Use manylinux2010 image to build linux python wheels (#1282)

* Update cuda for python wheels

* Update cuda for python wheels

* Update cuda for python wheels

* Update azure-pipelines-py-packaging.yml

* Update to cuda 10

* Only test win gpu

* Update cuda for python wheels

* Use manylinux2010 image to build linux python wheels

Allow wheels built to truly be compliant with a manylinux policy

cmake/onnxruntime_common.cmake

@@ -67,3 +67,18 @@ if(WIN32)
     # Add Code Analysis properties to enable C++ Core checks. Have to do it via a props file include.
     set_target_properties(onnxruntime_common PROPERTIES VS_USER_PROPS ${PROJECT_SOURCE_DIR}/EnableVisualStudioCodeAnalysis.props)
 endif()
+
+# check if we need to link against librt on Linux
+include(CheckLibraryExists)
+include(CheckFunctionExists)
+if ("${CMAKE_SYSTEM_NAME}" STREQUAL "Linux")
+  check_library_exists(rt clock_gettime "time.h" HAVE_CLOCK_GETTIME)
+
+  if (NOT HAVE_CLOCK_GETTIME)
+    set(CMAKE_EXTRA_INCLUDE_FILES time.h)
+    check_function_exists(clock_gettime HAVE_CLOCK_GETTIME)
+    set(CMAKE_EXTRA_INCLUDE_FILES)
+  else()
+    target_link_libraries(onnxruntime_common rt)
+  endif()
+endif()

cmake/onnxruntime_python.cmake

@@ -60,7 +60,7 @@ onnxruntime_add_include_to_target(onnxruntime_pybind11_state gsl)
 if(APPLE)
   set(ONNXRUNTIME_SO_LINK_FLAG "-Xlinker -exported_symbols_list ${ONNXRUNTIME_ROOT}/python/exported_symbols.lst")
 elseif(UNIX)
-  set(ONNXRUNTIME_SO_LINK_FLAG "-Xlinker --version-script=${ONNXRUNTIME_ROOT}/python/version_script.lds -Xlinker --no-undefined -Xlinker --gc-sections")
+  set(ONNXRUNTIME_SO_LINK_FLAG "-Xlinker --version-script=${ONNXRUNTIME_ROOT}/python/version_script.lds -Xlinker --gc-sections")
 else()
   set(ONNXRUNTIME_SO_LINK_FLAG "-DEF:${ONNXRUNTIME_ROOT}/python/pybind.def")
 endif()

onnxruntime/python/_ld_preload.py

@@ -0,0 +1,10 @@
+#-------------------------------------------------------------------------
+# Copyright (c) Microsoft Corporation. All rights reserved.
+# Licensed under the MIT License.
+#--------------------------------------------------------------------------
+
+# This file can be modified by setup.py when building a manylinux2010 wheel
+# When modified, it will preload some libraries needed for the python C extension
+# Do not remove or move the following comment
+
+# LD_PRELOAD_BEGIN_MARK

onnxruntime/python/_pybind_state.py

@@ -5,6 +5,7 @@
 import sys
 import os
 import warnings
+import onnxruntime.capi._ld_preload
 
 try:
     from onnxruntime.capi.onnxruntime_pybind11_state import *  # noqa

rename_manylinux.sh

@@ -1,14 +0,0 @@
-#!/usr/bin/env bash
-
-# Copyright (c) Microsoft Corporation. All rights reserved.
-# Licensed under the MIT License.
-
-# hack script to modify whl as manylinux whl
-whl=(*whl)
-renamed_whl=`echo $whl | sed --expression='s/linux/manylinux1/g'`
-basename=`echo $whl | awk -F'-cp3' '{print $1}'`
-unzip $whl
-sed -i 's/linux/manylinux1/g' ${basename}.dist-info/WHEEL
-# explicitly set file perms
-chmod 664 ${basename}.dist-info/*
-zip -r $renamed_whl ${basename}.data ${basename}.dist-info

setup.py

@@ -3,9 +3,14 @@
 # Licensed under the MIT License.
 #--------------------------------------------------------------------------
 
-from setuptools import setup, find_packages
-from os import path, getcwd
+from setuptools import setup, find_packages, Extension
+from distutils import log as logger
+from distutils.command.build_ext import build_ext as _build_ext
+from glob import glob
+from os import path, getcwd, environ, remove
+from shutil import copyfile
 import platform
+import subprocess
 import sys
 import datetime
 
@@ -38,12 +43,69 @@ if '--nightly_build' in sys.argv:
     nightly_build = True
     sys.argv.remove('--nightly_build')
 
+is_manylinux2010 = False
+if environ.get('AUDITWHEEL_PLAT', None) == 'manylinux2010_x86_64':
+    is_manylinux2010 = True
+
+
+class build_ext(_build_ext):
+    def build_extension(self, ext):
+        dest_file = self.get_ext_fullpath(ext.name)
+        logger.info('copying %s -> %s', ext.sources[0], dest_file)
+        copyfile(ext.sources[0], dest_file)
+
+
 try:
     from wheel.bdist_wheel import bdist_wheel as _bdist_wheel
     class bdist_wheel(_bdist_wheel):
         def finalize_options(self):
             _bdist_wheel.finalize_options(self)
-            self.root_is_pure = False
+            if not is_manylinux2010:
+                self.root_is_pure = False
+
+        def _rewrite_ld_preload(self, to_preload):
+            with open('onnxruntime/capi/_ld_preload.py', 'rt') as f:
+                ld_preload = f.read().splitlines()
+            with open('onnxruntime/capi/_ld_preload.py', 'wt') as f:
+                for line in ld_preload:
+                    f.write(line)
+                    f.write('\n')
+                    if 'LD_PRELOAD_BEGIN_MARK' in line:
+                        break
+                if len(to_preload) > 0:
+                    f.write('from ctypes import CDLL, RTLD_GLOBAL\n')
+                    for library in to_preload:
+                        f.write('_{} = CDLL("{}", mode=RTLD_GLOBAL)\n'.format(library.split('.')[0], library))
+
+        def run(self):
+            if is_manylinux2010:
+                source = 'onnxruntime/capi/onnxruntime_pybind11_state.so'
+                dest = 'onnxruntime/capi/onnxruntime_pybind11_state_manylinux2010.so'
+                logger.info('copying %s -> %s', source, dest)
+                copyfile(source, dest)
+                result = subprocess.run(['patchelf', '--print-needed', dest], check=True, stdout=subprocess.PIPE, universal_newlines=True)
+                cuda_dependencies = ['libcublas.so', 'libcudnn.so', 'libcudart.so']
+                to_preload = []
+                args = ['patchelf', '--debug']
+                for line in result.stdout.split('\n'):
+                    for dependency in cuda_dependencies:
+                        if dependency in line:
+                            to_preload.append(line)
+                            args.extend(['--remove-needed', line])
+                args.append(dest)
+                if len(to_preload) > 0:
+                    subprocess.run(args, check=True, stdout=subprocess.PIPE)
+                self._rewrite_ld_preload(to_preload)
+            _bdist_wheel.run(self)
+            if is_manylinux2010:
+                file = glob(path.join(self.dist_dir, '*linux*.whl'))[0]
+                logger.info('repairing %s for manylinux2010', file)
+                try:
+                    subprocess.run(['auditwheel', 'repair', '--plat', 'manylinux2010_x86_64', '-w', self.dist_dir, file], check=True, stdout=subprocess.PIPE)
+                finally:
+                    logger.info('removing %s', file)
+                    remove(file)
+
 except ImportError:
     bdist_wheel = None
 
@@ -57,7 +119,18 @@ elif platform.system() == "Darwin":
 else:
   libs = ['onnxruntime_pybind11_state.pyd', 'mkldnn.dll', 'mklml.dll', 'libiomp5md.dll']
 
-data = [path.join('capi', x) for x in libs if path.isfile(path.join('onnxruntime', 'capi', x))]
+if is_manylinux2010:
+    data = []
+    ext_modules = [
+        Extension(
+            'onnxruntime.capi.onnxruntime_pybind11_state',
+            ['onnxruntime/capi/onnxruntime_pybind11_state_manylinux2010.so'],
+        ),
+    ]
+else:
+    data = [path.join('capi', x) for x in libs if path.isfile(path.join('onnxruntime', 'capi', x))]
+    ext_modules = []
+
 
 python_modules_list = list()
 if '--use_openvino' in sys.argv:
@@ -98,7 +171,7 @@ setup(
     long_description=long_description,
     author='Microsoft Corporation',
     author_email='onnx@microsoft.com',
-    cmdclass={'bdist_wheel': bdist_wheel},
+    cmdclass={'bdist_wheel': bdist_wheel, 'build_ext': build_ext},
     license="MIT License",
     packages=['onnxruntime',
               'onnxruntime.backend',
@@ -106,6 +179,7 @@ setup(
               'onnxruntime.datasets',
               'onnxruntime.tools',
               ],
+    ext_modules=ext_modules,
     package_data={
         'onnxruntime': data + examples + extra,
     },

tools/ci_build/build.py

@@ -683,33 +683,20 @@ def run_server_model_tests(build_dir, configs):
 def build_python_wheel(source_dir, build_dir, configs, use_cuda, use_ngraph, use_tensorrt, use_openvino, nightly_build = False):
     for config in configs:
         cwd = get_config_build_dir(build_dir, config)
-
         if is_windows():
             cwd = os.path.join(cwd, config)
+        args = [sys.executable, os.path.join(source_dir, 'setup.py'), 'bdist_wheel']
         if nightly_build:
-            if use_tensorrt:
-                run_subprocess([sys.executable, os.path.join(source_dir, 'setup.py'), 'bdist_wheel', '--use_tensorrt', '--nightly_build'], cwd=cwd)
-            elif use_cuda:
-                run_subprocess([sys.executable, os.path.join(source_dir, 'setup.py'), 'bdist_wheel', '--use_cuda', '--nightly_build'], cwd=cwd)
-            elif use_ngraph:
-                run_subprocess([sys.executable, os.path.join(source_dir, 'setup.py'), 'bdist_wheel', '--use_ngraph', '--nightly-build'], cwd=cwd)
-            elif use_openvino:
-                run_subprocess([sys.executable, os.path.join(source_dir, 'setup.py'), 'bdist_wheel', '--use_openvino', '--nightly-build'], cwd=cwd)
-            else:
-                run_subprocess([sys.executable, os.path.join(source_dir, 'setup.py'), 'bdist_wheel', '--nightly_build'], cwd=cwd)
-        else:
-            if use_tensorrt:
-                run_subprocess([sys.executable, os.path.join(source_dir, 'setup.py'), 'bdist_wheel', '--use_tensorrt'], cwd=cwd)
-            elif use_cuda:
-                run_subprocess([sys.executable, os.path.join(source_dir, 'setup.py'), 'bdist_wheel', '--use_cuda'], cwd=cwd)
-            elif use_ngraph:
-                run_subprocess([sys.executable, os.path.join(source_dir, 'setup.py'), 'bdist_wheel', '--use_ngraph'], cwd=cwd)
-            elif use_openvino:
-                run_subprocess([sys.executable, os.path.join(source_dir, 'setup.py'), 'bdist_wheel', '--use_openvino'], cwd=cwd)
-            else:
-                run_subprocess([sys.executable, os.path.join(source_dir, 'setup.py'), 'bdist_wheel'], cwd=cwd)
-        if is_ubuntu_1604():
-            run_subprocess([os.path.join(source_dir, 'rename_manylinux.sh')], cwd=cwd+'/dist')
+            args.append('--nightly_build')
+        if use_tensorrt:
+            args.append('--use_tensorrt')
+        elif use_cuda:
+            args.append('--use_cuda')
+        elif use_ngraph:
+            args.append('--use_ngraph')
+        elif use_openvino:
+            args.append('--use_openvino')
+        run_subprocess(args, cwd=cwd)
 
 def build_protoc_for_windows_host(cmake_path, source_dir, build_dir):
     if not is_windows():

tools/ci_build/github/azure-pipelines/azure-pipelines-py-packaging.yml

@@ -1,5 +1,5 @@
 jobs:
-- job: Ubuntu1604_py_Wheels
+- job: Manylinux2010_py_Wheels
   pool: Linux-CPU
   strategy:
     matrix:
@@ -32,13 +32,13 @@ jobs:
       displayName: 'Run build script'
       inputs:
         scriptPath: 'tools/ci_build/github/linux/run_dockerbuild.sh'
-        args: '-c Release -o ubuntu16.04 -d cpu -r $(Build.BinariesDirectory) -p $(python.version) -x "--use_openmp --build_wheel"'
+        args: '-c Release -o manylinux2010 -d cpu -r $(Build.BinariesDirectory) -p $(python.version) -x "--use_openmp --build_wheel"'
 
     - task: CopyFiles@2
       displayName: 'Copy Python Wheel to:  $(Build.ArtifactStagingDirectory)'
       inputs:
         SourceFolder: '$(Build.BinariesDirectory)'
-        Contents: 'Release/dist/*-manylinux1_x86_64.whl'
+        Contents: 'Release/dist/*-manylinux2010_x86_64.whl'
         TargetFolder: '$(Build.ArtifactStagingDirectory)'
 
     - task: PublishBuildArtifacts@1
@@ -51,7 +51,7 @@ jobs:
 
     - template: templates/clean-agent-build-directory-step.yml
 
-- job: Ubuntu1604_py_GPU_Wheels
+- job: Manylinux2010_py_GPU_Wheels
   pool: Linux-GPU
   strategy:
     matrix:
@@ -94,13 +94,13 @@ jobs:
       displayName: 'Run build script'
       inputs:
         scriptPath: 'tools/ci_build/github/linux/run_dockerbuild.sh'
-        args: '-c Release -o ubuntu16.04 -d gpu -c cuda9.1-cudnn7.1 -r $(Build.BinariesDirectory) -p $(python.version) -x "--use_openmp --build_wheel"'
+        args: '-c Release -o manylinux2010 -d gpu -c cuda10.1 -r $(Build.BinariesDirectory) -p $(python.version) -x "--use_openmp --build_wheel"'
 
     - task: CopyFiles@2
       displayName: 'Copy Python Wheel to:  $(Build.ArtifactStagingDirectory)'
       inputs:
         SourceFolder: '$(Build.BinariesDirectory)'
-        Contents: 'Release/dist/*-manylinux1_x86_64.whl'
+        Contents: 'Release/dist/*-manylinux2010_x86_64.whl'
         TargetFolder: '$(Build.ArtifactStagingDirectory)'
 
     - task: PublishBuildArtifacts@1
@@ -161,7 +161,7 @@ jobs:
   pool: Win-GPU
   variables:
     buildDirectory: '$(Build.SourcesDirectory)\build'
-    CUDA_VERSION: '9.1'
+    CUDA_VERSION: '10.0'
   strategy:
     matrix:
       Python35:
@@ -178,12 +178,6 @@ jobs:
         packageSpecs: 'python=$(python.version)'
         cleanEnvironment: true
 
-    - task: PowerShell@1
-      displayName: 'Set CUDA path'
-      inputs:
-        scriptName: 'tools/ci_build/github/windows/set_cuda_path.ps1'
-        arguments: '-CudaMsbuildPath C:\local\cudaMsbuildIntegration-9.1.85-windows10-x64-0 -CudaVersion $(CUDA_VERSION)'
-
     - task: BatchScript@1
       displayName: 'Setup VS2017 env vars'
       inputs:
@@ -195,8 +189,8 @@ jobs:
       displayName: 'Run build script'
       inputs:
         filename: 'build.bat'
-        arguments: ' --use_cuda --cuda_home="C:\local\cuda-9.1.85-windows10-x64-0"
-      --cudnn_home="C:\local\cudnn-9.1-windows10-x64-v7.1\cuda" --build_dir $(buildDirectory) --config Release --use_openmp --build_wheel'
+        arguments: ' --use_cuda --cuda_home="C:\local\cuda_10.0.130_win10"
+      --cudnn_home="C:\local\cudnn-10.0-windows10-x64-v7.3.1.20\cuda" --build_dir $(buildDirectory) --config Release --use_openmp --build_wheel'
         workingFolder: "$(Build.SourcesDirectory)"
 
     - task: CopyFiles@2
@@ -215,45 +209,3 @@ jobs:
       displayName: 'Component Detection'
 
     - template: templates/clean-agent-build-directory-step.yml
-
-- job: MacOS_py_Wheels
-  pool:
-    vmImage: 'macOS-10.13'
-  strategy:
-    matrix:
-      Python35:
-        python.version: '3.5'
-      Python36:
-        python.version: '3.6'
-      Python37:
-        python.version: '3.7'
-  steps:
-    - task: CondaEnvironment@1
-      inputs:
-        createCustomEnvironment: true
-        environmentName: 'py$(python.version)'
-        packageSpecs: 'python=$(python.version)'
-        cleanEnvironment: true
-
-    - script: |
-        sudo python -m pip install numpy==1.15.0
-        sudo xcode-select --switch /Applications/Xcode_10.app/Contents/Developer
-        ./build.sh --config Release --skip_submodule_sync --parallel --use_openmp --build_wheel
-      displayName: 'Command Line Script'
-
-    - task: CopyFiles@2
-      displayName: 'Copy Python Wheel to: $(Build.ArtifactStagingDirectory)'
-      inputs:
-        SourceFolder: '$(Build.SourcesDirectory)'
-        Contents: '**/dist/*.whl'
-        TargetFolder: '$(Build.ArtifactStagingDirectory)'
-
-    - task: PublishBuildArtifacts@1
-      displayName: 'Publish Artifact: ONNXRuntime python wheel'
-      inputs:
-        ArtifactName: onnxruntime
-
-    - task: ms.vss-governance-buildtask.governance-build-task-component-detection.ComponentGovernanceComponentDetection@0
-      displayName: 'Component Detection'
-
-    - template: templates/clean-agent-build-directory-step.yml

tools/ci_build/github/azure-pipelines/c-api-packaging-pipelines.yml

@@ -180,11 +180,4 @@ jobs:
         artifactName: 'onnxruntime-win-$(buildArch)-gpu-$(OnnxRuntimeVersion)'
         commitId: $(OnnxRuntimeGitCommitHash)
 
-    - task: PowerShell@2
-      displayName: 'Clean up Cuda Path 9.1'
-      inputs:
-        targetType: 'filePath'
-        filePath: '$(Build.SourcesDirectory)/tools/ci_build/github/windows/clean_up_cuda_prop_files.ps1'
-        arguments: '-CudaVersion 9.1'
-
     - template: templates/clean-agent-build-directory-step.yml

tools/ci_build/github/linux/docker/Dockerfile.manylinux2010

@@ -0,0 +1,31 @@
+FROM quay.io/pypa/manylinux2010_x86_64:latest
+
+ARG PYTHON_VERSION=3.5
+
+ADD scripts/install_manylinux2010.sh /tmp/scripts/install_manylinux2010.sh
+RUN /tmp/scripts/install_manylinux2010.sh -p ${PYTHON_VERSION}
+ADD scripts/install_protobuf.sh /tmp/scripts/install_protobuf.sh
+RUN (source /opt/onnxruntime-python/bin/activate; pip install cmake && /tmp/scripts/install_protobuf.sh && pip uninstall -y cmake)
+ADD scripts /tmp/scripts
+RUN (source /opt/onnxruntime-python/bin/activate; /tmp/scripts/install_deps.sh)
+RUN rm -rf /tmp/scripts  # not useful at all except not to see the scripts
+
+RUN echo "#!/bin/bash" > /opt/entrypoint.sh && \
+    echo "set -e" >> /opt/entrypoint.sh && \
+    echo "source /opt/onnxruntime-python/bin/activate" >> /opt/entrypoint.sh && \
+    echo "exec \"\$@\"" >> /opt/entrypoint.sh
+RUN cat /opt/entrypoint.sh
+RUN chmod +x /opt/entrypoint.sh
+
+WORKDIR /root
+
+ENV LD_LIBRARY_PATH /usr/local/openblas/lib:$LD_LIBRARY_PATH
+
+ARG BUILD_UID=1000
+ARG BUILD_USER=onnxruntimedev
+WORKDIR /home/$BUILD_USER
+# --disabled-password
+RUN adduser --comment 'onnxruntime Build User' $BUILD_USER --uid $BUILD_UID
+USER $BUILD_USER
+
+ENTRYPOINT ["/opt/entrypoint.sh"]

tools/ci_build/github/linux/docker/Dockerfile.manylinux2010_gpu

@@ -0,0 +1,66 @@
+FROM quay.io/pypa/manylinux2010_x86_64:latest
+
+ARG PYTHON_VERSION=3.5
+
+ADD scripts/install_manylinux2010.sh /tmp/scripts/install_manylinux2010.sh
+RUN /tmp/scripts/install_manylinux2010.sh -p ${PYTHON_VERSION}
+ADD scripts/install_protobuf.sh /tmp/scripts/install_protobuf.sh
+RUN (source /opt/onnxruntime-python/bin/activate; pip install cmake && /tmp/scripts/install_protobuf.sh && pip uninstall -y cmake)
+ADD scripts /tmp/scripts
+RUN (source /opt/onnxruntime-python/bin/activate; /tmp/scripts/install_deps.sh)
+RUN rm -rf /tmp/scripts  # not useful at all except not to see the scripts
+
+RUN echo "#!/bin/bash" > /opt/entrypoint.sh && \
+    echo "set -e" >> /opt/entrypoint.sh && \
+    echo "source /opt/onnxruntime-python/bin/activate" >> /opt/entrypoint.sh && \
+    echo "exec \"\$@\"" >> /opt/entrypoint.sh
+RUN cat /opt/entrypoint.sh
+RUN chmod +x /opt/entrypoint.sh
+
+RUN NVIDIA_GPGKEY_SUM=d1be581509378368edeec8c1eb2958702feedf3bc3d17011adbf24efacce4ab5 && \
+    curl -fsSL https://developer.download.nvidia.com/compute/cuda/repos/rhel6/x86_64/7fa2af80.pub | sed '/^Version/d' > /etc/pki/rpm-gpg/RPM-GPG-KEY-NVIDIA && \
+    echo "$NVIDIA_GPGKEY_SUM  /etc/pki/rpm-gpg/RPM-GPG-KEY-NVIDIA" | sha256sum -c -
+
+COPY cuda_manylinux2010.repo /etc/yum.repos.d/cuda.repo
+
+ENV CUDA_VERSION 10.1.168
+ENV CUDA_PKG_VERSION 10-1-$CUDA_VERSION-1
+
+# For libraries in the cuda-compat-* package: https://docs.nvidia.com/cuda/eula/index.html#attachment-a
+RUN yum install -y \
+        cuda-cudart-$CUDA_PKG_VERSION \
+        cuda-compat-10-1 \
+        cuda-libraries-$CUDA_PKG_VERSION \
+        cuda-nvtx-$CUDA_PKG_VERSION \
+        cuda-libraries-dev-$CUDA_PKG_VERSION \
+        cuda-nvml-dev-$CUDA_PKG_VERSION \
+        cuda-minimal-build-$CUDA_PKG_VERSION \
+        cuda-command-line-tools-$CUDA_PKG_VERSION \
+    && \
+    ln -s cuda-10.1 /usr/local/cuda && \
+    rm -rf /var/cache/yum/*
+
+# cuDNN license: https://developer.nvidia.com/cudnn/license_agreement
+RUN CUDNN_DOWNLOAD_SUM=e956c6f9222fcb867a10449cfc76dee5cfd7c7531021d95fe9586d7e043b57d7 && \
+    curl -fsSL http://developer.download.nvidia.com/compute/redist/cudnn/v7.6.0/cudnn-10.1-linux-x64-v7.6.0.64.tgz -O && \
+    echo "$CUDNN_DOWNLOAD_SUM  cudnn-10.1-linux-x64-v7.6.0.64.tgz" | sha256sum -c - && \
+    tar --no-same-owner -xzf cudnn-10.1-linux-x64-v7.6.0.64.tgz -C /usr/local && \
+    rm cudnn-10.1-linux-x64-v7.6.0.64.tgz && \
+    ldconfig
+
+ENV LD_LIBRARY_PATH /usr/local/openblas/lib:/usr/local/cuda/lib64/stubs:/usr/local/nvidia/lib:/usr/local/nvidia/lib64:$LD_LIBRARY_PATH
+ENV PATH /usr/local/nvidia/bin:/usr/local/cuda/bin:${PATH}
+
+# nvidia-container-runtime
+ENV NVIDIA_VISIBLE_DEVICES all
+ENV NVIDIA_DRIVER_CAPABILITIES compute,utility
+ENV NVIDIA_REQUIRE_CUDA "cuda>=10.1 brand=tesla,driver>=384,driver<385 brand=tesla,driver>=396,driver<397 brand=tesla,driver>=410,driver<411"
+
+ARG BUILD_UID=1000
+ARG BUILD_USER=onnxruntimedev
+WORKDIR /home/$BUILD_USER
+# --disabled-password
+RUN adduser --comment 'onnxruntime Build User' $BUILD_USER --uid $BUILD_UID
+USER $BUILD_USER
+
+ENTRYPOINT ["/opt/entrypoint.sh"]

tools/ci_build/github/linux/docker/cuda_manylinux2010.repo

@@ -0,0 +1,6 @@
+[cuda]
+name=cuda
+baseurl=http://developer.download.nvidia.com/compute/cuda/repos/rhel6/x86_64
+enabled=1
+gpgcheck=1
+gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-NVIDIA

tools/ci_build/github/linux/docker/scripts/install_deps.sh

@@ -37,6 +37,9 @@ rm -rf /tmp/src
 DISTRIBUTOR=$(lsb_release -i -s)
 if [ "$DISTRIBUTOR" = "Ubuntu" ]; then
   apt-get -y remove libprotobuf-dev protobuf-compiler
+elif [ "$AUDITWHEEL_PLAT" = "manylinux2010_x86_64" ]; then
+  # we did not install protobuf 2.x no need to uninstall
+  :
 else
   dnf remove -y protobuf-devel protobuf-compiler
 fi

tools/ci_build/github/linux/docker/scripts/install_manylinux2010.sh

@@ -0,0 +1,42 @@
+#!/bin/bash
+set -e
+set -x
+
+while getopts p: parameter_Option
+do case "${parameter_Option}"
+in
+p) PYTHON_VER=${OPTARG};;
+esac
+done
+
+PYTHON_VER=${PYTHON_VER:=3.5}
+CPYTHON_VER=cp${PYTHON_VER//./}
+
+# need to install rpmforge in order to get aria2
+curl -fsSLo /tmp/rpmforge.rpm http://repository.it4i.cz/mirrors/repoforge/redhat/el6/en/x86_64/rpmforge/RPMS/rpmforge-release-0.5.3-1.el6.rf.x86_64.rpm
+yum -y install /tmp/rpmforge.rpm
+rm -f /tmp/rpmforge.rpm
+
+yum -y install openblas-devel zlib-devel curl-devel expat-devel aria2 rsync redhat-lsb-core
+yum -y clean all
+
+/opt/python/${CPYTHON_VER}-${CPYTHON_VER}m/bin/python -m venv /opt/onnxruntime-python
+source /opt/onnxruntime-python/bin/activate
+if [ ! -f /opt/onnxruntime-python/bin/python${PYTHON_VER} ]; then
+  ln -s python /opt/onnxruntime-python/bin/python${PYTHON_VER}
+fi
+python -m pip install --upgrade --force-reinstall pip==19.1.1
+python -m pip install --upgrade --force-reinstall numpy==1.16.3
+python -m pip install --upgrade --force-reinstall requests==2.21.0
+python -m pip install --upgrade --force-reinstall wheel==0.31.1
+python -m pip install --upgrade --force-reinstall setuptools==41.0.1
+python -m pip install --upgrade --force-reinstall pytest==4.6.2
+
+ls -al /opt/onnxruntime-python/bin
+
+echo "#!/bin/sh" > /opt/entrypoint.sh
+echo "source /opt/onnxruntime-python/bin/activate" >> /opt/entrypoint.sh
+echo "exec \"$@\"" >> /opt/entrypoint.sh
+
+mkdir -p $HOME/.aria2
+echo "ca-certificate=/opt/_internal/certs.pem" > $HOME/.aria2/aria2.conf

tools/ci_build/github/linux/docker/scripts/install_onnx.sh

@@ -1,10 +1,10 @@
 #!/bin/bash
 
 # The script is to generate all supported versions of onnx models which will be tested by onnx_test_runner
-# in the end of ci build pipeline. The purpose is to make sure latest onnxruntime has no regressions. Note 
-# that the order of installation must be onnx123, onnx130, onnx141, onnx150 and onnxtip since we want 
+# in the end of ci build pipeline. The purpose is to make sure latest onnxruntime has no regressions. Note
+# that the order of installation must be onnx123, onnx130, onnx141, onnx150 and onnxtip since we want
 # to keep the tip of master on script exit for onnx backend test which is also a part of build pipeline.
-# One possible improvement here is to keep the models saved to some public storage instead of generating 
+# One possible improvement here is to keep the models saved to some public storage instead of generating
 # on the fly every time.
 
 set -e
@@ -21,14 +21,14 @@ for v2t in ${version2tag[*]}; do
     echo "first pass";
   else
     echo "deleting old onnx-${lastest_onnx_version}";
-    /usr/bin/python${PYTHON_VER} -m pip uninstall -y onnx
+    python${PYTHON_VER} -m pip uninstall -y onnx
   fi
   lastest_onnx_version=$onnx_version
   aria2c -q -d /tmp/src  https://github.com/onnx/onnx/archive/$onnx_version.tar.gz
   tar -xf /tmp/src/onnx-$onnx_version.tar.gz -C /tmp/src
   cd /tmp/src/onnx-$onnx_version
   git clone https://github.com/pybind/pybind11.git third_party/pybind11
-  /usr/bin/python${PYTHON_VER} -m pip install .
+  python${PYTHON_VER} -m pip install .
   mkdir -p /data/onnx/${onnx_tag}
   backend-test-tools generate-data -o /data/onnx/$onnx_tag
 done

tools/ci_build/github/linux/run_build.sh

@@ -21,13 +21,29 @@ if [ $BUILD_OS = "android" ]; then
     make -j$(nproc)
 else
     COMMON_BUILD_ARGS="--skip_submodule_sync --enable_onnx_tests --parallel --build_shared_lib --use_openmp --cmake_path /usr/bin/cmake --ctest_path /usr/bin/ctest"
+    if [ $BUILD_OS = "manylinux2010" ]; then
+        # FindPython3 does not work on manylinux2010 image, define things manually
+        # ask python where to find includes
+        COMMON_BUILD_ARGS="${COMMON_BUILD_ARGS} --cmake_extra_defines PYTHON_INCLUDE_DIR=$(python3 -c 'import distutils.sysconfig; print(distutils.sysconfig.get_python_inc())')"
+        # Python does not provide a shared library on manylinux, use another library
+        COMMON_BUILD_ARGS="${COMMON_BUILD_ARGS} PYTHON_LIBRARY=/usr/lib64/librt.so"
+
+    fi
     if [ $BUILD_DEVICE = "gpu" ]; then
-        _CUDNN_VERSION=$(echo $CUDNN_VERSION | cut -d. -f1-2)
-        python3 $SCRIPT_DIR/../../build.py --build_dir /build \
-            --config Debug Release $COMMON_BUILD_ARGS \
-            --use_cuda \
-            --cuda_home /usr/local/cuda \
-            --cudnn_home /usr/local/cudnn-$_CUDNN_VERSION/cuda $BUILD_EXTR_PAR
+        if [ $BUILD_OS = "manylinux2010" ]; then
+            python3 $SCRIPT_DIR/../../build.py --build_dir /build \
+                --config Debug Release $COMMON_BUILD_ARGS \
+                --use_cuda \
+                --cuda_home /usr/local/cuda \
+                --cudnn_home /usr/local/cuda $BUILD_EXTR_PAR
+        else
+            _CUDNN_VERSION=$(echo $CUDNN_VERSION | cut -d. -f1-2)
+            python3 $SCRIPT_DIR/../../build.py --build_dir /build \
+                --config Debug Release $COMMON_BUILD_ARGS \
+                --use_cuda \
+                --cuda_home /usr/local/cuda \
+                --cudnn_home /usr/local/cudnn-$_CUDNN_VERSION/cuda $BUILD_EXTR_PAR
+        fi
     elif [ $BUILD_DEVICE = "tensorrt" ]; then
         _CUDNN_VERSION=$(echo $CUDNN_VERSION | cut -d. -f1-2)
         python3 $SCRIPT_DIR/../../build.py --build_dir /build \

tools/ci_build/github/linux/run_dockerbuild.sh

@@ -8,7 +8,7 @@ CUDA_VER=cuda10.0-cudnn7.3
 while getopts c:o:d:r:p:x:a:v: parameter_Option
 do case "${parameter_Option}"
 in
-#android, ubuntu16.04
+#android, ubuntu16.04, manylinux2010
 o) BUILD_OS=${OPTARG};;
 #cpu, gpu, tensorrt
 d) BUILD_DEVICE=${OPTARG};;
@@ -22,7 +22,7 @@ c) CUDA_VER=${OPTARG};;
 # x86 or other, only for ubuntu16.04 os
 a) BUILD_ARCH=${OPTARG};;
 # openvino version tag: 2018_R5, 2019_R1 (Default is 2018_R5)
-v) OPENVINO_VERSION=${OPTARG};; 
+v) OPENVINO_VERSION=${OPTARG};;
 esac
 done
 
@@ -35,6 +35,15 @@ if [ $BUILD_OS = "android" ]; then
     IMAGE="android"
     DOCKER_FILE=Dockerfile.ubuntu_for_android
     docker build -t "onnxruntime-$IMAGE" --build-arg BUILD_USER=onnxruntimedev --build-arg BUILD_UID=$(id -u) --build-arg PYTHON_VERSION=${PYTHON_VER} -f $DOCKER_FILE .
+elif [ $BUILD_OS = "manylinux2010" ]; then
+    if [ $BUILD_DEVICE = "gpu" ]; then
+        IMAGE="manylinux2010-cuda10.1"
+        DOCKER_FILE=Dockerfile.manylinux2010_gpu
+    else
+        IMAGE="manylinux2010"
+        DOCKER_FILE=Dockerfile.manylinux2010
+    fi
+    docker build -t "onnxruntime-$IMAGE" --build-arg BUILD_USER=onnxruntimedev --build-arg BUILD_UID=$(id -u) --build-arg PYTHON_VERSION=${PYTHON_VER} -f $DOCKER_FILE .
 else
     if [ $BUILD_DEVICE = "gpu" ]; then
         IMAGE="ubuntu16.04-$CUDA_VER"
@@ -67,35 +76,26 @@ mkdir -p ~/.cache/onnxruntime
 mkdir -p ~/.onnx
 
 if [ -z "$NIGHTLY_BUILD" ]; then
-set NIGHTLY_BUILD=0
+    set NIGHTLY_BUILD=0
 fi
 
 if [ $BUILD_DEVICE = "cpu" ] || [ $BUILD_DEVICE = "ngraph" ] || [ $BUILD_DEVICE = "openvino" ]; then
-    docker rm -f "onnxruntime-$BUILD_DEVICE" || true
-    docker run -h $HOSTNAME \
-        --name "onnxruntime-$BUILD_DEVICE" \
-        --volume "$SOURCE_ROOT:/onnxruntime_src" \
-        --volume "$BUILD_DIR:/build" \
-        --volume "$HOME/.cache/onnxruntime:/home/onnxruntimedev/.cache/onnxruntime" \
-        --volume "$HOME/.onnx:/home/onnxruntimedev/.onnx" \
-        -e NIGHTLY_BUILD \
-        "onnxruntime-$IMAGE" \
-        /bin/bash /onnxruntime_src/tools/ci_build/github/linux/run_build.sh \
-         -d $BUILD_DEVICE -x "$BUILD_EXTR_PAR" -o $BUILD_OS &
+    ONNX_DOCKER=docker
 else
-    docker rm -f "onnxruntime-$BUILD_DEVICE" || true
-    nvidia-docker run --rm -h $HOSTNAME \
-        --rm \
-        --name "onnxruntime-$BUILD_DEVICE" \
-        --volume "$SOURCE_ROOT:/onnxruntime_src" \
-        --volume "$BUILD_DIR:/build" \
-        --volume "$HOME/.cache/onnxruntime:/home/onnxruntimedev/.cache/onnxruntime" \
-        --volume "$HOME/.onnx:/home/onnxruntimedev/.onnx" \
-        -e NIGHTLY_BUILD \
-        "onnxruntime-$IMAGE" \
-        /bin/bash /onnxruntime_src/tools/ci_build/github/linux/run_build.sh \
-        -d $BUILD_DEVICE -x "$BUILD_EXTR_PAR" -o $BUILD_OS &
+    ONNX_DOCKER=nvidia-docker
 fi
+
+docker rm -f "onnxruntime-$BUILD_DEVICE" || true
+$ONNX_DOCKER run -h $HOSTNAME \
+    --name "onnxruntime-$BUILD_DEVICE" \
+    --volume "$SOURCE_ROOT:/onnxruntime_src" \
+    --volume "$BUILD_DIR:/build" \
+    --volume "$HOME/.cache/onnxruntime:/home/onnxruntimedev/.cache/onnxruntime" \
+    --volume "$HOME/.onnx:/home/onnxruntimedev/.onnx" \
+    -e NIGHTLY_BUILD \
+    "onnxruntime-$IMAGE" \
+    /bin/bash /onnxruntime_src/tools/ci_build/github/linux/run_build.sh \
+    -d $BUILD_DEVICE -x "$BUILD_EXTR_PAR" -o $BUILD_OS &
 wait $!
 
 EXIT_CODE=$?