[ET-VK] Implement SDPA with fused ops#14130
Merged
facebook-github-bot merged 3 commits intogh/SS-JIA/324/basefrom Sep 10, 2025
Merged
[ET-VK] Implement SDPA with fused ops#14130facebook-github-bot merged 3 commits intogh/SS-JIA/324/basefrom
facebook-github-bot merged 3 commits intogh/SS-JIA/324/basefrom
Conversation
## Context As title; optimize the SDPA operator by introducing shaders to perform the operation in 3 steps: 1. Compute attention weights, multiplying QT x K_cache, and applying scale and mask 2. Compute softmax normalization of computed attention weights 3. Compute final output by multiplying attention weights with V cache This new implementation is much more efficient than the existing one, which performed slicing, repeat_interleave, and transposition of projected and cache tensors as separate steps. The fusion of scale and mask with the computation of attention weights also allows for the computation of elements within the mask region to be skipped. ## Impact Decode latency for LLMs is much improved. For llama 3.2 3B generating ~250 tokens, decode latency increases from ~15 tok/s to ~21.5 tok/s Differential Revision: [D82053493](https://our.internmc.facebook.com/intern/diff/D82053493/) [ghstack-poisoned]
SS-JIA
pushed a commit
that referenced
this pull request
Sep 9, 2025
## Context As title; optimize the SDPA operator by introducing shaders to perform the operation in 3 steps: 1. Compute attention weights, multiplying QT x K_cache, and applying scale and mask 2. Compute softmax normalization of computed attention weights 3. Compute final output by multiplying attention weights with V cache This new implementation is much more efficient than the existing one, which performed slicing, repeat_interleave, and transposition of projected and cache tensors as separate steps. The fusion of scale and mask with the computation of attention weights also allows for the computation of elements within the mask region to be skipped. ## Impact Decode latency for LLMs is much improved. For llama 3.2 3B generating ~250 tokens, decode latency increases from ~15 tok/s to ~21.5 tok/s Differential Revision: [D82053493](https://our.internmc.facebook.com/intern/diff/D82053493/) ghstack-source-id: 308592117 Pull Request resolved: #14130
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D82053493 |
This PR needs a
|
manuelcandales
approved these changes
Sep 9, 2025
## Context As title; optimize the SDPA operator by introducing shaders to perform the operation in 3 steps: 1. Compute attention weights, multiplying QT x K_cache, and applying scale and mask 2. Compute softmax normalization of computed attention weights 3. Compute final output by multiplying attention weights with V cache This new implementation is much more efficient than the existing one, which performed slicing, repeat_interleave, and transposition of projected and cache tensors as separate steps. The fusion of scale and mask with the computation of attention weights also allows for the computation of elements within the mask region to be skipped. ## Impact Decode latency for LLMs is much improved. For llama 3.2 3B generating ~250 tokens, decode latency increases from ~15 tok/s to ~21.5 tok/s Differential Revision: [D82053493](https://our.internmc.facebook.com/intern/diff/D82053493/) [ghstack-poisoned]
SS-JIA
pushed a commit
that referenced
this pull request
Sep 9, 2025
Pull Request resolved: #14130 ## Context As title; optimize the SDPA operator by introducing shaders to perform the operation in 3 steps: 1. Compute attention weights, multiplying QT x K_cache, and applying scale and mask 2. Compute softmax normalization of computed attention weights 3. Compute final output by multiplying attention weights with V cache This new implementation is much more efficient than the existing one, which performed slicing, repeat_interleave, and transposition of projected and cache tensors as separate steps. The fusion of scale and mask with the computation of attention weights also allows for the computation of elements within the mask region to be skipped. ## Impact Decode latency for LLMs is much improved. For llama 3.2 3B generating ~250 tokens, decode latency increases from ~15 tok/s to ~21.5 tok/s ghstack-source-id: 308621243 @exported-using-ghexport Differential Revision: [D82053493](https://our.internmc.facebook.com/intern/diff/D82053493/)
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D82053493 |
## Context As title; optimize the SDPA operator by introducing shaders to perform the operation in 3 steps: 1. Compute attention weights, multiplying QT x K_cache, and applying scale and mask 2. Compute softmax normalization of computed attention weights 3. Compute final output by multiplying attention weights with V cache This new implementation is much more efficient than the existing one, which performed slicing, repeat_interleave, and transposition of projected and cache tensors as separate steps. The fusion of scale and mask with the computation of attention weights also allows for the computation of elements within the mask region to be skipped. ## Impact Decode latency for LLMs is much improved. For llama 3.2 3B generating ~250 tokens, decode latency increases from ~15 tok/s to ~21.5 tok/s Differential Revision: [D82053493](https://our.internmc.facebook.com/intern/diff/D82053493/) [ghstack-poisoned]
SS-JIA
pushed a commit
that referenced
this pull request
Sep 10, 2025
Pull Request resolved: #14130 ## Context As title; optimize the SDPA operator by introducing shaders to perform the operation in 3 steps: 1. Compute attention weights, multiplying QT x K_cache, and applying scale and mask 2. Compute softmax normalization of computed attention weights 3. Compute final output by multiplying attention weights with V cache This new implementation is much more efficient than the existing one, which performed slicing, repeat_interleave, and transposition of projected and cache tensors as separate steps. The fusion of scale and mask with the computation of attention weights also allows for the computation of elements within the mask region to be skipped. ## Impact Decode latency for LLMs is much improved. For llama 3.2 3B generating ~250 tokens, decode latency increases from ~15 tok/s to ~21.5 tok/s ghstack-source-id: 308660072 @exported-using-ghexport Differential Revision: [D82053493](https://our.internmc.facebook.com/intern/diff/D82053493/)
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D82053493 |
711ccd8
into
gh/SS-JIA/324/base
246 of 288 checks passed
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
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):
Context
As title; optimize the SDPA operator by introducing shaders to perform the operation in 3 steps:
This new implementation is much more efficient than the existing one, which performed slicing, repeat_interleave, and transposition of projected and cache tensors as separate steps. The fusion of scale and mask with the computation of attention weights also allows for the computation of elements within the mask region to be skipped.
Impact
Decode latency for LLMs is much improved. For llama 3.2 3B generating ~250 tokens, decode latency increases from ~15 tok/s to ~21.5 tok/s
Differential Revision: D82053493