These same instructions should work for native compilation on both x86_64 and aarch64 architectures.
To ensure everything is consistent for redistributable builds we build all redistributable components from source with a specific version of gcc.
You will need the following environment variables to be defined:
JAVA_HOME- Should point to the JDK you want to use to run Gradle.CPP_SRC_HOME- Only required if building the C++ code directly usingcmake, as Gradle sets it automatically.PATH- Must have/usr/local/gcc103/binbefore/usr/binand/bin.LD_LIBRARY_PATH- Must have/usr/local/gcc103/lib64and/usr/local/gcc103/libbefore/usr/liband/lib.
For example, you might create a .bashrc file in your home directory containing something like this:
umask 0002
export JAVA_HOME=/usr/local/jdk1.8.0_121
export LD_LIBRARY_PATH=/usr/local/gcc103/lib64:/usr/local/gcc103/lib:/usr/lib:/lib
export PATH=$JAVA_HOME/bin:/usr/local/gcc103/bin:/usr/local/cmake/bin:/usr/bin:/bin:/usr/sbin:/sbin:/home/vagrant/bin
# Only required if building the C++ code directly using cmake - adjust depending on the location of your Git clone
export CPP_SRC_HOME=$HOME/ml-cpp
You need the C++ compiler and the headers for the zlib library that comes with the OS. You also need the archive utilities unzip, bzip2 and xz. libffi-devel and openssl-devel are dependencies for building PyTorch. Finally, the unit tests for date/time parsing require the tzdata package that contains the Linux timezone database. On RHEL/CentOS these can be installed using:
sudo yum install bzip2 gcc-c++ libffi-devel openssl-devel texinfo tzdata unzip xz zlib-devel
On other Linux distributions the package names are generally the same and you just need to use the correct package manager to install these packages.
Most of the tools are built via a GNU "configure" script. There are some environment variables that affect the behaviour of this. Therefore, when building ANY tool on Linux, set the following environment variables:
export CFLAGS='-g -O3 -fstack-protector -D_FORTIFY_SOURCE=2 -msse4.2 -mfpmath=sse'
export CXX='g++ -std=gnu++17'
export CXXFLAGS='-g -O3 -fstack-protector -D_FORTIFY_SOURCE=2 -msse4.2 -mfpmath=sse'
export LDFLAGS='-Wl,-z,relro -Wl,-z,now'
export LDFLAGS_FOR_TARGET='-Wl,-z,relro -Wl,-z,now'
unset LIBRARY_PATH
For aarch64 replace -msse4.2 -mfpmath=sse with -march=armv8-a+crc+crypto.
These environment variables only need to be set when building tools on Linux. They should NOT be set when compiling the Machine Learning source code (as this should pick up all settings from our build system).
We have to build on old Linux versions to enable our software to run on the older versions of Linux that users have. However, this means the default compiler on our Linux build servers is also very old. To enable use of more modern C++ features, we use the default compiler to build a newer version of gcc and then use that to build all our other dependencies.
Download gcc-10.3.0.tar.gz from http://ftpmirror.gnu.org/gcc/gcc-10.3.0/gcc-10.3.0.tar.gz.
Unlike most automake-based tools, gcc must be built in a directory adjacent to the directory containing its source code, so build and install it like this:
tar zxvf gcc-10.3.0.tar.gz
cd gcc-10.3.0
contrib/download_prerequisites
sed -i -e 's/$(SHLIB_LDFLAGS)/-Wl,-z,relro -Wl,-z,now $(SHLIB_LDFLAGS)/' libgcc/config/t-slibgcc
cd ..
mkdir gcc-10.3.0-build
cd gcc-10.3.0-build
unset CXX
unset LD_LIBRARY_PATH
export PATH=/usr/bin:/bin:/usr/sbin:/sbin
../gcc-10.3.0/configure --prefix=/usr/local/gcc103 --enable-languages=c,c++ --enable-vtable-verify --with-system-zlib --disable-multilib
make -j 6
sudo make install
It's important that gcc itself is built using the system compiler in C++98 mode, hence the adjustment to PATH and unsetting of CXX and LD_LIBRARY_PATH.
After the gcc build is complete, if you are going to work through the rest of these instructions in the same shell remember to reset the CXX environment variable so that the remaining C++ components get built with C++17:
export CXX='g++ -std=gnu++17'
To confirm that everything works correctly run:
g++ --version
It should print:
g++ (GCC) 10.3.0
in the first line of the output. If it doesn't then double check that /usr/local/gcc103/bin is near the beginning of your PATH.
Also due to building on old Linux versions yet wanting to use modern libraries we have to install an up-to-date version of binutils.
Download binutils-2.37.tar.bz2 from http://ftpmirror.gnu.org/binutils/binutils-2.37.tar.bz2.
Uncompress and untar the resulting file. Then run:
./configure --prefix=/usr/local/gcc103 --enable-vtable-verify --with-system-zlib --disable-libstdcxx --with-gcc-major-version-only
This should build an appropriate Makefile. Assuming it does, type:
make
sudo make install
to install.
Modern versions of Linux will come with Git in their package repositories, and (since we're not redistributing it so don't really care about the exact version used) this is the easiest way to install it. The command will be:
sudo yum install git
on RHEL clones. However, shallow clones do not work correctly before version 1.8.3 of Git, so if the version that yum installs is older you'll still have to build it from scratch. In this case, you may need to uninstall the version that yum installed:
git --version
sudo yum remove git
If you have to build Git from source in order to get version 1.8.3 or above, this is what to do:
Make sure you install the packages python-devel, curl-devel and openssl-devel using yum or similar before you start this.
Start by running:
./configure
as usual.
Then run:
make prefix=/usr all
sudo make prefix=/usr install
Without the prefix=/usr bit, you'll end up with a personal Git build in ~/bin instead of one everyone on the machine can use.
Download libxml2-2.9.14.tar.xz from https://download.gnome.org/sources/libxml2/2.9/libxml2-2.9.14.tar.xz.
Uncompress and untar the resulting file. Then run:
./configure --prefix=/usr/local/gcc103 --without-python --without-readline
This should build an appropriate Makefile. Assuming it does, type:
make
sudo make install
to install.
Download version 1.83.0 of Boost from https://boostorg.jfrog.io/artifactory/main/release/1.83.0/source/boost_1_83_0.tar.bz2. You must get this exact version, as the Machine Learning build system requires it.
Assuming you chose the .bz2 version, extract it to a temporary directory:
bzip2 -cd boost_1_83_0.tar.bz2 | tar xvf -
In the resulting boost_1_83_0 directory, run:
./bootstrap.sh --without-libraries=context --without-libraries=coroutine --without-libraries=graph_parallel --without-libraries=mpi --without-libraries=python --without-icu
This should build the b2 program, which in turn is used to build Boost.
Edit boost/unordered/detail/prime_fmod.hpp and change line 134 from:
(13ul)(29ul)(53ul)(97ul)(193ul)(389ul)(769ul)(1543ul)(3079ul)(6151ul)( \
to:
(3ul)(13ul)(29ul)(53ul)(97ul)(193ul)(389ul)(769ul)(1543ul)(3079ul)(6151ul)( \
Finally, run:
./b2 -j6 --layout=versioned --disable-icu pch=off optimization=speed inlining=full define=BOOST_MATH_NO_LONG_DOUBLE_MATH_FUNCTIONS define=BOOST_LOG_WITHOUT_DEBUG_OUTPUT define=BOOST_LOG_WITHOUT_EVENT_LOG define=BOOST_LOG_WITHOUT_SYSLOG define=BOOST_LOG_WITHOUT_IPC define=_FORTIFY_SOURCE=2 cxxflags='-std=gnu++17 -fstack-protector -msse4.2 -mfpmath=sse' cflags='-D__STDC_FORMAT_MACROS' linkflags='-std=gnu++17 -Wl,-z,relro -Wl,-z,now'
sudo env PATH="$PATH" LD_LIBRARY_PATH="$LD_LIBRARY_PATH" ./b2 install --prefix=/usr/local/gcc103 --layout=versioned --disable-icu pch=off optimization=speed inlining=full define=BOOST_MATH_NO_LONG_DOUBLE_MATH_FUNCTIONS define=BOOST_LOG_WITHOUT_DEBUG_OUTPUT define=BOOST_LOG_WITHOUT_EVENT_LOG define=BOOST_LOG_WITHOUT_SYSLOG define=BOOST_LOG_WITHOUT_IPC define=_FORTIFY_SOURCE=2 cxxflags='-std=gnu++17 -fstack-protector -msse4.2 -mfpmath=sse' cflags='-D__STDC_FORMAT_MACROS' linkflags='-std=gnu++17 -Wl,-z,relro -Wl,-z,now'
to install the Boost headers and libraries. (Note the env PATH="$PATH" bit in the install command - this is because sudo usually resets PATH and that will cause Boost to rebuild everything again with the default compiler as part of the install!)
For aarch64 replace -msse4.2 -mfpmath=sse with -march=armv8-a+crc+crypto.
Obtain patchelf from https://github.com/NixOS/patchelf/releases/download/0.13/patchelf-0.13.tar.bz2.
Extract it to a temporary directory using:
bzip2 -cd patchelf-0.13.tar.bz2 | tar xvf -
In the resulting patchelf-0.13.20210805.a949ff2 directory, run the:
./configure --prefix=/usr/local/gcc103
script. This should build an appropriate Makefile. Assuming it does, run:
make
sudo make install
to complete the build.
CMake version 3.19.2 is the minimum required to build ml-cpp. Download version 3.23.2 from https://github.com/Kitware/CMake/releases/download/v3.23.2/cmake-3.23.2-Linux-x86_64.sh and install:
chmod +x cmake-3.23.2-Linux-x86_64.sh
sudo mkdir /usr/local/cmake
sudo ./cmake-3.23.2-Linux-x86_64.sh --skip-license --prefix=/usr/local/cmake
Please ensure /usr/local/cmake/bin is in your PATH environment variable.
Python 3.10 requires OpenSSL 1.1.1. No other version is acceptable.
If the openssl-devel package for your distribution happens to be version 1.1.1 then you can skip this step. Otherwise, you need to build OpenSSL 1.1.1 from source.
Download openssl-1.1.1q.tar.gz from https://www.openssl.org/source/old/1.1.1/openssl-1.1.1q.tar.gz, then build as follows:
tar zxvf openssl-1.1.1q.tar.gz
cd openssl-1.1.1q
./Configure --prefix=/usr/local/gcc103 shared linux-`uname -m`
make
sudo make install
PyTorch currently requires Python 3.7 or higher; we use version 3.10. If your system does not have a requisite version of Python install it with a package manager or build the last 3.10 release from source by downloading Python-3.10.9.tgz from https://www.python.org/ftp/python/3.10.9/Python-3.10.9.tgz then extract as follows:
tar xzf Python-3.10.9.tgz
cd Python-3.10.9
If the distribution you are building on uses OpenSSL 1.1.1 as its built in OpenSSL version then configure as follows:
./configure --prefix=/usr/local/gcc103 --enable-optimizations
If you had to build OpenSSL 1.1.1 yourself then on x86_64 configure like this:
sed -i -e 's~ssldir/lib~ssldir/lib64~' configure
./configure --prefix=/usr/local/gcc103 --enable-optimizations --with-openssl=/usr/local/gcc103 --with-openssl-rpath=/usr/local/gcc103/lib64
or on aarch64 configure like this:
./configure --prefix=/usr/local/gcc103 --enable-optimizations --with-openssl=/usr/local/gcc103 --with-openssl-rpath=/usr/local/gcc103/lib
Finally, build as follows:
make
sudo make altinstall
Skip this step if you are building on aarch64.
Intel Maths Kernel Library is an optimized BLAS library which greatly improves the performance of PyTorch CPU inference when PyTorch is built with it.
tee > /tmp/oneAPI.repo << EOF
[oneAPI]
name=Intel oneAPI repository
baseurl=https://yum.repos.intel.com/oneapi
enabled=1
gpgcheck=1
repo_gpgcheck=1
gpgkey=https://yum.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
EOF
sudo cp /tmp/oneAPI.repo /etc/yum.repos.d
sudo yum -y install intel-oneapi-mkl-devel-2024.0
The process is different for distributions that use other package managers, for example APT. More instructions can be found at: https://www.intel.com/content/www/us/en/developer/tools/oneapi/onemkl-download.html?operatingsystem=linux
Then copy the shared libraries to the system directory:
(cd /opt/intel/oneapi/mkl/2024.0 && tar cf - lib) | (cd /usr/local/gcc103 && sudo tar xvf -)
(This step requires a reasonable amount of memory. It failed on a machine with 8GB of RAM. It succeeded on a 16GB machine. You can specify the number of parallel jobs using environment variable MAX_JOBS. Lower number of jobs will reduce memory usage.)
PyTorch requires that certain Python modules are installed. Install these modules with pip using the same Python version you will build PyTorch with. If you followed the instructions above and built Python from source use python3.10:
sudo /usr/local/gcc103/bin/python3.10 -m pip install install numpy ninja pyyaml setuptools cffi typing_extensions future six requests dataclasses
For aarch64 the ninja module is not available, so use:
sudo /usr/local/gcc103/bin/python3.10 -m pip install install numpy pyyaml setuptools cffi typing_extensions future six requests dataclasses
Then obtain the PyTorch code:
git clone --depth=1 --branch=v2.3.1 git@github.com:pytorch/pytorch.git
cd pytorch
git submodule sync
git submodule update --init --recursive
Edit torch/csrc/jit/codegen/fuser/cpu/fused_kernel.cpp and replace all
occurrences of system( with strlen(. This file is used to compile
fused CPU kernels, which we do not expect to be doing and never want to
do for security reasons. Replacing the calls to system() ensures that
a heuristic virus scanner looking for potentially dangerous function
calls in our shipped product will not encounter these functions that run
external processes.
Build as follows:
[ $(uname -m) = x86_64 ] && export BLAS=MKL || export BLAS=Eigen
export BUILD_TEST=OFF
[ $(uname -m) = x86_64 ] && export BUILD_CAFFE2=OFF
[ $(uname -m) != x86_64 ] && export USE_FBGEMM=OFF
[ $(uname -m) != x86_64 ] && export USE_KINETO=OFF
[ $(uname -m) = x86_64 ] && export USE_NUMPY=OFF
export USE_DISTRIBUTED=OFF
export USE_MKLDNN=ON
export USE_QNNPACK=OFF
export USE_PYTORCH_QNNPACK=OFF
[ $(uname -m) = x86_64 ] && export USE_XNNPACK=OFF
export PYTORCH_BUILD_VERSION=2.3.1
export PYTORCH_BUILD_NUMBER=1
/usr/local/gcc103/bin/python3.10 setup.py install
Once built copy headers and libraries to system directories:
sudo mkdir -p /usr/local/gcc103/include/pytorch
sudo cp -r torch/include/* /usr/local/gcc103/include/pytorch/
sudo cp torch/lib/libtorch_cpu.so /usr/local/gcc103/lib
sudo cp torch/lib/libc10.so /usr/local/gcc103/lib
valgrind is not required to build the code. However, since we build gcc ourselves, if you want
to debug with valgrind then you'll get better results if you build a version that's compatible
with our gcc instead of using the version you can get via your package manager.
If you find yourself needing to do this, download valgrind from http://valgrind.org/downloads/ - the
download file will be valgrind-3.17.0.tar.bz2.
Extract it to a temporary directory using:
tar jxvf valgrind-3.17.0.tar.bz2
In the resulting valgrind-3.17.0 directory, run:
unset CFLAGS
unset CXXFLAGS
./configure --prefix=/usr/local/gcc103 --disable-dependency-tracking --enable-only64bit
The reason for unsetting the compiler flags is that valgrind does not build correctly
with the fortified options we have to use for libraries we ship.
This should build an appropriate Makefile. Assuming it does, run:
make
sudo make install
to complete the build.