Skip to content

Latest commit

 

History

History
56 lines (41 loc) · 1.78 KB

build.rst

File metadata and controls

56 lines (41 loc) · 1.78 KB
Raw

Building from source

Compiler version

You must use GCC v5.4 or newer. We tested RaNNC with GCC v5.4 and v7.1. Note that, however, RaNNC must be built complying ABI of PyTorch.

RaNNC is built with Pre-cxx11 ABI (_GLIBCXX_USE_CXX11_ABI=0) as default because PyTorch installed via conda is built with Pre-cxx11 ABI. You can change the ABI setting in CMakeLists.txt. PyTorch provides you with a function below to know how the binary is compiled.

Build and Install

You need to set some environment variables before building RaNNC to help cmake find 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 lib directory.

The building process refers to PyTorch installed with conda. Therefore, install PyTorch with your python and run setup.py. The following script shows configurations 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 to build and install RaNNC. They are used to build pip packages.