Add non-flash SDPA path gated by ET_USE_UNFUSED_SDPA#18717
Open
kimishpatel wants to merge 1 commit into
Open
Conversation
Benchmarks show ET's flash attention is 10-12% slower than non-tiled GEMM-based SDPA for decode (SeqLen=1) due to within-head tiling overhead (multiple small GEMMs + online softmax rescaling). This adds an alternative non-flash code path that computes full Q@K^T, standard softmax, then scores@V using two GEMM calls, gated by #ifdef ET_USE_UNFUSED_SDPA so it can be tested without disrupting the existing flash path. The new cpu_sdpa function reuses existing SeqDim, stride extraction, cpublas::gemm, and parallel_for infrastructure. Float-only (no quantized input support). Threading granularity is one (batch, head) per work unit. Differential Revision: [D99677685](https://our.internmc.facebook.com/intern/diff/D99677685/) [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/18717
Note: Links to docs will display an error until the docs builds have been completed. ❌ 126 New Failures, 1 Unrelated FailureAs of commit 867ba3e with merge base fb1618e ( NEW FAILURES - The following jobs have failed:
FLAKY - The following job failed but was likely due to flakiness present on trunk:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
This was referenced Apr 6, 2026
Open
This PR needs a
|
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):
Benchmarks show ET's flash attention is 10-12% slower than non-tiled
GEMM-based SDPA for decode (SeqLen=1) due to within-head tiling overhead
(multiple small GEMMs + online softmax rescaling). This adds an alternative
non-flash code path that computes full Q@K^T, standard softmax, then
scores@V using two GEMM calls, gated by #ifdef ET_USE_UNFUSED_SDPA so it
can be tested without disrupting the existing flash path.
The new cpu_sdpa function reuses existing SeqDim, stride extraction,
cpublas::gemm, and parallel_for infrastructure. Float-only (no quantized
input support). Threading granularity is one (batch, head) per work unit.
Differential Revision: D99677685