Reserve KV cache capacity after the first model run #408
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
Hugging Face models with separate branches for the first and subsequent iterations do not use the input KV cache buffer on the first run. Thus they did not benefit from the pre-allocated capacity and ended up re-allocating a new KV cache buffer on each run. To resolve this, change the KV cache growth strategy to grow the buffer after the model runs, if the capacity limit has been reached. Also replace the hard-coded capacity with a growth strategy that doubles the capacity each time. This amortizes the costs of copying the old KV cache into the new buffer.
robertknight
force-pushed
the
kv-cache-grow-after-run
branch
from
b239c1d
to
c24e15f
Compare
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Outdated suggestions cannot be applied.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.
Hugging Face models with separate branches for the first and subsequent iterations do not use the input KV cache buffer on the first run. Thus they did not benefit from the pre-allocated capacity and ended up re-allocating a new KV cache buffer on each run.
To resolve this, change the KV cache growth strategy to grow the buffer after the model runs, if the capacity limit has been reached. Also replace the hard-coded capacity with a growth strategy that doubles the capacity each time. This amortizes the costs of copying the old KV cache into the new buffer.
This impacts the Whisper and TrOCR examples. For the Whisper example, this reduced time for transcribing a 2-minute audio clip with the base model from ~6.1s to ~5.7s (~6%). With the larger models the benefit will be greater since the caches are larger.