Building from source

Compiler version

You must use GCC v5.4 or newer. RaNNC has been tested with GCC v8.2. Note that RaNNC must be built in compliance with ABI of PyTorch.

RaNNC is built with Pre-cxx11 ABI (_GLIBCXX_USE_CXX11_ABI=0) following the binary packages of PyTorch. You can change the ABI setting in CMakeLists.txt. PyTorch provides you with function below to understand how the binary is compiled.

Build and Install

Clone the repository with the submodules (--recursive is required).

git clone --recursive https://github.com/nict-wisdom/rannc

You need to set some environment variables before building RaNNC to help cmake find the dependent libraries.

Variables for building configurations

Variable
CUDA_HOME	Path to a CUDA runtime directory.
MPI_DIR	Path to an MPI installation directory.
BOOST_DIR	Path to a Boost libraries directory.
CUDNN_ROOT_DIR	Path to a cuDNN libraries directory.
LD_LIBRARY_PATH	Must contain the path to NCCL library directory.

The building process refers to PyTorch installed with conda. Therefore, install PyTorch using your python and run setup.py. The following script shows the configurations used to install RaNNC from the source.

#!/usr/bin/env bash

 # Activate conda
 source [CONDA_PATH]/etc/profile.d/conda.sh
 conda activate rannc

 # Set dependencies
 export CUDA_HOME="$(dirname $(which nvcc))/../"
 export MPI_DIR="$(dirname $(which ompi_info))/../"
 export BOOST_DIR=[BOOST_DIR_PATH]
 export CUDNN_ROOT_DIR=[YOUR_CUDNN_DIR_PATH]

 python setup.py build -g install

Makefiles under docker/ show the complete process used to build and install RaNNC. These are used to build the pip packages.