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

[WIP] Add all gather op and all reduce op #9713

Conversation

chengduoZH
Copy link
Contributor

@chengduoZH chengduoZH commented Apr 7, 2018

This PR attempt to resolve the sparse gradient on multi-cards.

@chengduoZH chengduoZH force-pushed the feature/add_all_reduce_and_all_gather_op branch from 7f595aa to 22a064a Compare April 7, 2018 14:37
@chengduoZH chengduoZH force-pushed the feature/add_all_reduce_and_all_gather_op branch from 22a064a to 231a219 Compare April 7, 2018 15:26
@panyx0718
Copy link
Contributor

Briefly discussed. Customized gather and reduce might allow us to fully overlap computation with memcpy. It also allows us to do customized sparse gradient aggregation. Nice!

@reyoung
Copy link
Collaborator

reyoung commented Apr 8, 2018

I do not believe that implementation is faster than NCCL, especially for dense parameters. Also, I think AllReduce operators must contain GPU kernels. It cannot be implemented by only memcpies.

@chengduoZH chengduoZH force-pushed the feature/add_all_reduce_and_all_gather_op branch 2 times, most recently from f78783d to 5f337c9 Compare April 8, 2018 07:43
@chengduoZH chengduoZH force-pushed the feature/add_all_reduce_and_all_gather_op branch from 5f337c9 to 1e29c2e Compare April 8, 2018 09:40
@chengduoZH
Copy link
Contributor Author

The PR did not reach the expectation, so it can be turned off.

@chengduoZH chengduoZH closed this Apr 13, 2018
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

Successfully merging this pull request may close these issues.

3 participants