8000 GitHub - rbpittman/horovod: Distributed training framework for TensorFlow.
[go: up one dir, main page]
More Web Proxy on the site http://driver.im/
Skip to content

rbpittman/horovod

 
 

Repository files navigation

Horovod Groups

In this package, the simple installation for Horovod (no GPU/NCCL support) has been modified to support custom MPI groups, as well as a Gather operation.

Example

In order to create 2 groups, one containing ranks 0, 1, and 2, and another group containing ranks 2, 3, and 4, initialize HVD with:

hvd.init([[0,1,2], [2,3,4]])

A subsequent allgather for group 0 might look like:

all_images = hvd.allgather(tensor, group=0, name="my_allgather")

About

Distributed training framework for TensorFlow.

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages

  • C++ 80.1%
  • Python 19.8%
  • Shell 0.1%
0