[Executorch][llama] Add custom_sdpa and use that instead of sdpa_with_kv_cache #5669
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
…_kv_cache sdpa_with_kv_cache updates kv cache. In quantized kv cache, cache updates happens separately. Then the quantized cache is dequantized. After that we call sdpa_with_kv_cache which copies k and v data into dequantized cache. Although this is not needed because the actual cache is the one that is quantized. For very large context length this will add significant amount data copy. Subsequent diffs will deprecate sdpa_with_kv_cache op and deconstruct that using a) update_cache op and b) custom_sdpa op. Differential Revision: [D62623241](https://our.internmc.facebook.com/intern/diff/D62623241/) [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/5669
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit f43dc24 with merge base ee32848 ( 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: D62623241 |
This was referenced Sep 25, 2024
swolchok
approved these changes
Sep 26, 2024
…f sdpa_with_kv_cache" sdpa_with_kv_cache updates kv cache. In quantized kv cache, cache updates happens separately. Then the quantized cache is dequantized. After that we call sdpa_with_kv_cache which copies k and v data into dequantized cache. Although this is not needed because the actual cache is the one that is quantized. For very large context length this will add significant amount data copy. Subsequent diffs will deprecate sdpa_with_kv_cache op and deconstruct that using a) update_cache op and b) custom_sdpa op. Differential Revision: [D62623241](https://our.internmc.facebook.com/intern/diff/D62623241/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D62623241 |
…f sdpa_with_kv_cache" sdpa_with_kv_cache updates kv cache. In quantized kv cache, cache updates happens separately. Then the quantized cache is dequantized. After that we call sdpa_with_kv_cache which copies k and v data into dequantized cache. Although this is not needed because the actual cache is the one that is quantized. For very large context length this will add significant amount data copy. Subsequent diffs will deprecate sdpa_with_kv_cache op and deconstruct that using a) update_cache op and b) custom_sdpa op. Differential Revision: [D62623241](https://our.internmc.facebook.com/intern/diff/D62623241/) [ghstack-poisoned]
kimishpatel
added a commit
that referenced
this pull request
Sep 27, 2024
…_kv_cache Pull Request resolved: #5669 sdpa_with_kv_cache updates kv cache. In quantized kv cache, cache updates happens separately. Then the quantized cache is dequantized. After that we call sdpa_with_kv_cache which copies k and v data into dequantized cache. Although this is not needed because the actual cache is the one that is quantized. For very large context length this will add significant amount data copy. Subsequent diffs will deprecate sdpa_with_kv_cache op and deconstruct that using a) update_cache op and b) custom_sdpa op. ghstack-source-id: 245150344 @exported-using-ghexport Differential Revision: [D62623241](https://our.internmc.facebook.com/intern/diff/D62623241/)
|
This pull request was exported from Phabricator. Differential Revision: D62623241 |
…f sdpa_with_kv_cache" sdpa_with_kv_cache updates kv cache. In quantized kv cache, cache updates happens separately. Then the quantized cache is dequantized. After that we call sdpa_with_kv_cache which copies k and v data into dequantized cache. Although this is not needed because the actual cache is the one that is quantized. For very large context length this will add significant amount data copy. Subsequent diffs will deprecate sdpa_with_kv_cache op and deconstruct that using a) update_cache op and b) custom_sdpa op. Differential Revision: [D62623241](https://our.internmc.facebook.com/intern/diff/D62623241/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D62623241 |
…f sdpa_with_kv_cache" sdpa_with_kv_cache updates kv cache. In quantized kv cache, cache updates happens separately. Then the quantized cache is dequantized. After that we call sdpa_with_kv_cache which copies k and v data into dequantized cache. Although this is not needed because the actual cache is the one that is quantized. For very large context length this will add significant amount data copy. Subsequent diffs will deprecate sdpa_with_kv_cache op and deconstruct that using a) update_cache op and b) custom_sdpa op. Differential Revision: [D62623241](https://our.internmc.facebook.com/intern/diff/D62623241/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D62623241 |
…f sdpa_with_kv_cache" sdpa_with_kv_cache updates kv cache. In quantized kv cache, cache updates happens separately. Then the quantized cache is dequantized. After that we call sdpa_with_kv_cache which copies k and v data into dequantized cache. Although this is not needed because the actual cache is the one that is quantized. For very large context length this will add significant amount data copy. Subsequent diffs will deprecate sdpa_with_kv_cache op and deconstruct that using a) update_cache op and b) custom_sdpa op. Differential Revision: [D62623241](https://our.internmc.facebook.com/intern/diff/D62623241/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D62623241 |
…f sdpa_with_kv_cache" sdpa_with_kv_cache updates kv cache. In quantized kv cache, cache updates happens separately. Then the quantized cache is dequantized. After that we call sdpa_with_kv_cache which copies k and v data into dequantized cache. Although this is not needed because the actual cache is the one that is quantized. For very large context length this will add significant amount data copy. Subsequent diffs will deprecate sdpa_with_kv_cache op and deconstruct that using a) update_cache op and b) custom_sdpa op. Differential Revision: [D62623241](https://our.internmc.facebook.com/intern/diff/D62623241/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D62623241 |
|
This pull request has been merged in 43d7662. |
kedarnath03
pushed a commit
to kedarnath03/executorch
that referenced
this pull request
Jun 25, 2025
…_kv_cache Pull Request resolved: pytorch/executorch#5669 sdpa_with_kv_cache updates kv cache. In quantized kv cache, cache updates happens separately. Then the quantized cache is dequantized. After that we call sdpa_with_kv_cache which copies k and v data into dequantized cache. Although this is not needed because the actual cache is the one that is quantized. For very large context length this will add significant amount data copy. Subsequent diffs will deprecate sdpa_with_kv_cache op and deconstruct that using a) update_cache op and b) custom_sdpa op. ghstack-source-id: 245751544 @exported-using-ghexport Differential Revision: [D62623241](https://our.internmc.facebook.com/intern/diff/D62623241/)
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.
Stack from ghstack (oldest at bottom):
sdpa_with_kv_cache updates kv cache. In quantized kv cache, cache updates
happens separately. Then the quantized cache is dequantized. After that
we call sdpa_with_kv_cache which copies k and v data into dequantized cache.
Although this is not needed because the actual cache is the one that is
quantized.
For very large context length this will add significant amount data copy.
Subsequent diffs will deprecate sdpa_with_kv_cache op and deconstruct that
using a) update_cache op and b) custom_sdpa op.
Differential Revision: D62623241