Support separate K/V seq dim in custom_sdpa op#18714
Open
kimishpatel wants to merge 1 commit into
Open
Conversation
Previously custom_sdpa used a single is_seq_at_dim_2 flag for all tensors. This meant v_only transpose required a runtime transpose copy for K (converting from [B,H,S,D] to [B,S,H,D]), which caused a 2.3x decode slowdown (15.35 vs 35.63 tok/s). Now the C++ op accepts separate is_seq_dim_2, is_k_seq_dim_2, is_v_seq_dim_2 flags so Q, K, V can each have independent layouts. The Python layer passes K and V in their native cache layout without any transpose, and the flash attention kernel handles the mixed strides directly. Changes: - op_sdpa_impl.h: cpu_flash_attention takes q_seq_dim, k_seq_dim, v_seq_dim instead of single seq_dim - op_sdpa.cpp/h: custom_sdpa_out takes 3 bool params - op_sdpa_aot.cpp: Updated schema strings and wrappers - sdpa.py: SDPACustom uses is_k_seq_at_dim_2 / is_v_seq_at_dim_2, Q always at dim 2, no input transposes - custom_kv_cache.py: update() returns native cache layout, added is_seq_at_dim_2 compat property - export_llama_lib.py: passes separate K/V flags Differential Revision: [D99677678](https://our.internmc.facebook.com/intern/diff/D99677678/) [ghstack-poisoned]
kimishpatel
requested review from
larryliu0820,
lucylq and
mergennachin
as code owners
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/18714
Note: Links to docs will display an error until the docs builds have been completed. ❌ 126 New Failures, 2 Cancelled JobsAs of commit 45468e1 with merge base fb1618e ( NEW FAILURES - The following jobs have failed:
CANCELLED JOBS - The following jobs were cancelled. Please retry:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
This was referenced Apr 6, 2026
This PR needs a
|
This was referenced Apr 6, 2026
Open
meta-cla
Bot
added
the
CLA Signed
Contributor
Author
|
submitted by accident, not meant to land immedidately |
|
Looks like this PR hasn't been updated in a while so we're going to go ahead and mark this as |
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
1 participant
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):
Previously custom_sdpa used a single is_seq_at_dim_2 flag for all
tensors. This meant v_only transpose required a runtime transpose
copy for K (converting from [B,H,S,D] to [B,S,H,D]), which caused
a 2.3x decode slowdown (15.35 vs 35.63 tok/s).
Now the C++ op accepts separate is_seq_dim_2, is_k_seq_dim_2,
is_v_seq_dim_2 flags so Q, K, V can each have independent layouts.
The Python layer passes K and V in their native cache layout
without any transpose, and the flash attention kernel handles the
mixed strides directly.
Changes:
v_seq_dim instead of single seq_dim
Q always at dim 2, no input transposes
added is_seq_at_dim_2 compat property
Differential Revision: D99677678