To Top
首页 > 常用平台 > 正文

horovod

标签:horovod


目录

https://juejin.im/post/5cbc6dbd5188253236619ccb#heading-9

参考[深度学习] 分布式Horovod介绍(四)

主页:

https://github.com/uber/horovod

安装

有一个装了tf的py

例如:

/home/work/tools/python-2.7.8-tf1.4-gpu

安装openmpi

下载地址: https://www.open-mpi.org/software/ompi/v3.0/

解压后

./configure --prefix=/home/work/tools/openmpi
make && make install

pip安装

确保PATH里有gcc48,以及/home/work/tools/openmpi/bin/

如果是GPU的,确保export LD_LIBRARY_PATH=/home/work/cudnnv6/cuda/lib64/:$LD_LIBRARY_PATH

因为安装时要用到-lpython2.7,如果没有root权限,去报错的gcc命令里找-L的路径,发现最简单粗暴的方法就是

cp  /home/work/tools/python-2.7.8-tf1.4-gpu/lib/libpython2.7.so* /home/work/tools/openmpi/lib

或者

cp /home/work/tools/python-2.7.8-tf1.4-gpu/lib/ /home/work/tools/python-2.7.8-tf1.4-gpu/lib/python2.7/site-packages/tensorflow

然后pip安装:

/home/work/tools/python-2.7.8-tf1.4-gpu/bin/python  /home/work/tools/python-2.7.8-tf1.4-gpu/bin/pip install horovod

原创文章,转载请注明出处!
本文链接:http://daiwk.github.io/posts/platform-horovod.html
上篇: gan with the wind
下篇: alphago-zero

comment here..