add attention_sink.py #6579
Merged
add attention_sink.py #6579
Conversation
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
This PR adds `KVCacheWithAttentionSink`, which is required for `AttentionSink`. It keeps the first `sink_size` tokens as attention sinks and maintains a sliding window with `window_size` for new tokens. Note: I am trying to implement and verify `AttentionSink` in eager mode first. So the current implementation may still have some lower errors or performance issue. For example, it does not support the case when dynamic shape is disabled. Will leave these problems to resolve when we are ready to deploy `AttentionSink` to edge. Differential Revision: [D65235798](https://our.internmc.facebook.com/intern/diff/D65235798/) [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/6579
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 9105c8f with merge base c726a9b ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
facebook-github-bot
added
the
CLA Signed
|
This pull request was exported from Phabricator. Differential Revision: D65235798 |
This PR adds `KVCacheWithAttentionSink`, which is required for `AttentionSink`. It keeps the first `sink_size` tokens as attention sinks and maintains a sliding window with `window_size` for new tokens. Note: I am trying to implement and verify `AttentionSink` in eager mode first. So the current implementation may still have some lower errors or performance issue. For example, it does not support the case when dynamic shape is disabled. Will leave these problems to resolve when we are ready to deploy `AttentionSink` to edge. Differential Revision: [D65235798](https://our.internmc.facebook.com/intern/diff/D65235798/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D65235798 |
This PR adds `KVCacheWithAttentionSink`, which is required for `AttentionSink`. It keeps the first `sink_size` tokens as attention sinks and maintains a sliding window with `window_size` for new tokens. Note: I am trying to implement and verify `AttentionSink` in eager mode first. So the current implementation may still have some lower errors or performance issue. For example, it does not support the case when dynamic shape is disabled. Will leave these problems to resolve when we are ready to deploy `AttentionSink` to edge. Differential Revision: [D65235798](https://our.internmc.facebook.com/intern/diff/D65235798/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D65235798 |
This was referenced Nov 7, 2024
This PR adds `KVCacheWithAttentionSink`, which is required for `AttentionSink`. It keeps the first `sink_size` tokens as attention sinks and maintains a sliding window with `window_size` for new tokens. Note: I am trying to implement and verify `AttentionSink` in eager mode first. So the current implementation may still have some lower errors or performance issue. For example, it does not support the case when dynamic shape is disabled. Will leave these problems to resolve when we are ready to deploy `AttentionSink` to edge. Differential Revision: [D65235798](https://our.internmc.facebook.com/intern/diff/D65235798/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D65235798 |
This PR adds `KVCacheWithAttentionSink`, which is required for `AttentionSink`. It keeps the first `sink_size` tokens as attention sinks and maintains a sliding window with `window_size` for new tokens. Note: I am trying to implement and verify `AttentionSink` in eager mode first. So the current implementation may still have some lower errors or performance issue. For example, it does not support the case when dynamic shape is disabled. Will leave these problems to resolve when we are ready to deploy `AttentionSink` to edge. Differential Revision: [D65235798](https://our.internmc.facebook.com/intern/diff/D65235798/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D65235798 |
This PR adds `KVCacheWithAttentionSink`, which is required for `AttentionSink`. It keeps the first `sink_size` tokens as attention sinks and maintains a sliding window with `window_size` for new tokens. Note: I am trying to implement and verify `AttentionSink` in eager mode first. So the current implementation may still have some lower errors or performance issue. For example, it does not support the case when dynamic shape is disabled. Will leave these problems to resolve when we are ready to deploy `AttentionSink` to edge. Differential Revision: [D65235798](https://our.internmc.facebook.com/intern/diff/D65235798/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D65235798 |
This was referenced Nov 22, 2024
This PR adds `KVCacheWithAttentionSink`, which is required for `AttentionSink`. It keeps the first `sink_size` tokens as attention sinks and maintains a sliding window with `window_size` for new tokens. Note: I am trying to implement and verify `AttentionSink` in eager mode first. So the current implementation may still have some lower errors or performance issue. For example, it does not support the case when dynamic shape is disabled. Will leave these problems to resolve when we are ready to deploy `AttentionSink` to edge. Differential Revision: [D65235798](https://our.internmc.facebook.com/intern/diff/D65235798/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D65235798 |
This PR adds `KVCacheWithAttentionSink`, which is required for `AttentionSink`. It keeps the first `sink_size` tokens as attention sinks and maintains a sliding window with `window_size` for new tokens. Note: I am trying to implement and verify `AttentionSink` in eager mode first. So the current implementation may still have some lower errors or performance issue. For example, it does not support the case when dynamic shape is disabled. Will leave these problems to resolve when we are ready to deploy `AttentionSink` to edge. Differential Revision: [D65235798](https://our.internmc.facebook.com/intern/diff/D65235798/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D65235798 |
larryliu0820
approved these changes
Nov 27, 2024
This PR adds `KVCacheWithAttentionSink`, which is required for `AttentionSink`. It keeps the first `sink_size` tokens as attention sinks and maintains a sliding window with `window_size` for new tokens. Note: I am trying to implement and verify `AttentionSink` in eager mode first. So the current implementation may still have some lower errors or performance issue. For example, it does not support the case when dynamic shape is disabled. Will leave these problems to resolve when we are ready to deploy `AttentionSink` to edge. Differential Revision: [D65235798](https://our.internmc.facebook.com/intern/diff/D65235798/) [ghstack-poisoned]
helunwencser
added a commit
that referenced
this pull request
Nov 28, 2024
Pull Request resolved: #6579 This PR adds `KVCacheWithAttentionSink`, which is required for `AttentionSink`. It keeps the first `sink_size` tokens as attention sinks and maintains a sliding window with `window_size` for new tokens. Note: I am trying to implement and verify `AttentionSink` in eager mode first. So the current implementation may still have some lower errors. Will leave these problems to resolve when we are ready to deploy `AttentionSink` to edge. ghstack-source-id: 255715047 @exported-using-ghexport Differential Revision: [D65235798](https://our.internmc.facebook.com/intern/diff/D65235798/)
|
This pull request was exported from Phabricator. Differential Revision: D65235798 |
facebook-github-bot
merged commit 8d30fc1
into
gh/helunwencser/70/base
41 of 43 checks passed
facebook-github-bot
temporarily deployed
to
cherry-pick-bot
GitHub Actions
Inactive
Merged
kirklandsign
pushed a commit
that referenced
this pull request
Dec 2, 2024
add KVCacheWithAttentionSink Pull Request resolved: #6579 This PR adds `KVCacheWithAttentionSink`, which is required for `AttentionSink`. It keeps the first `sink_size` tokens as attention sinks and maintains a sliding window with `window_size` for new tokens. Note: I am trying to implement and verify `AttentionSink` in eager mode first. So the current implementation may still have some lower errors. Will leave these problems to resolve when we are ready to deploy `AttentionSink` to edge. ghstack-source-id: 255715047 @exported-using-ghexport Differential Revision: [D65235798](https://our.internmc.facebook.com/intern/diff/D65235798/) Co-authored-by: Lunwen He <lwhecser@gmail.com>
kedarnath03
pushed a commit
to kedarnath03/executorch
that referenced
this pull request
Jun 25, 2025
Pull Request resolved: pytorch/executorch#6579 This PR adds `KVCacheWithAttentionSink`, which is required for `AttentionSink`. It keeps the first `sink_size` tokens as attention sinks and maintains a sliding window with `window_size` for new tokens. Note: I am trying to implement and verify `AttentionSink` in eager mode first. So the current implementation may still have some lower errors. Will leave these problems to resolve when we are ready to deploy `AttentionSink` to edge. ghstack-source-id: 254019779 @exported-using-ghexport Differential Revision: [D65235798](https://our.internmc.facebook.com/intern/diff/D65235798/)
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.
Stack from ghstack (oldest at bottom):
This PR adds
KVCacheWithAttentionSink, which is required forAttentionSink. It keeps the firstsink_sizetokens as attention sinks and maintains a sliding window withwindow_sizefor new tokens.Note: I am trying to implement and verify
AttentionSinkin eager mode first. So the current implementation may still have some lower errors or performance issue. For example, it does not support the case when dynamic shape is disabled. Will leave these problems to resolve when we are ready to deployAttentionSinkto edge.Differential Revision: D65235798