Flash-attn performance: remove cuda sync during inference #33570
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
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
|
cc @ArthurZucker, forgot to ping you |
ArthurZucker
approved these changes
Sep 20, 2024
NielsRogge
pushed a commit
to NielsRogge/transformers
that referenced
this pull request
Oct 21, 2024
…e#33570) Switch conditions to use short-circuit during inference
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Outdated suggestions cannot be applied.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.
What does this PR do?
#31629 & #32241 introduced a functionality in FA2 intended for training efficiency. However, it adds unnecessary cuda synchronization at inference time in every forward pass due to always checking
(torch.diff(position_ids, dim=-1) >= 0).all()in theelifcondition. This PR fixes the performance issue by simply switching the order of the different checks in theelifcondition, to make good use of Python's default short-circuit evaluation. Indeed, at inference time,query_lengthwill always be 1 except during prefill, thus we will short-circuit torch synchronization all the time.Performance degradation was not so significant, but this PR allows to win back around 5-10% speed at inference time from the quick tests I ran.