Training packaging pipeline for cuda12 (#18524)

### Description  Build ORT-training packaging pipeline for CUDA 12.2 ### Motivation and Context  This will help any customer using CUDA 12 and would not need to build ORT-training from source Test run: https://dev.azure.com/aiinfra/Lotus/_build/results?buildId=382993&view=logs&s=130be951-c2f3-5601-5709-434b5e50ddb0
2026-07-22 19:23:30 +00:00 · 2023-11-21 13:19:21 -08:00 · 2023-11-21 13:19:21 -08:00 · 680a526e73
commit 680a526e73
parent 2a01622536
3 changed files with 209 additions and 0 deletions
--- a/tools/ci_build/github/azure-pipelines/orttraining-py-packaging-pipeline-cuda12.yml
+++ b/tools/ci_build/github/azure-pipelines/orttraining-py-packaging-pipeline-cuda12.yml
@ -0,0 +1,22 @@
+trigger: none
+
+resources:
+  repositories:
+  - repository: manylinux
+    type: Github
+    endpoint: Microsoft
+    name: pypa/manylinux
+    ref: 5eda9aded5462201e6310105728d33016e637ea7
+
+stages:
+- template: templates/py-packaging-training-cuda-stage.yml
+  parameters:
+    build_py_parameters: --enable_training --update --build
+    torch_version: '2.1.0'
+    opset_version: '15'
+    cuda_version: '12.2'
+    cmake_cuda_architectures: 70;75;80;86;90
+    docker_file: Dockerfile.manylinux2_28_training_cuda12_2
+    agent_pool: Onnxruntime-Linux-GPU
+    upload_wheel: 'yes'
+    debug_build: false
--- a/tools/ci_build/github/linux/docker/Dockerfile.manylinux2_28_training_cuda12_2
+++ b/tools/ci_build/github/linux/docker/Dockerfile.manylinux2_28_training_cuda12_2
@ -0,0 +1,180 @@
+ARG BASEIMAGE=nvidia/cuda:12.2.2-cudnn8-devel-ubi8
+ARG POLICY=manylinux2014
+ARG PLATFORM=x86_64
+ARG DEVTOOLSET_ROOTPATH=
+ARG LD_LIBRARY_PATH_ARG=
+ARG PREPEND_PATH=
+
+#We need both CUDA and manylinux. But the CUDA Toolkit End User License Agreement says NVIDIA CUDA Driver Libraries(libcuda.so, libnvidia-ptxjitcompiler.so) are only distributable in applications that meet this criteria:
+#1. The application was developed starting from a NVIDIA CUDA container obtained from Docker Hub or the NVIDIA GPU Cloud, and
+#2. The resulting application is packaged as a Docker container and distributed to users on Docker Hub or the NVIDIA GPU Cloud only.
+#So we use CUDA as the base image then add manylinux on top of it.
+
+#Build manylinux2014 docker image begin
+FROM $BASEIMAGE AS runtime_base
+ARG POLICY
+ARG PLATFORM
+ARG DEVTOOLSET_ROOTPATH
+ARG LD_LIBRARY_PATH_ARG
+ARG PREPEND_PATH
+LABEL maintainer="The ManyLinux project"
+
+ENV AUDITWHEEL_POLICY=${POLICY} AUDITWHEEL_ARCH=${PLATFORM} AUDITWHEEL_PLAT=${POLICY}_${PLATFORM}
+ENV LC_ALL=en_US.UTF-8 LANG=en_US.UTF-8 LANGUAGE=en_US.UTF-8
+ENV DEVTOOLSET_ROOTPATH=${DEVTOOLSET_ROOTPATH}
+ENV LD_LIBRARY_PATH=${LD_LIBRARY_PATH_ARG}
+ENV PATH=${PREPEND_PATH}${PATH}
+ENV PKG_CONFIG_PATH=/usr/local/lib/pkgconfig
+
+# first copy the fixup mirrors script, keep the script around
+COPY build_scripts/fixup-mirrors.sh /usr/local/sbin/fixup-mirrors
+
+# setup entrypoint, this will wrap commands with `linux32` with i686 images
+COPY build_scripts/install-entrypoint.sh \
+     build_scripts/build_utils.sh \
+     /build_scripts/
+
+RUN /build_scripts/install-entrypoint.sh && rm -rf /build_scripts
+COPY manylinux-entrypoint /usr/local/bin/manylinux-entrypoint
+ENTRYPOINT ["manylinux-entrypoint"]
+
+COPY build_scripts/install-runtime-packages.sh \
+     build_scripts/build_utils.sh \
+     /build_scripts/
+RUN manylinux-entrypoint /build_scripts/install-runtime-packages.sh && rm -rf /build_scripts/
+
+COPY build_scripts/build_utils.sh /build_scripts/
+
+COPY build_scripts/install-autoconf.sh /build_scripts/
+RUN export AUTOCONF_ROOT=autoconf-2.71 && \
+    export AUTOCONF_HASH=431075ad0bf529ef13cb41e9042c542381103e80015686222b8a9d4abef42a1c && \
+    export AUTOCONF_DOWNLOAD_URL=http://ftp.gnu.org/gnu/autoconf && \
+    manylinux-entrypoint /build_scripts/install-autoconf.sh
+
+COPY build_scripts/install-automake.sh /build_scripts/
+RUN export AUTOMAKE_ROOT=automake-1.16.5 && \
+    export AUTOMAKE_HASH=07bd24ad08a64bc17250ce09ec56e921d6343903943e99ccf63bbf0705e34605 && \
+    export AUTOMAKE_DOWNLOAD_URL=http://ftp.gnu.org/gnu/automake && \
+    manylinux-entrypoint /build_scripts/install-automake.sh
+
+COPY build_scripts/install-libtool.sh /build_scripts/
+RUN export LIBTOOL_ROOT=libtool-2.4.7 && \
+    export LIBTOOL_HASH=04e96c2404ea70c590c546eba4202a4e12722c640016c12b9b2f1ce3d481e9a8 && \
+    export LIBTOOL_DOWNLOAD_URL=http://ftp.gnu.org/gnu/libtool && \
+    manylinux-entrypoint /build_scripts/install-libtool.sh
+
+COPY build_scripts/install-libxcrypt.sh /build_scripts/
+RUN export LIBXCRYPT_VERSION=4.4.28 && \
+    export LIBXCRYPT_HASH=db7e37901969cb1d1e8020cb73a991ef81e48e31ea5b76a101862c806426b457 && \
+    export LIBXCRYPT_DOWNLOAD_URL=https://github.com/besser82/libxcrypt/archive && \
+    export PERL_ROOT=perl-5.34.0 && \
+    export PERL_HASH=551efc818b968b05216024fb0b727ef2ad4c100f8cb6b43fab615fa78ae5be9a && \
+    export PERL_DOWNLOAD_URL=https://www.cpan.org/src/5.0 && \
+    manylinux-entrypoint /build_scripts/install-libxcrypt.sh
+
+FROM runtime_base AS build_base
+COPY build_scripts/install-build-packages.sh /build_scripts/
+RUN manylinux-entrypoint /build_scripts/install-build-packages.sh
+
+
+FROM build_base AS build_git
+COPY build_scripts/build-git.sh /build_scripts/
+RUN export GIT_ROOT=git-2.36.2 && \
+    export GIT_HASH=6dc2cdea5fb23d823ba4871cc23222c1db31dfbb6d6c6ff74c4128700df57c68 && \
+    export GIT_DOWNLOAD_URL=https://www.kernel.org/pub/software/scm/git && \
+    manylinux-entrypoint /build_scripts/build-git.sh
+
+
+FROM build_base AS build_cpython
+COPY build_scripts/build-sqlite3.sh /build_scripts/
+RUN export SQLITE_AUTOCONF_ROOT=sqlite-autoconf-3390200 && \
+    export SQLITE_AUTOCONF_HASH=852be8a6183a17ba47cee0bbff7400b7aa5affd283bf3beefc34fcd088a239de && \
+    export SQLITE_AUTOCONF_DOWNLOAD_URL=https://www.sqlite.org/2022 && \
+    manylinux-entrypoint /build_scripts/build-sqlite3.sh
+
+COPY build_scripts/build-openssl.sh /build_scripts/
+RUN export OPENSSL_ROOT=openssl-1.1.1q && \
+    export OPENSSL_HASH=d7939ce614029cdff0b6c20f0e2e5703158a489a72b2507b8bd51bf8c8fd10ca && \
+    export OPENSSL_DOWNLOAD_URL=https://www.openssl.org/source && \
+    manylinux-entrypoint /build_scripts/build-openssl.sh
+
+COPY build_scripts/build-cpython.sh /build_scripts/
+
+
+FROM build_cpython AS build_cpython38
+COPY build_scripts/ambv-pubkey.txt /build_scripts/cpython-pubkeys.txt
+RUN manylinux-entrypoint /build_scripts/build-cpython.sh 3.8.13
+
+
+FROM build_cpython AS build_cpython39
+COPY build_scripts/ambv-pubkey.txt /build_scripts/cpython-pubkeys.txt
+RUN manylinux-entrypoint /build_scripts/build-cpython.sh 3.9.13
+
+
+FROM build_cpython AS build_cpython310
+COPY build_scripts/cpython-pubkey-310-311.txt /build_scripts/cpython-pubkeys.txt
+RUN manylinux-entrypoint /build_scripts/build-cpython.sh 3.10.5
+
+FROM build_cpython AS build_cpython311
+COPY build_scripts/cpython-pubkey-310-311.txt /build_scripts/cpython-pubkeys.txt
+RUN manylinux-entrypoint /build_scripts/build-cpython.sh 3.11.2
+
+FROM build_cpython AS all_python
+COPY build_scripts/install-pypy.sh \
+     build_scripts/pypy.sha256 \
+     build_scripts/finalize-python.sh \
+     /build_scripts/
+RUN manylinux-entrypoint /build_scripts/install-pypy.sh 3.8 7.3.9
+RUN manylinux-entrypoint /build_scripts/install-pypy.sh 3.9 7.3.9
+COPY --from=build_cpython38 /opt/_internal /opt/_internal/
+COPY --from=build_cpython39 /opt/_internal /opt/_internal/
+COPY --from=build_cpython310 /opt/_internal /opt/_internal/
+COPY --from=build_cpython311 /opt/_internal /opt/_internal/
+RUN manylinux-entrypoint /build_scripts/finalize-python.sh
+
+
+FROM runtime_base
+COPY --from=build_git /manylinux-rootfs /
+COPY --from=build_cpython /manylinux-rootfs /
+COPY --from=all_python /opt/_internal /opt/_internal/
+COPY build_scripts/finalize.sh \
+     build_scripts/python-tag-abi-tag.py \
+     build_scripts/requirements3.8.txt \
+     build_scripts/requirements3.9.txt \
+     build_scripts/requirements3.10.txt \
+     build_scripts/requirements3.11.txt \
+     build_scripts/requirements-base-tools.txt \
+     /build_scripts/
+COPY build_scripts/requirements-tools/* /build_scripts/requirements-tools/
+RUN manylinux-entrypoint /build_scripts/finalize.sh && rm -rf /build_scripts
+
+ENV SSL_CERT_FILE=/opt/_internal/certs.pem
+
+CMD ["/bin/bash"]
+
+#Build manylinux2014 docker image end
+ARG PYTHON_VERSION=3.9
+ARG TORCH_VERSION=2.1.0
+ARG OPSET_VERSION=15
+ARG INSTALL_DEPS_EXTRA_ARGS
+
+#Add our own dependencies
+ADD scripts /tmp/scripts
+RUN cd /tmp/scripts && \
+    /tmp/scripts/manylinux/install_centos.sh && \
+    /tmp/scripts/install_os_deps.sh -d gpu $INSTALL_DEPS_EXTRA_ARGS && \
+    /tmp/scripts/install_rust.sh
+
+ENV PATH="/root/.cargo/bin/:$PATH"
+
+RUN /tmp/scripts/install_ninja.sh && \
+    /tmp/scripts/install_python_deps.sh -d gpu -v 12.2 -p $PYTHON_VERSION -h $TORCH_VERSION $INSTALL_DEPS_EXTRA_ARGS && \
+    rm -rf /tmp/scripts
+
+ARG BUILD_UID=1001
+ARG BUILD_USER=onnxruntimedev
+RUN adduser --uid $BUILD_UID $BUILD_USER
+WORKDIR /home/$BUILD_USER
+USER $BUILD_USER
+ENV PATH /usr/local/dotnet:$PATH
+ENV ORTMODULE_ONNX_OPSET_VERSION=$OPSET_VERSION
--- a/tools/ci_build/github/linux/docker/scripts/training/ortmodule/stage1/requirements_torch2.1.0_cu12.2/requirements.txt
+++ b/tools/ci_build/github/linux/docker/scripts/training/ortmodule/stage1/requirements_torch2.1.0_cu12.2/requirements.txt
@ -0,0 +1,7 @@
+--pre
+-f https://download.pytorch.org/whl/torch_stable.html
+torch==2.1.0+cu121
+torchvision==0.16.0+cu121
+torchtext==0.16.0
+packaging==23.1
+setuptools>=68.2.2