Fix attention caching to make transcription run 30% faster #370
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Outdated suggestions cannot be applied.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.
For a
dict.get(a, f())
call,f()
is run even when thedict
already hasa
. This is causing the bug in attention caching logic.So making this change fixes the caching logic, and gets a sweet (expected) speedup of close to 30%.
For a close to 5 min test audio, on an Nvidia Quadro RTX 8000 gpu:
large
model took 102.8s, and after this fix it takes 68.9smedium.en
model took 55.1s, and after this fix it takes 39.7s