Add ROCm CI job

Author: NicolasHug

.github/workflows/linux_rocm.yaml

@@ -0,0 +1,134 @@
+name: Linux ROCm x86
+
+on:
+  pull_request:
+  push:
+    branches:
+      - nightly
+      - main
+      - release/*
+    tags:
+        - v[0-9]+.[0-9]+.[0-9]+-rc[0-9]+
+        - v[0-9]+.[0-9]+.[0-9]+
+  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: read
+
+defaults:
+  run:
+    shell: bash -l -eo pipefail {0}
+
+jobs:
+  generate-matrix:
+    uses: pytorch/test-infra/.github/workflows/generate_binary_build_matrix.yml@main
+    with:
+      package-type: wheel
+      os: linux
+      test-infra-repository: pytorch/test-infra
+      test-infra-ref: main
+      with-cpu: disable
+      with-xpu: disable
+      with-cuda: disable
+      with-rocm: enable
+      build-python-only: "disable"
+
+  build:
+    needs: generate-matrix
+    strategy:
+      fail-fast: false
+    name: Build and Upload wheel
+    uses: pytorch/test-infra/.github/workflows/build_wheels_linux.yml@main
+    with:
+      repository: meta-pytorch/torchcodec
+      ref: ""
+      test-infra-repository: pytorch/test-infra
+      test-infra-ref: main
+      build-matrix: ${{ needs.generate-matrix.outputs.matrix }}
+      pre-script: packaging/pre_build_script.sh
+      post-script: packaging/post_build_script.sh
+      smoke-test-script: packaging/fake_smoke_test.py
+      package-name: torchcodec
+      trigger-event: ${{ github.event_name }}
+      build-platform: "python-build-package"
+      # CPU-only compile: we build against torch-ROCm, but no ROCm/GPU decode
+      # sources exist yet (rocJPEG is a follow-up), so ENABLE_CUDA/ENABLE_ROCM
+      # are intentionally unset and this produces a CPU-only wheel.
+      build-command: "BUILD_AGAINST_ALL_FFMPEG_FROM_S3=1 python -m build --wheel -vvv --no-isolation"
+
+  install-and-test:
+    # ROCm runners are not on EC2 and authenticate to AWS (ECR) via GitHub OIDC.
+    # Fork PRs don't get an OIDC id-token from GitHub, so the job can't pull the
+    # docker image and always fails. Skip it for fork PRs; it still runs on
+    # same-repo PRs, push, and workflow_dispatch.
+    if: github.event_name != 'pull_request' || github.event.pull_request.head.repo.full_name == github.repository
+    needs: build
+    strategy:
+      fail-fast: false
+      matrix:
+        python-version: ['3.10']
+        rocm-version: ['7.1']
+        ffmpeg-version-for-tests: ['7']
+    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
+    permissions:
+      id-token: write
+      contents: read
+    with:
+      repository: meta-pytorch/torchcodec
+      runner: "linux.rocm.gpu.gfx942.1"
+      gpu-arch-type: "rocm"
+      gpu-arch-version: ${{ matrix.rocm-version }}
+      # ROCm self-hosted runners are not on EC2, so they can't pull the default
+      # ECR-hosted docker image (no AWS credentials). Use a Docker Hub image and
+      # run without sudo, matching pytorch/vision's ROCm job.
+      docker-image: pytorch/manylinux2_28-builder:rocm${{ matrix.rocm-version }}
+      no-sudo: true
+      timeout: 120
+      test-infra-ref: main
+      download-artifact: meta-pytorch_torchcodec__${{ matrix.python-version }}_rocm${{ matrix.rocm-version }}_x86_64
+      script: |
+        set -euo pipefail
+
+        python -m pip install --upgrade pip
+
+        echo '::group::Install PyTorch (ROCm)'
+        pip install --progress-bar=off --pre torch \
+          --index-url https://download.pytorch.org/whl/nightly/rocm${{ matrix.rocm-version }}
+        echo '::endgroup::'
+
+        echo '::group::Assert we are on ROCm'
+        rocm-smi
+        python -c "
+        import torch
+        print('torch:', torch.__version__, '| hip:', torch.version.hip)
+        assert torch.version.hip is not None, 'torch is not a ROCm/HIP build'
+        assert torch.cuda.is_available(), 'no GPU visible to torch'
+        print('device:', torch.cuda.get_device_name(0))
+        "
+        echo '::endgroup::'
+
+        # Prevent the checked-out src/ tree from shadowing the installed wheel.
+        bash packaging/remove_src.sh
+
+        echo '::group::Install torchcodec from the wheel'
+        python -m pip install "${RUNNER_ARTIFACT_DIR}"/*.whl -vvv
+        echo '::endgroup::'
+
+        echo '::group::Install FFmpeg and test dependencies'
+        bash packaging/install_ffmpeg.sh ${{ matrix.ffmpeg-version-for-tests }}
+        bash packaging/install_test_dependencies.sh
+        echo '::endgroup::'
+
+        echo '::group::Run Python tests'
+        # torch.cuda.is_available() is True on ROCm (HIP masquerades as CUDA), so
+        # conftest.py won't auto-skip needs_cuda tests. This wheel has no GPU
+        # decode compiled, so we deselect them explicitly. When the rocJPEG
+        # decoder lands, its GPU tests get wired in here.
+        FAIL_WITHOUT_JPEG=1 FAIL_WITHOUT_PNG=1 FAIL_WITHOUT_WEBP=1 FAIL_WITHOUT_AVIF=1 \
+          pytest -m "not needs_cuda" --override-ini="addopts=-v" test --tb=short
+        echo '::endgroup::'