how to enable multi-gpu training in each node

Question

how to enable multi-gpu training in each node

Dav-Jay opened this issue 8 years ago · 2 comments

Thanks for the excellent work!

When I tried the provided train_inception.sh for CIFAR10 on two nodes, I found only one GPU is invoked in each node (GPU 0), is there a way to use multiple gpus in each node? In original Caffe, I can use --gpu=0,1,2,3 which seems not work in geeps.

Answer 1 · 2016-12-02T08:10:25.000Z

Seems current geeps does not support multi-gpu in each node?

Answer 2 · 2016-12-14T19:05:19.000Z

Unfortunately, the current GeePS implementation does not support multi-gpu in each node. Feel free to implement that and create a pull request :)