This should fix this issue. Preliminary tests look like this works but I would appreciate any feedbacks.
Make sure to use cmake for now and add -DUSE_MPI=1. Also make sure to install OpenMPI and NCCL2 beforehand. Also perhaps should check the NCCL version and complain if USE_MPI does not have NCCL2.
The code is partly from: https://github.com/Caffe-MPI/Caffe-MPI.github.io which is an older fork of nvcaffe, without a traceable git history.
drnikolaev
commented
6 years ago
This should fix this issue. Preliminary tests look like this works but I would appreciate any feedbacks. Make sure to use cmake for now and add
-DUSE_MPI=1. Also make sure to install OpenMPI and NCCL2 beforehand. Also perhaps should check the NCCL version and complain if USE_MPI does not have NCCL2. The code is partly from: https://github.com/Caffe-MPI/Caffe-MPI.github.io which is an older fork of nvcaffe, without a traceable git history.