配置torch、nccl环境遇到的问题

问题一描述

Tried loading libnccl.so.1 but got error /home/april/torch/install/share/lua/5.1/nccl/ffi.lua:192: libnccl.so.1: cannot open shared object file: No such file or directory
Tried loading libnccl.1.dylib but got error /home/april/torch/install/share/lua/5.1/nccl/ffi.lua:192: libnccl.1.dylib: cannot open shared object file: No such file or directory
| [en] Dictionary: 24738 types
| [de] Dictionary: 35474 types
Tried loading libnccl.so.1 but got error /home/april/torch/install/share/lua/5.1/nccl/ffi.lua:192: libnccl.so.1: cannot open shared object file: No such file or directory
Tried loading libnccl.1.dylib but got error /home/april/torch/install/share/lua/5.1/nccl/ffi.lua:192: libnccl.1.dylib: cannot open shared object file: No such file or directory
Tried loading libnccl.so.1 but got error /home/april/torch/install/share/lua/5.1/nccl/ffi.lua:192: libnccl.so.1: cannot open shared object file: No such file or directory

解决方案

分析问题：找不到libnccl的文件，所以要给系统指定路径
nccl源码下载后的路径为：/home/april/torch/nccl/nccl/
1.讲该路径添加到LD_LIBRARY_PATH中
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/home/april/torch/nccl/nccl/build/lib
2.更新torch的路径
source /home/april/torch/.bashrc

问题二描述

找不到epel-release-7-9.noarch.rpm的下载路径
方法：直接下载epel-release-7-9.noarch.rpm https://www.rpmfind.net/linux/rpm2html/search.php?query=epel-release
安装：rpm -ivh epel-release-6-8.noarch.rpm

问题三描述

配置nccl
git clone https://github.com/NVIDIA/nccl.git
cd nccl
export CUDA_HOME=/usr/local/cuda-8.0
make CUDA_HOME=${CUDA_HOME} test