Reorder channelwise gated delta rule chunked hot loops for autovectorization (#21021)#21021
Open
JakeStevens wants to merge 3 commits into
Open
Reorder channelwise gated delta rule chunked hot loops for autovectorization (#21021)#21021JakeStevens wants to merge 3 commits into
JakeStevens wants to merge 3 commits into
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/21021
Note: Links to docs will display an error until the docs builds have been completed. ❗ 2 Active SEVsThere are 2 currently active SEVs. If your PR is affected, please view them below:
❌ 129 New Failures, 6 Unclassified FailuresAs of commit b8d62b5 with merge base 931ba24 ( NEW FAILURES - The following jobs have failed:
UNCLASSIFIED FAILURES - DrCI could not classify the following jobs because the workflow did not run on the merge base. The failures may be pre-existing on trunk or introduced by this PR:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
Contributor
|
@JakeStevens has exported this pull request. If you are a Meta employee, you can view the originating Diff in D112598714. |
This PR needs a
|
…h#21020) Summary: Reduce the channelwise gated delta rule kernel from four state traversals to two per token. The first pass computes the predicted value while folding in decay; the second applies decay and the rank-one update while accumulating the output. This preserves operation grouping and numerical behavior while reducing state-memory traffic. Add a standalone microbenchmark covering decode and representative prefill lengths. Differential Revision: D112596724
Summary: Route the channelwise gated delta rule by sequence length: T == 1 keeps the two-pass token recurrence for autoregressive decode, while T != 1 uses a chunkwise WY/UT formulation for prefill. The chunked path computes per-channel log-decay prefixes, causal query/key terms, the beta-folded triangular transform, WY pseudo-keys and pseudo-values, and inter-chunk state carry. It handles a ragged final chunk without a separate tail implementation. Parallelize independent (batch, head) work across the ExecuTorch threadpool. Each worker receives a disjoint slice of one temporary scratch arena, avoiding shared mutable buffers while amortizing allocation across chunks. Differential Revision: D112597348
…ization (pytorch#21021) Summary: Reorder the chunked prefill inner loops (steps 1, 4, 5, 6) so the innermost loop runs contiguously over the head dimension (k or v) instead of striding down a column of the state / pv. This lets the compiler autovectorize the now-unit-stride AXPYs; hand-written at::vec was tried and was slower than the compiler output, so the loops stay scalar. Differential Revision: D112598714
meta-codesync
Bot
changed the title
JakeStevens
force-pushed
the
export-D112598714
branch
from
July 17, 2026 20:49
b8d62b5 to
dac4505
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
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.
Summary:
Reorder the chunked prefill inner loops (steps 1, 4, 5, 6) so the innermost loop runs contiguously over the head dimension (k or v) instead of striding down a column of the state / pv. This lets the compiler autovectorize the now-unit-stride AXPYs; hand-written at::vec was tried and was slower than the compiler output, so the loops stay scalar.
Differential Revision: D112598714