[ET-VK][sdpa] Use numerically-stable softmax in attention weights#18460
Merged
Conversation
The SDPA attention weights softmax shader computed naive softmax:
exp(x) / sum(exp(x)). When attention weights are large (e.g., 151.29 for
Phi-4-mini with head_dim=128), exp(x) overflows float32 (threshold ~88.7),
producing Infinity and then NaN from inf/inf in the normalization step.
This replaces the naive softmax with the standard numerically-stable variant:
exp(x - max(x)) / sum(exp(x - max(x))). The implementation adds a cooperative
max-finding pass (same workgroup reduction pattern as the existing exp_sum pass)
before the exp_sum and normalization passes. The max subtraction ensures that the
largest exponent is 0, preventing overflow.
This fixes Phi-4-mini Vulkan inference which previously produced garbage output
due to NaN propagation from the first transformer layer's attention.
On-device A/B benchmarks on Samsung Galaxy S24 (Adreno 750) with Llama 3.2 1B
(8da4w g128 q4emb, 677 MB) confirm no performance regression:
Llama 3.2 1B (short prompt, 4 tokens, --warmup):
Prefill: 67.2 tok/s | Decode: 59.4 tok/s | TTFT: 60 ms
Llama 3.2 1B (medium prompt, 197 tokens, --warmup):
Prefill: 723.5 tok/s | Decode: 53.3 tok/s | TTFT: 273 ms
These numbers are within run-to-run variance of the baseline (no fix) measurements,
confirming the additional max-finding pass has negligible overhead.
Differential Revision: [D97757920](https://our.internmc.facebook.com/intern/diff/D97757920/)
ghstack-source-id: 356136427
Pull Request resolved: #18407
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/18460
Note: Links to docs will display an error until the docs builds have been completed. ⏳ No Failures, 121 PendingAs of commit 186ad12 with merge base 60d57e5 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
This PR needs a
|
…nt op (#18461) This PR was created by the merge bot to help merge the original PR into the main branch. ghstack PR number: #18408 by @SS-JIA ^ Please use this as the source of truth for the PR details, comments, and reviews ghstack PR base: https://github.com/pytorch/executorch/tree/gh/SS-JIA/501/base ghstack PR head: https://github.com/pytorch/executorch/tree/gh/SS-JIA/501/head Merge bot PR base: https://github.com/pytorch/executorch/tree/gh/SS-JIA/500/orig Merge bot PR head: https://github.com/pytorch/executorch/tree/gh/SS-JIA/501/orig Differential Revision: [D97757921](https://our.internmc.facebook.com/intern/diff/D97757921/) @diff-train-skip-merge Co-authored-by: ssjia <ssjia@devvm26340.ftw0.facebook.com>
SS-JIA
approved these changes
Mar 24, 2026
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.
This PR was created by the merge bot to help merge the original PR into the main branch.
ghstack PR number: #18407 by @SS-JIA
^ Please use this as the source of truth for the PR details, comments, and reviews
ghstack PR base: https://github.com/pytorch/executorch/tree/gh/SS-JIA/500/base
ghstack PR head: https://github.com/pytorch/executorch/tree/gh/SS-JIA/500/head
Merge bot PR base: https://github.com/pytorch/executorch/tree/main
Merge bot PR head: https://github.com/pytorch/executorch/tree/gh/SS-JIA/500/orig
Differential Revision: D97757920
@diff-train-skip-merge