Skip to content

[Executorch][llm] Add ring buffer based kv cache and mask calculation to MHA #10609

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 3 commits into
base: gh/kimishpatel/186/base
Choose a base branch
from

Conversation

kimishpatel
Copy link
Contributor

@kimishpatel kimishpatel commented May 1, 2025

Stack from ghstack (oldest at bottom):

Leveraging previous work now we allow MHA to have ring buffer cache. If ring buffer cache is used
then we query the mask from kv cache and use that for sdpa instead of using precalculated mask.

In this process we had to adjsut ring buffer implementation to allow keeping the context of
full sliding window. See code for comment.

Differential Revision: D73891425

… to MHA

Leveraging previous work now we allow MHA to have ring buffer cache. If ring buffer cache is used
then we query the mask from kv cache and use that for sdpa instead of using precalculated mask.

In this process we had to adjsut ring buffer implementation to allow keeping the context of
full sliding window. See code for comment.

Differential Revision: [D73891425](https://our.internmc.facebook.com/intern/diff/D73891425/)

[ghstack-poisoned]
Copy link

pytorch-bot bot commented May 1, 2025

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/10609

Note: Links to docs will display an error until the docs builds have been completed.

❌ 1 New Failure, 1 Unrelated Failure

As of commit 7d3fdd7 with merge base 1ae8c2c (image):

NEW FAILURE - The following job has failed:

BROKEN TRUNK - The following job failed but were present on the merge base:

👉 Rebase onto the `viable/strict` branch to avoid these failures

This comment was automatically generated by Dr. CI and updates every 15 minutes.

@facebook-github-bot
Copy link
Contributor

This pull request was exported from Phabricator. Differential Revision: D73891425

…calculation to MHA"

Leveraging previous work now we allow MHA to have ring buffer cache. If ring buffer cache is used
then we query the mask from kv cache and use that for sdpa instead of using precalculated mask.

In this process we had to adjsut ring buffer implementation to allow keeping the context of
full sliding window. See code for comment.

Differential Revision: [D73891425](https://our.internmc.facebook.com/intern/diff/D73891425/)

[ghstack-poisoned]
@facebook-github-bot
Copy link
Contributor

This pull request was exported from Phabricator. Differential Revision: D73891425

@kimishpatel kimishpatel added the release notes: examples Changes to any of our example LLMs integrations, such as Llama3 and Llava label May 5, 2025
…calculation to MHA"

Leveraging previous work now we allow MHA to have ring buffer cache. If ring buffer cache is used
then we query the mask from kv cache and use that for sdpa instead of using precalculated mask.

In this process we had to adjsut ring buffer implementation to allow keeping the context of
full sliding window. See code for comment.

Differential Revision: [D73891425](https://our.internmc.facebook.com/intern/diff/D73891425/)

[ghstack-poisoned]
@facebook-github-bot
Copy link
Contributor

This pull request was exported from Phabricator. Differential Revision: D73891425

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. fb-exported release notes: examples Changes to any of our example LLMs integrations, such as Llama3 and Llava
Projects
None yet
Development

Successfully merging this pull request may close these issues.

3 participants