Skip to content

[ROCm] Test ROCm failures on PR#1554 #42

[ROCm] Test ROCm failures on PR#1554

[ROCm] Test ROCm failures on PR#1554 #42

Workflow file for this run

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"
# ENABLE_ROCM=1 compiles the rocJPEG GPU JPEG decoder into
# libtorchcodec_image.so (the ROCm counterpart of nvJPEG). Video decoding
# stays on CPU. The FFmpeg core libs are still built against all FFmpeg.
build-command: "BUILD_AGAINST_ALL_FFMPEG_FROM_S3=1 ENABLE_ROCM=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: ['7.14']
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.meta-pytorch.mi350.1"
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
# We use conda only to get a Python env matching the cp310 wheel (the
# image's base env may be a different Python). Point conda at conda-forge
# only, with libmamba, so env creation is fast and never touches the
# 'defaults'/pkgs/main channel (repo.anaconda.com), which hangs from the
# ROCm runners.
echo '::group::Configure conda (libmamba solver, conda-forge only)'
conda install -n base -y -c conda-forge conda-libmamba-solver
conda config --set solver libmamba
conda config --remove-key channels 2>/dev/null || true
conda config --add channels conda-forge
conda config --set channel_priority strict
conda config --show channels
echo '::endgroup::'
# linux_job_v2 runs the script in the manylinux container's base conda
# env, whose Python isn't necessarily 3.10. The wheel is built cp310, so
# pip would reject it ("not a supported wheel on this platform"). Create
# and activate a matching Python env (conda is sourced by linux_job_v2).
echo '::group::Create Python ${{ matrix.python-version }} env'
conda create -y -n test -c conda-forge python=${{ matrix.python-version }} pip
conda activate test
python --version
echo '::endgroup::'
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${{ matrix.rocm-version }}
echo '::endgroup::'
echo '::group::Assert we are on ROCm'
rocm-smi
python -c "
import torch
print('torch:', torch.__version__, '| hip:', torch.version.hip)
assert torch.version.hip is not None, 'torch is not a ROCm/HIP build'
assert torch.cuda.is_available(), 'no GPU visible to torch'
print('device:', torch.cuda.get_device_name(0))
"
echo '::endgroup::'
# Prevent the checked-out src/ tree from shadowing the installed wheel.
bash packaging/remove_src.sh
echo '::group::Install rocJPEG runtime'
# librocjpeg is NOT bundled into the wheel. Instead, libtorchcodec_image.so
# has an RPATH entry pointing to _rocm_sdk_core/lib (ROCm 7.14 pip-wheel) and
# /opt/rocm/lib (ROCm <= 7.2 system install). install_rocjpeg.sh ensures the
# runtime side-deps (libva) are present regardless of layout.
bash packaging/install_rocjpeg.sh
echo '::endgroup::'
echo '::group::Install torchcodec from the wheel'
python -m pip install "${RUNNER_ARTIFACT_DIR}"/*.whl -vvv
echo '::endgroup::'
# This job intentionally does NOT install FFmpeg: the rocJPEG decoder and
# the other image decoders don't need it, and conda's heavy ffmpeg
# install (~110MB, 100+ pkgs) is unreliable on the ROCm runners. So we
# run only the FFmpeg-free image-decoder tests, mirroring the no-ffmpeg
# CUDA job. Video decoding on ROCm (which needs FFmpeg) is a later step.
echo '::group::Install test dependencies (no FFmpeg)'
# false: skip the optional libheif conda install (HEIC is CPU-only and
# covered elsewhere); this leaves a pip-only install (numpy/pytest/pillow).
bash packaging/install_test_dependencies.sh false
echo '::endgroup::'
echo '::group::Run FFmpeg-free image decoder tests (incl. rocJPEG GPU)'
rocm_core_lib=$(python -c "import importlib.util, pathlib, os; spec = importlib.util.find_spec('_rocm_sdk_core'); p = pathlib.Path(spec.submodule_search_locations[0]) if spec else None; dirs = [str(p/'lib'), str(p/'lib'/'rocm_sysdeps'/'lib')] if p else []; print(os.pathsep.join(d for d in dirs if pathlib.Path(d).is_dir()))" 2>/dev/null || true)
if [ -n "${rocm_core_lib}" ]; then
export LD_LIBRARY_PATH="${rocm_core_lib}:${LD_LIBRARY_PATH:-}"
fi
rocm_core_dri=$(python -c "import importlib.util, pathlib; spec = importlib.util.find_spec('_rocm_sdk_core'); p = pathlib.Path(spec.submodule_search_locations[0]) if spec else None; hits = list((p/'lib').rglob('*_drv_video.so')) if p else []; print(str(hits[0].parent)) if hits else None" 2>/dev/null || true)
rocm_va_driver=$(python -c "import importlib.util, pathlib; spec = importlib.util.find_spec('_rocm_sdk_core'); p = pathlib.Path(spec.submodule_search_locations[0]) if spec else None; hits = list((p/'lib').rglob('*_drv_video.so')) if p else []; print(hits[0].name.replace('_drv_video.so','')) if hits else None" 2>/dev/null || true)
if [ -n "${rocm_core_dri}" ] && [ -n "${rocm_va_driver}" ]; then
export LIBVA_DRIVERS_PATH="${rocm_core_dri}"
export LIBVA_DRIVER_NAME="${rocm_va_driver}"
fi
# torch.cuda.is_available() is True on ROCm (HIP masquerades as CUDA), so
# the needs_cuda image tests run here and exercise the GPU JPEG decoder,
# which is rocJPEG on ROCm: the jpeg_cuda params call
# decode_jpeg(device="cuda"). FAIL_WITHOUT_CUDA=1 makes the job fail
# loudly rather than silently skip them if no GPU is visible.
# FAIL_WITHOUT_IMAGE_CODECS=1 is the catch-all requiring every image codec
# to be present; FAIL_WITHOUT_HEIC=0 opts out of HEIC (libheif isn't
# installed here), so HEIC tests skip instead of failing.
FAIL_WITHOUT_CUDA=1 FAIL_WITHOUT_IMAGE_CODECS=1 FAIL_WITHOUT_HEIC=0 \
pytest --override-ini="addopts=-v" \
test/test_ffmpeg_optional.py test/test_decoders.py::TestImageDecoder --tb=short
echo '::endgroup::'