implement position encoding for shifted tokens #6646
Merged
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
In AttentionSink, it uses tokens' positions in the KVCache instead of the actual text. When tokens get shifted in KVCache, it needs to update q and k's position embedding. In the original [implementation](https://github.com/mit-han-lab/streaming-llm) of AttentionSink with Rope, it caches the original q and k in KVCache and apply position embedding during inference. This PR adds `RopeWithAttentionSink`. It assumes that q and k are already encoded with their original position. When we shift tokens, we reapply the position delta. This has two benefits: - minimize our code since our existing `llama_transformer` applies rope embedding before doing KVCache update - avoid performance regression when tokens are not shifted because we don't need to reapply position encoding in KVCache for them Differential Revision: [D65366440](https://our.internmc.facebook.com/intern/diff/D65366440/) [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/6646
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit a574663 with merge base b8fbc48 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
facebook-github-bot
added
the
CLA Signed
This was referenced Nov 4, 2024
Merged
|
This pull request was exported from Phabricator. Differential Revision: D65366440 |
In AttentionSink, it uses tokens' positions in the KVCache instead of the actual text. When tokens get shifted in KVCache, it needs to update q and k's position embedding. In the original [implementation](https://github.com/mit-han-lab/streaming-llm) of AttentionSink with Rope, it caches the original q and k in KVCache and apply position embedding during inference. This PR adds `RopeWithAttentionSink`. It assumes that q and k are already encoded with their original position. When we shift tokens, we reapply the position delta. This has two benefits: - minimize our code since our existing `llama_transformer` applies rope embedding before doing KVCache update - avoid performance regression when tokens are not shifted because we don't need to reapply position encoding in KVCache for them Differential Revision: [D65366440](https://our.internmc.facebook.com/intern/diff/D65366440/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D65366440 |
In AttentionSink, it uses tokens' positions in the KVCache instead of the actual text. When tokens get shifted in KVCache, it needs to update q and k's position embedding. In the original [implementation](https://github.com/mit-han-lab/streaming-llm) of AttentionSink with Rope, it caches the original q and k in KVCache and apply position embedding during inference. This PR adds `RopeWithAttentionSink`. It assumes that q and k are already encoded with their original position. When we shift tokens, we reapply the position delta. This has two benefits: - minimize our code since our existing `llama_transformer` applies rope embedding before doing KVCache update - avoid performance regression when tokens are not shifted because we don't need to reapply position encoding in KVCache for them Differential Revision: [D65366440](https://our.internmc.facebook.com/intern/diff/D65366440/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D65366440 |
In AttentionSink, it uses tokens' positions in the KVCache instead of the actual text. When tokens get shifted in KVCache, it needs to update q and k's position embedding. In the original [implementation](https://github.com/mit-han-lab/streaming-llm) of AttentionSink with Rope, it caches the original q and k in KVCache and apply position embedding during inference. This PR adds `RopeWithAttentionSink`. It assumes that q and k are already encoded with their original position. When we shift tokens, we reapply the position delta. This has two benefits: - minimize our code since our existing `llama_transformer` applies rope embedding before doing KVCache update - avoid performance regression when tokens are not shifted because we don't need to reapply position encoding in KVCache for them Differential Revision: [D65366440](https://our.internmc.facebook.com/intern/diff/D65366440/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D65366440 |
This was referenced Nov 7, 2024
In AttentionSink, it uses tokens' positions in the KVCache instead of the actual text. When tokens get shifted in KVCache, it needs to update q and k's position embedding. In the original [implementation](https://github.com/mit-han-lab/streaming-llm) of AttentionSink with Rope, it caches the original q and k in KVCache and apply position embedding during inference. This PR adds `RopeWithAttentionSink`. It assumes that q and k are already encoded with their original position. When we shift tokens, we reapply the position delta. This has two benefits: - minimize our code since our existing `llama_transformer` applies rope embedding before doing KVCache update - avoid performance regression when tokens are not shifted because we don't need to reapply position encoding in KVCache for them Differential Revision: [D65366440](https://our.internmc.facebook.com/intern/diff/D65366440/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D65366440 |
In AttentionSink, it uses tokens' positions in the KVCache instead of the actual text. When tokens get shifted in KVCache, it needs to update q and k's position embedding. In the original [implementation](https://github.com/mit-han-lab/streaming-llm) of AttentionSink with Rope, it caches the original q and k in KVCache and apply position embedding during inference. This PR adds `RopeWithAttentionSink`. It assumes that q and k are already encoded with their original position. When we shift tokens, we reapply the position delta. This has two benefits: - minimize our code since our existing `llama_transformer` applies rope embedding before doing KVCache update - avoid performance regression when tokens are not shifted because we don't need to reapply position encoding in KVCache for them Differential Revision: [D65366440](https://our.internmc.facebook.com/intern/diff/D65366440/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D65366440 |
This was referenced Nov 22, 2024
In AttentionSink, it uses tokens' positions in the KVCache instead of the actual text. When tokens get shifted in KVCache, it needs to update q and k's position embedding. In the original [implementation](https://github.com/mit-han-lab/streaming-llm) of AttentionSink with Rope, it caches the original q and k in KVCache and apply position embedding during inference. This PR adds `RopeWithAttentionSink`. It assumes that q and k are already encoded with their original position. When we shift tokens, we reapply the position delta. This has two benefits: - minimize our code since our existing `llama_transformer` applies rope embedding before doing KVCache update - avoid performance regression when tokens are not shifted because we don't need to reapply position encoding in KVCache for them Differential Revision: [D65366440](https://our.internmc.facebook.com/intern/diff/D65366440/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D65366440 |
In AttentionSink, it uses tokens' positions in the KVCache instead of the actual text. When tokens get shifted in KVCache, it needs to update q and k's position embedding. In the original [implementation](https://github.com/mit-han-lab/streaming-llm) of AttentionSink with Rope, it caches the original q and k in KVCache and apply position embedding during inference. This PR adds `RopeWithAttentionSink`. It assumes that q and k are already encoded with their original position. When we shift tokens, we reapply the position delta. This has two benefits: - minimize our code since our existing `llama_transformer` applies rope embedding before doing KVCache update - avoid performance regression when tokens are not shifted because we don't need to reapply position encoding in KVCache for them Differential Revision: [D65366440](https://our.internmc.facebook.com/intern/diff/D65366440/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D65366440 |
larryliu0820
approved these changes
Nov 27, 2024
facebook-github-bot
merged commit 630ab40
into
gh/helunwencser/71/base
41 of 44 checks passed
facebook-github-bot
temporarily deployed
to
cherry-pick-bot
GitHub Actions
Inactive
kirklandsign
pushed a commit
that referenced
this pull request
Nov 27, 2024
Pull Request resolved: #6646 In AttentionSink, it uses tokens' positions in the KVCache instead of the actual text. When tokens get shifted in KVCache, it needs to update q and k's position embedding. In the original [implementation](https://github.com/mit-han-lab/streaming-llm) of AttentionSink with Rope, it caches the original q and k in KVCache and apply position embedding during inference. This PR adds `RopeWithAttentionSink`. It assumes that q and k are already encoded with their original position. When we shift tokens, we reapply the position delta. This has two benefits: - minimize our code since our existing `llama_transformer` applies rope embedding before doing KVCache update - avoid performance regression when tokens are not shifted because we don't need to reapply position encoding in KVCache for them ghstack-source-id: 255579838 Differential Revision: [D65366440](https://our.internmc.facebook.com/intern/diff/D65366440/) --------- Co-authored-by: Lunwen He <[email protected]>
mergennachin
pushed a commit
that referenced
this pull request
Dec 2, 2024
Pull Request resolved: #6646 In AttentionSink, it uses tokens' positions in the KVCache instead of the actual text. When tokens get shifted in KVCache, it needs to update q and k's position embedding. In the original [implementation](https://github.com/mit-han-lab/streaming-llm) of AttentionSink with Rope, it caches the original q and k in KVCache and apply position embedding during inference. This PR adds `RopeWithAttentionSink`. It assumes that q and k are already encoded with their original position. When we shift tokens, we reapply the position delta. This has two benefits: - minimize our code since our existing `llama_transformer` applies rope embedding before doing KVCache update - avoid performance regression when tokens are not shifted because we don't need to reapply position encoding in KVCache for them ghstack-source-id: 255579838 Differential Revision: [D65366440](https://our.internmc.facebook.com/intern/diff/D65366440/) --------- Co-authored-by: Lunwen He <[email protected]>
kedarnath03
pushed a commit
to kedarnath03/executorch
that referenced
this pull request
Jun 25, 2025
Pull Request resolved: pytorch/executorch#6646 In AttentionSink, it uses tokens' positions in the KVCache instead of the actual text. When tokens get shifted in KVCache, it needs to update q and k's position embedding. In the original [implementation](https://github.com/mit-han-lab/streaming-llm) of AttentionSink with Rope, it caches the original q and k in KVCache and apply position embedding during inference. This PR adds `RopeWithAttentionSink`. It assumes that q and k are already encoded with their original position. When we shift tokens, we reapply the position delta. This has two benefits: - minimize our code since our existing `llama_transformer` applies rope embedding before doing KVCache update - avoid performance regression when tokens are not shifted because we don't need to reapply position encoding in KVCache for them ghstack-source-id: 254019780 Differential Revision: [D65366440](https://our.internmc.facebook.com/intern/diff/D65366440/)
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):
In AttentionSink, it uses tokens' positions in the KVCache instead of the actual text. When tokens get shifted in KVCache, it needs to update q and k's position embedding.
In the original implementation of AttentionSink with Rope, it caches the original q and k in KVCache and apply position embedding during inference.
This PR adds
RopeWithAttentionSink. It assumes that q and k are already encoded with their original position. When we shift tokens, we reapply the position delta. This has two benefits:llama_transformerapplies rope embedding before doing KVCache updateDifferential Revision: D65366440