Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
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
Add a way to configure the required process group structure for a model #153
Add a way to configure the required process group structure for a model #153
Changes from 2 commits
3f2f06c
e115dda
63bc4ef
File filter
Filter by extension
Conversations
Jump to
There are no files selected for viewing
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I really like this Akshay, I'm mainly curious of how the distributed manager will know what groups to make internode vs intranode?
For example in the example you gave theres three groups ("channel_parallel", "spatial_parallel", and "data_parallel"). How would we know to stick the first two on the same node?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Currently there isn't such a way. In your example, you'd have a
model_parallel
group withdata_parallel
being the orthogonal group.model_parallel
would havechannel_parallel
andspatial_parallel
child groups. When we create the process groups, we first create themodel_parallel
group which would use a contiguous set of ranks. So if the totalmodel_parallel
group size was 8, that would be on a single node. See #188 for the specifics of how this works.But I want to emphasize that this is just the logical group config. It is up to the
DistributedManager
to actually instantiate this group config and that's where this kind of topology optimization needs to come in. Right now it is this hard coded topology mapping, but there isn't a reason we cannot add the ability for a user to specify a custom topology mapping when passing this config to theDistributedManager
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I think the MR looks good but Nick is right: there needs to be some notion of what is close and what is not. So basically an order of initialization. So for example the leaf node list should be ordered, then it should be clear.