Attention Perf: Transpose blocked K right before QK instead of pre-transposing before the kernel#2374
Merged
Merged
Conversation
AmesingFlank
added a commit
that referenced
this pull request
May 9, 2026
…ansposing before the kernel stack-info: PR: #2374, branch: AmesingFlank/stack/50
AmesingFlank
force-pushed
the
AmesingFlank/stack/50
branch
from
0bdb0b9 to
4577603
Compare
AmesingFlank
force-pushed
the
AmesingFlank/stack/49
branch
from
1b0111f to
cdfb1ce
Compare
meta-cla
Bot
added
the
CLA Signed
AmesingFlank
force-pushed
the
AmesingFlank/stack/50
branch
from
4577603 to
a2abca6
Compare
AmesingFlank
added a commit
that referenced
this pull request
May 9, 2026
…ansposing before the kernel stack-info: PR: #2374, branch: AmesingFlank/stack/50
AmesingFlank
force-pushed
the
AmesingFlank/stack/50
branch
from
a2abca6 to
6ad5991
Compare
jansel
approved these changes
May 9, 2026
…ansposing before the kernel stack-info: PR: #2374, branch: AmesingFlank/stack/50
AmesingFlank
force-pushed
the
AmesingFlank/stack/50
branch
from
6ad5991 to
f9e9453
Compare
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
2 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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Optimization found by claude, by comparing the current Helion kernel with this reference impl
Previously, The Helion-generated kernel reshapes and transposes K before the kernel:
k_view = k_in.reshape([-1, n_dim, head_dim]).transpose(0, 2, 1)producing shape[B*H, D, S]. And then during matmul (jax.dot_general) it then uses a standard contractiondimension_numbers=(((2,), (1,)), ((0,), (0,))).This causes sub-optimal DMA patterns because the transposed K layout in HBM has non-contiguous memory access for the pipeline's sequential block reads along the sequence dimension.
The PR modifies the kernel by keeping K contiguous, and only transposes right before the matmul. On top of the previous PR (#2373) this improves the TFLOPs from 633 to 652