fix(cpu): default scaledDotProductAttention scale to 1/sqrt(headDim) (#860) - #880
Merged
Conversation
The op's scale parameter defaults to 0f, documented as meaning 1/sqrt(headDim). The CPU backend applied it literally, so calling SDPA without an explicit scale multiplied all scores by zero — the softmax collapsed to a uniform average and the attention pattern was silently discarded. Resolve scale == 0f to 1/sqrt(headDim) in the CPU kernel. Adds a regression test: the default-scale result equals an explicit 1/sqrt(headDim) call and differs from the near-zero (uniform) result. Closes #860
aharakal
approved these changes
Jul 24, 2026
MacOS
pushed a commit
to MacOS/SKaiNET
that referenced
this pull request
Jul 27, 2026
Bump version 0.36.0 -> 0.37.0 (gradle.properties, docs/antora.yml, README quickstart). Promote CHANGELOG [Unreleased] to [0.37.0]: Lstm layer (SKaiNET-developers#824), real Dropout masking (SKaiNET-developers#867), LR schedules and mutable optimizer lr (SKaiNET-developers#866), optional-bias and open Linear (SKaiNET-developers#870, SKaiNET-developers#875), androidNative IO targets (SKaiNET-developers#836, SKaiNET-developers#842, SKaiNET-developers#845), the SDPA default-scale fix (SKaiNET-developers#880), three autograd fixes (SKaiNET-developers#877), the argMax DAG output spec (SKaiNET-developers#878), tokenizer BPE inference and N-D gather (SKaiNET-developers#879), plus the CI/docs supply-chain hardening and toolchain bumps. Refresh README "What's New" and add a Contributors (0.37.0) section. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
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.
Closes #860.
Problem
TensorOps.scaledDotProductAttention'sscaleparameter defaults to0f, documented as "defaults to 1/sqrt(headDim)". The CPU backend applied the value literally, so any call that relied on the default multiplied every attention score by zero — the softmax collapsed to a uniform distribution and the attention pattern was silently discarded. No error, plausible-looking output shapes; only the numbers are wrong.Fix
Resolve
scale == 0fto1 / sqrt(headDim)inside the CPU kernel, matching the documented contract. An explicitly passed scale (including a genuinely tiny one) is untouched.Test
SDPAShapeValidationTest.default_scale_uses_one_over_sqrt_head_dim_not_zero: the default-scale (scale = 0f) result equals an explicit1/sqrt(headDim)call, and differs from the near-zero (uniform) degenerate result. Fullskainet-backend-cpujvmTestgreen.Found while validating a from-scratch multi-head attention against the fused op — the equivalence test caught the silent flattening. This is the last of the eight issues surfaced by that port.