cache enc kv proj for cross-attention#106
Open
tingqli wants to merge 1 commit intoFireRedTeam:mainfrom
Open
Conversation
Collaborator
|
Thanks for your PR, we will review. |
super-yuki
reviewed
Jan 20, 2026
| self.w_ks(k).view(bs, -1, self.n_head, self.d_k), | ||
| self.w_vs(v).view(bs, -1, self.n_head, self.d_k) | ||
| ) | ||
| k,v = self.kv_proj |
There was a problem hiding this comment.
layer 0 and layer 1 has different linear weight, so u should calculate every weight and save in self.kv_proj
This file contains hidden or 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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
The kv-projection in cross-attention is calculated in every decoding step which is redundant since encoder_outputs doesn't change during whole decoding phase, this PR add a simple caching mechanism in cross-attn to avoid recomputing. in my test case (batch-size=32 beam-size=3 audio-length=20s), the e2e latency reduced from 20seconds to 6.1 seconds on H20.