PGTransport: add inplace transport (3x faster) #119
Merged
+253
−81
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
facebook-github-bot
added
the
CLA Signed
fegin
reviewed
Feb 24, 2025
d4l3k
force-pushed
the
d4l3k/pg_inplace
branch
from
841ace6 to
c7f028f
Compare
d4l3k
force-pushed
the
d4l3k/pg_inplace
branch
from
c7f028f to
2f162a5
Compare
fegin
approved these changes
Feb 25, 2025
|
|
||
| for w in work: | ||
| w.wait(timeout) | ||
|
|
||
| def recv_checkpoint( | ||
| self, src_rank: int, metadata: str, step: int, timeout: timedelta | ||
| ) -> T: | ||
| state_dict = self._state_dict() if self._state_dict else {} | ||
| state_dict_leaves, _ = tree_flatten_with_path(state_dict) |
There was a problem hiding this comment.
Is tree_flatten_with_path a new one? Is it going to give you the FQN?
There was a problem hiding this comment.
It's been there for a few versions of torch -- it gives a path like:
(MappingKey(key='user'), MappingKey(key='optimizer'), MappingKey(key='state.layers.7.feed_forward.w2.weight.step'))
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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Outdated suggestions cannot be applied.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.
This adds a new
state_dictargument toPGTransportthat when provided will give a state_dict to use for doing in-place tensor operations. This has been measured at ~3x faster.Test plan:
Correctness:
The improvements have been measured via
pg_transport_bench.For inplace operation we see it at ~3x faster (15s -> 4-5s) for 12GB with 3MB tensors size. The remaining overhead is primarily from torchft ProcessGroupBaby queue communication and not proportional to the size of the tensors.
Reducing this overhead requires some careful consideration and will be addressed in a follow up PR.
inplace
12GB/3MB (4k tensors)
16KB/4B (4k tensors)
48GB/12MB (4k tensors)
not inplace
12GB/3MB (4k tensors)