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Outdated suggestions cannot be applied.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
[fix] OSS - enforce cuda parameters for state consolidation if NCCL backend #573
[fix] OSS - enforce cuda parameters for state consolidation if NCCL backend #573
Changes from 2 commits
be4c7f6
b97ae7e
05e8cd8
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.
no choice with NCCL, needs to be cuda
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.
if the model is moved back to the cpu and the optimizer state reflects it, why do we call broadcast? The optimizer state is not sharded anymore right? Maybe i am missing something.
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.
the framework is the one calling .consolidate(), it can do so at any time basically. We could add a skip mechanism for when it's called twice in a row (would be even more foolproof actually), but that would not solve the case of train -> move to cpu -> call .consolidate(), which can be legitimate, if unfortunate
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.
(complement) the issue was that if the model is moved to cpu, then some tensors in the optimizer dict are cpu. When consolidating the shards are exchanged towards a specific rank (or all), which breaks with NCCL since it always expects cuda for communication primitives
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.
This seems wasteful. Why not skip the broadcast in this case instead of sending a zero? In the
else
below you could check ifrank != recipient
.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.
without the fix, this unit test does fail with the same error that the user mentioned