheel artifact
-        uses: actions/download-artifact@v4
-        with:
-          name: paddlecodec-wheel-linux-arm64-py${{ matrix.python-version }}
-          path: dist/
-
-      - name: Download cached test video artifact
-        uses: actions/download-artifact@v4
-        with:
-          name: paddlecodec-test-video
-          path: .github/test-assets/
-
-      - name: Install FFmpeg via conda
-        uses: conda-incubator/setup-miniconda@v3
-        with:
-          auto-update-conda: true
-          miniforge-version: latest
-          activate-environment: test
-          python-version: ${{ matrix.python-version }}
-
-      - name: Install FFmpeg from conda-forge
-        run: |
-          conda install "ffmpeg=${{ matrix.ffmpeg-version }}" -c conda-forge -y
-          ffmpeg -version
-
-      - name: Install PaddlePaddle nightly in conda env
-        run: |
-          pip install --pre paddlepaddle -i https://www.paddlepaddle.org.cn/packages/nightly/cpu/
-
-      - name: Install paddlecodec from wheel
-        run: |
-          wheel_path=$(find dist -type f -name "*.whl")
-          echo "Installing $wheel_path"
-          pip install $wheel_path -vvv
-
-      - name: Install test dependencies
-        run: |
-          pip install numpy pytest pillow
-
-      - name: Delete src folder
-        run: |
-          # Delete src/ to ensure we're testing the installed wheel, not source code
-          rm -rf src/
-          ls -la
-
-      - name: Run tests
-        run: |
-          pytest --override-ini="addopts=-v" -s test_paddle
-
-  publish-pypi-arm:
-    runs-on: ubuntu-latest
-    name: Publish ARM Wheels to PyPI
-    if: "startsWith(github.ref, 'refs/tags/')"
-    needs:
-      - install-and-test
-    permissions:
-      id-token: write
-
-    steps:
-      - name: Retrieve ARM release distributions
-        uses: actions/download-artifact@v6
-        with:
-          pattern: paddlecodec-wheel-linux-arm64-*
-          path: dist/
-          merge-multiple: true
-
-      - name: Publish ARM release distributions to PyPI
-        uses: pypa/gh-action-pypi-publish@release/v1
-
-  publish-release-arm:
-    runs-on: ubuntu-latest
-    name: Publish ARM Wheels to GitHub
-    if: "startsWith(github.ref, 'refs/tags/')"
-    needs:
-      - install-and-test
-    permissions:
-      contents: write
-    steps:
-      - uses: actions/download-artifact@v6
-        with:
-          pattern: paddlecodec-wheel-linux-arm64-*
-          path: dist/
-          merge-multiple: true
-      - name: Publish ARM wheels to GitHub
-        uses: softprops/action-gh-release@v2
-        with:
-          draft: true
-          files: dist/*
-          tag_name: ${{ github.ref_name }}
diff --git a/.github/workflows/paddle_wheel.yaml b/.github/workflows/paddle_wheel.yaml
index 86dd9c0e0..3ce2964c6 100644
--- a/.github/workflows/paddle_wheel.yaml
+++ b/.github/workflows/paddle_wheel.yaml
@@ -53,13 +53,28 @@ jobs:
           if-no-files-found: error
 
   build-paddlecodec-wheel:
-    runs-on: ubuntu-latest
+    name: Build and upload Paddle wheel (${{ matrix.arch-name }}, py${{ matrix.python-version }})
+    runs-on: ${{ matrix.runner }}
     container:
-      image: pytorch/manylinux2_28-builder:cpu
+      image: ${{ matrix.container-image }}
     strategy:
       fail-fast: false
       matrix:
         python-version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
+        arch: ["x86_64", "arm64"]
+        include:
+          - arch: x86_64
+            arch-name: x86_64
+            runner: ubuntu-latest
+            container-image: pytorch/manylinux2_28-builder:cpu
+            artifact-prefix: paddlecodec-wheel-linux
+            wheel-platform: manylinux_2_28_x86_64
+          - arch: arm64
+            arch-name: arm64
+            runner: ubuntu-24.04-arm
+            container-image: pytorch/manylinux2_28_aarch64-builder:cpu-aarch64
+            artifact-prefix: paddlecodec-wheel-linux-arm64
+            wheel-platform: manylinux_2_28_aarch64
     permissions:
       id-token: write
       contents: read
@@ -90,8 +105,8 @@ jobs:
 
       - name: Build wheel
         run: |
-          # Use pre-built FFmpeg from PyTorch S3
           export BUILD_AGAINST_ALL_FFMPEG_FROM_S3=1
+          export I_CONFIRM_THIS_IS_NOT_A_LICENSE_VIOLATION=1
           export TORCHCODEC_CMAKE_BUILD_DIR=$(pwd)/build_cmake
           python -m build --wheel -vvv --no-isolation
 
@@ -99,29 +114,18 @@ jobs:
         run: |
           pip install auditwheel
 
-          # 1. Extract internal libraries from the wheel to a temporary directory
-          # This allows auditwheel to find them when checking dependencies
           mkdir -p temp_libs
           unzip -j dist/*.whl "torchcodec/*.so" -d temp_libs || true
 
-          # 2. Prepare LD_LIBRARY_PATH
-          # FFmpeg libraries
-          FFMPEG_LIB_PATHS=$(find $(pwd)/build_cmake/_deps -type d -name "lib" | tr '\n' ':')
-          # PaddlePaddle libraries
+          FFMPEG_LIB_PATHS=$(find "$(pwd)/build_cmake/_deps" -type d -name "lib" -print | paste -sd: -)
           PADDLE_PATH=$(python -c "import paddle; print(paddle.__path__[0])")
           PADDLE_LIB_PATHS="$PADDLE_PATH/base:$PADDLE_PATH/libs"
-          # Wheel internal libraries
           INTERNAL_LIB_PATH=$(pwd)/temp_libs
 
-          export LD_LIBRARY_PATH=${FFMPEG_LIB_PATHS}${PADDLE_LIB_PATHS}:${INTERNAL_LIB_PATH}:${LD_LIBRARY_PATH}
+          export LD_LIBRARY_PATH=${FFMPEG_LIB_PATHS}:${PADDLE_LIB_PATHS}:${INTERNAL_LIB_PATH}:${LD_LIBRARY_PATH}
 
-          # 3. Repair wheel with auditwheel
-          # We exclude all external libraries because we want to rely on system libraries (like FFmpeg)
-          # or libraries provided by other packages (like PaddlePaddle).
-          # auditwheel 6.1.0+ supports wildcards in --exclude.
-          auditwheel repair dist/*.whl --plat manylinux_2_28_x86_64 -w wheelhouse/ --exclude "*"
+          auditwheel repair dist/*.whl --plat ${{ matrix.wheel-platform }} -w wheelhouse/ --exclude "*"
 
-          # Cleanup
           rm -rf temp_libs
           rm dist/*.whl
           mv wheelhouse/*.whl dist/
@@ -130,7 +134,7 @@ jobs:
       - name: Upload wheel artifact
         uses: actions/upload-artifact@v5
         with:
-          name: paddlecodec-wheel-linux-py${{ matrix.python-version }}
+          name: ${{ matrix.artifact-prefix }}-py${{ matrix.python-version }}
           path: dist/*.whl
 
       - name: Run post-build script
@@ -144,31 +148,41 @@ jobs:
           unzip -l $wheel_path
 
   test-paddlecodec-wheel:
+    name: Install and test Paddle wheel (${{ matrix.arch-name }}, py${{ matrix.python-version }}, ffmpeg ${{ matrix.ffmpeg-version }})
     needs: [prepare-test-video, build-paddlecodec-wheel]
-    runs-on: ubuntu-latest
+    runs-on: ${{ matrix.runner }}
+    container:
+      image: ${{ matrix.container-image }}
     env:
       PADDLECODEC_TEST_VIDEO: .github/test-assets/example_video.mp4
     strategy:
       fail-fast: false
       matrix:
         python-version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
+        arch: ["x86_64", "arm64"]
         # FFmpeg 8.0 depends on libopenvino.so.2520, PaddlePaddle CPU depends on libopenvino.so.2500
         # There has some conflict causing test failures, but it works with PaddlePaddle GPU.
         # We skip FFmpeg 8.0 tests for PaddlePaddle CPU builds for now.
         ffmpeg-version: ["4.4.2", "5.1.2", "6.1.1", "7.0.1"]
+        include:
+          - arch: x86_64
+            arch-name: x86_64
+            runner: ubuntu-latest
+            container-image: pytorch/manylinux2_28-builder:cpu
+            artifact-prefix: paddlecodec-wheel-linux
+          - arch: arm64
+            arch-name: arm64
+            runner: ubuntu-24.04-arm
+            container-image: pytorch/manylinux2_28_aarch64-builder:cpu-aarch64
+            artifact-prefix: paddlecodec-wheel-linux-arm64
     steps:
       - name: Checkout repository
         uses: actions/checkout@v6
 
-      - name: Set up Python ${{ matrix.python-version }}
-        uses: actions/setup-python@v5
-        with:
-          python-version: ${{ matrix.python-version }}
-
       - name: Download wheel artifact
         uses: actions/download-artifact@v4
         with:
-          name: paddlecodec-wheel-linux-py${{ matrix.python-version }}
+          name: ${{ matrix.artifact-prefix }}-py${{ matrix.python-version }}
           path: dist/
 
       - name: Download cached test video artifact
@@ -206,7 +220,6 @@ jobs:
 
       - name: Delete src folder
         run: |
-          # Delete src/ to ensure we're testing the installed wheel, not source code
           rm -rf src/
           ls -la
 
@@ -216,18 +229,17 @@ jobs:
 
   publish-pypi:
     runs-on: ubuntu-latest
-    name: Publish to PyPI
+    name: Publish Paddle wheels to PyPI
     if: "startsWith(github.ref, 'refs/tags/')"
     needs:
       - test-paddlecodec-wheel
     permissions:
       id-token: write
-
     steps:
       - name: Retrieve release distributions
         uses: actions/download-artifact@v6
         with:
-          pattern: paddlecodec-wheel-linux-*
+          pattern: paddlecodec-wheel-linux*
           path: dist/
           merge-multiple: true
 
@@ -236,7 +248,7 @@ jobs:
 
   publish-release:
     runs-on: ubuntu-latest
-    name: Publish to GitHub
+    name: Publish Paddle wheels to GitHub
     if: "startsWith(github.ref, 'refs/tags/')"
     needs:
       - test-paddlecodec-wheel
@@ -245,7 +257,7 @@ jobs:
     steps:
       - uses: actions/download-artifact@v6
         with:
-          pattern: paddlecodec-wheel-linux-*
+          pattern: paddlecodec-wheel-linux*
           path: dist/
           merge-multiple: true
       - name: Get tag name

From b62ae46935aeea7706d8c8cd078261cbd02b977a Mon Sep 17 00:00:00 2001
From: SigureMo <sigure.qaq@gmail.com>
Date: Sun, 12 Apr 2026 01:44:58 +0800
Subject: [PATCH 14/14] Make pixel value guard cross platform

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
---
 test_paddle/test_video_decode.py | 37 +++++++++++++++++++++++++++++++-
 1 file changed, 36 insertions(+), 1 deletion(-)

diff --git a/test_paddle/test_video_decode.py b/test_paddle/test_video_decode.py
index 1be61bd3e..240944cc2 100644
--- a/test_paddle/test_video_decode.py
+++ b/test_paddle/test_video_decode.py
@@ -2,6 +2,7 @@
 paddle.enable_compat(scope={"torchcodec"})
 
 import pytest
+import subprocess
 from dataclasses import dataclass, fields
 from io import BytesIO
 from typing import Callable, Mapping, Optional, Union
@@ -11,6 +12,40 @@
 import numpy as np
 
 
+def ffmpeg_rgb_sum(video_path_or_url: str) -> int:
+    # Use the local FFmpeg build as the oracle because YUV->RGB conversion is
+    # architecture- and FFmpeg-build-dependent.
+    proc = subprocess.Popen(
+        [
+            "ffmpeg",
+            "-v",
+            "error",
+            "-vsync",
+            "0",
+            "-i",
+            video_path_or_url,
+            "-f",
+            "rawvideo",
+            "-pix_fmt",
+            "rgb24",
+            "-",
+        ],
+        stdout=subprocess.PIPE,
+        stderr=subprocess.PIPE,
+    )
+    assert proc.stdout is not None
+    assert proc.stderr is not None
+
+    total = 0
+    while chunk := proc.stdout.read(16 * 1024 * 1024):
+        total += sum(chunk)
+
+    stderr = proc.stderr.read().decode("utf-8", "ignore")
+    if proc.wait() != 0:
+        raise RuntimeError(f"ffmpeg failed to decode video: {stderr}")
+    return total
+
+
 @dataclass
 class VideoMetadata(Mapping):
     total_num_frames: int
@@ -174,7 +209,7 @@ def test_video_decode():
         "https://paddlenlp.bj.bcebos.com/datasets/paddlemix/demo_video/example_video.mp4",
     )
     video, metadata = load_video(video_path, backend="torchcodec")
-    assert video.to(paddle.int64).sum().item() == 247759890390
+    assert video.to(paddle.int64).sum().item() == ffmpeg_rgb_sum(video_path)
     assert metadata.total_num_frames == 263
     assert metadata.fps == pytest.approx(29.99418249715141)
     assert metadata.width == 1920