This folder contains scripts for building the dependencies from source. The static libraries from the build artifacts can be used to create self-contained shared object for mxnet through static linking.
The scripts use the following environment variables for setting behavior:
DEPS_PATH: the location in which the libraries are downloaded, built, and installed. PLATFORM: name of the OS in lower case. Supported options are ‘linux’ and ‘darwin’.
It also expects the following build tools in path: make, cmake, tar, unzip, autoconf, nasm
Currently, we only support gcc-4.8 build. It's your own choice to use a higher version of gcc. Please make sure your gcc, g++ and gfortran always have the same version in order to eliminate build failure.
This issue appeared in the OSX build with XCode version 8.0 above (reproduced on 9.2). Please add the following build flag in curl.sh if your XCode version is more than 8.0:
--without-libidn2
MXNet is built on top of many dependencies. Managing these dependencies could be a big headache. This goal of this document is to give a overview of those dependencies and how to upgrade when new version of those are rolled out.
The dependencies could be categorized by several groups: BLAS libraries, CPU-based performance boost library, i.e. oneDNN and GPU-based performance boosting library including CUDA, cuDNN, NCCL. and others including OpenCV, Numpy, S3-related, PS-lite dependencies. The list below shows all the dependencies and their version. Except for CUDA, cuDNN, NCCL which the user is required to install on their environments, we statically link those dependencies into libmxnet.so when we build PyPi package. By doing this, the user can take advantage of these dependencies without being worry about it.
| Dependencies | MXNet Version |
|---|---|
| OpenBLAS | 0.3.9 |
| oneDNN | 2.6 |
| CUDA | 10.1 |
| cuDNN | 7.5.1 |
| NCCL | 2.4.2 |
| numpy | >1.16.0,<2.0.0 |
| request | >=2.20.0,< 3.0.0 |
| graphviz | <0.9.0,>=0.8.1 |
| OpenCV | 3.4.2 |
| zlib | 1.2.6 |
| libjpeg-turbo | 2.0.2 |
| libpng | 1.6.35 |
| libtiff | 4-0-10 |
| eigen | 3.3.4 |
| libcurl | 7.61.0 |
| libssl-dev | 1.1.1b |
| zmq | 4.2.2 |
| protobuf | 3.5.1 |
| lz4 | r130 |
| cityhash | 1.1.1 |
sudo apt update
sudo apt-get install -y git \
cmake \
libcurl4-openssl-dev \
unzip \
gcc-4.8 \
g++-4.8 \
gfortran \
gfortran-4.8 \
binutils \
nasm \
libtool \
curl \
wget \
sudo \
gnupg \
gnupg2 \
gnupg-agent \
pandoc \
python3-pip \
automake \
pkg-config
@pengzhao-intel (https://github.com/apache/mxnet/commits?author=pengzhao-intel) and his team are tracking and updating these versions. Kudos to them!
We will install all the prerequsite software. We demonstrate with CUDA10/cuDNN7.5/NCCL 2.4.2. You might want to change these versions to suit your needs.
# CUDA installation
# Take CUDA 10 for example, please follow the instructions on https://developer.nvidia.com/cuda-downloads
# Install NVIDIA Accelerated Graphics Driver for Linux-x86_64 410.48?
# (y)es/(n)o/(q)uit: y
#
# Do you want to install the OpenGL libraries?
# (y)es/(n)o/(q)uit [ default is yes ]:
#
# Do you want to run nvidia-xconfig?
# This will update the system X configuration file so that the NVIDIA X driver
# is used. The pre-existing X configuration file will be backed up.
# This option should not be used on systems that require a custom
# X configuration, such as systems with multiple GPU vendors.
# (y)es/(n)o/(q)uit [ default is no ]:
#
# Install the CUDA 10.0 Toolkit?
# (y)es/(n)o/(q)uit: y
#
# Enter Toolkit Location
# [ default is /usr/local/cuda-10.0 ]:
#
# Do you want to install a symbolic link at /usr/local/cuda?
# (y)es/(n)o/(q)uit: y
#
# Install the CUDA 10.0 Samples?
# (y)es/(n)o/(q)uit: n
# Set LD_LIBRARY_PATH
export LD_LIBRARY_PATH=/usr/local/cuda/lib64:${LD_LIBRARY_PATH}
# Check installation
nvidia-smi
# cuDNN Setup
# Take cuDNN 7.5.0 with CUDA 10 for example
# https://developer.nvidia.com/rdp/cudnn-download
# Register with NVIDIA and download cudnn-10.0-linux-x64-v7.5.0.56.tgz
# scp it to your instance
# https://docs.nvidia.com/deeplearning/sdk/cudnn-install/index.html
tar -xvzf cudnn-10.0-linux-x64-v7.5.0.56.tgz
sudo cp cuda/include/cudnn.h /usr/local/cuda/include
sudo cp cuda/lib64/libcudnn* /usr/local/cuda/lib64
sudo chmod a+r /usr/local/cuda/include/cudnn.h /usr/local/cuda/lib64/libcudnn*
# Check cuDNN version
cat /usr/local/cuda/include/cudnn.h | grep CUDNN_MAJOR -A 2
# #define CUDNN_MAJOR 7
# #define CUDNN_MINOR 5
# #define CUDNN_PATCHLEVEL 0
# --
# #define CUDNN_VERSION (CUDNN_MAJOR * 1000 + CUDNN_MINOR * 100 + CUDNN_PATCHLEVEL)
#
# #include "driver_types.h"
# install NCCL
# take NCCL 2.4.2 for example
# https://developer.nvidia.com/nccl/nccl2-download-survey
# Register with NVIDIA and download nccl-repo-ubuntu1604-2.4.2-ga-cuda10.0_1-1_amd64.deb
sudo dpkg -i nccl-repo-ubuntu1604-2.4.2-ga-cuda10.0_1-1_amd64.deb
sudo apt-key add /var/nccl-repo-2.4.2-ga-cuda10.0/7fa2af80.pub
sudo apt update
sudo apt install libnccl2 libnccl-dev
# we will check the NCCL version later
We will build MXNet with statically linked dependencies.
# Clone MXNet repo
git clone --recursive https://github.com/apache/mxnet.git
cd mxnet
# Make sure you pin to specific commit for all the performance sanity check to make fair comparison
# Make corresponding change on tools/setup_gpu_build_tools.sh
# to upgrade CUDA version, please refer to PR #14887.
# Make sure you add new makefile and right debs CUDA uses on the website
# http://developer.download.nvidia.com/compute/cuda/repos/ubuntu1604/x86_64/
# Build PyPi package
tools/staticbuild/build.sh cu100mkl
# Wait for 10 - 30 mins, you will find libmxnet.so under the mxnet/lib
# Install python frontend
pip install -e python
# Test MXNet
>>> import mxnet as mx
>>> mx.runtime.feature_list()
# Test NCCL version
export NCCL_DEBUG=VERSION
vim tests/python/gpu/test_nccl.py
# Remove @unittest.skip("Test requires NCCL library installed and enabled during build") then run
pytest --verbose tests/python/gpu/test_nccl.py
# test_nccl.test_nccl_pushpull ... NCCL version 2.4.2+cuda10.0
# ok
# ----------------------------------------------------------------------
# Ran 1 test in 67.666s
OK
Please run performance test aginast the MXNet you build before raising the PR.
# Compile mxnet to get libmxnet.so pip install -e python python >>> import mxnet as mx >>> mx.nd.ones((1, 2)) [[1. 1.]] <NDArray 1x2 @cpu(0)>
tools/dependencies and update the version.# Take Ubuntu 16.04 for example sudo apt update sudo apt-get install -y git \ cmake \ libcurl4-openssl-dev \ unzip \ gcc-4.8 \ g++-4.8 \ gfortran \ gfortran-4.8 \ binutils \ nasm \ libtool \ curl \ wget \ sudo \ gnupg \ gnupg2 \ gnupg-agent \ pandoc \ python3-pip \ automake \ pkg-config
# Update the dependency under tools/dependencies, then tools/staticbuild/build.sh mkl # Wait for 10 - 30 mins, you will find libmxnet.so under the mxnet/lib # Install python frontend pip install -e python # Test MXNet >>> import mxnet as mx >>> mx.runtime.feature_list()
tools/dependencies and update the version.# Take Ubuntu 16.04 for example sudo apt update sudo apt-get install -y git \ cmake \ libcurl4-openssl-dev \ unzip \ gcc-4.8 \ g++-4.8 \ gfortran \ gfortran-4.8 \ binutils \ nasm \ libtool \ curl \ wget \ sudo \ gnupg \ gnupg2 \ gnupg-agent \ pandoc \ python3-pip \ automake \ pkg-config
# Update the dependency under tools/dependencies, then tools/staticbuild/build.sh mkl # Wait for 10 - 30 mins, you will find libmxnet.so under the mxnet/lib # Install python frontend pip install -e python # Test MXNet >>> import mxnet as mx >>> mx.runtime.feature_list()