Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

how do I get the ordered list of reduction ops to do the layer-wise all-to-all reducetion #5

Open
yogurfrul opened this issue Nov 15, 2017 · 1 comment

Comments

@yogurfrul
Copy link

I ‘confused with the MPI Allreduce Operator
the paper said that MaTEx-TensorFlow use the allreduce ops to synchronize each layer across ranks.
I think one AllReduce op is to reduce all the gradients generated one the same layer(also returns at the same time )
Only got the graph defined by user, I do not know which layer one node belongs to
so how do I know which gradients are on the same layer?
I mean how do I get the ordered list of reduction ops to do the layer-wise all-to-all reducetion.
I have not find some code do this work in this resouce

"Example of a MaTEx TensorFlow executing on four MPI ranks. Each rank will run a model replica and communicate
at each of the reduction points (i.e. the orange bars). Each model is initialized identically due to the broadcast operator at the
beginning (i.e the blue bar)."
https://www.researchgate.net/profile/Abhinav_Vishnu/publication/316184213/figure/fig2/AS:484334135189506@1492485664993/Fig-2-Example-of-a-MaTEx-TensorFlow-executing-on-four-MPI-ranks-Each-rank-will-run-a.png

@charlesmsiegel
Copy link
Collaborator

MaTEx-Tensorflow handles all of this automatically on the back end, at the time the gradients are computed. There is no need for sample code to use the all-to-all reduction, as the user needs to make no code changes in setting up the Tensorflow graph.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants