Skip to content

onnx: add com.microsoft SimplifiedLayerNormalization + SkipSimplifiedLayerNormalization#2288

Open
czoli1976 wants to merge 1 commit into
sonos:mainfrom
czoli1976:feature/onnx-rmsnorm-contrib
Open

onnx: add com.microsoft SimplifiedLayerNormalization + SkipSimplifiedLayerNormalization#2288
czoli1976 wants to merge 1 commit into
sonos:mainfrom
czoli1976:feature/onnx-rmsnorm-contrib

Conversation

@czoli1976
Copy link
Copy Markdown
Contributor

SimplifiedLayerNormalization (RMSNorm; reuses the RMSNormalization handler) and SkipSimplifiedLayerNormalization (residual input+skip then RMSNorm*gamma, optional input_skip_bias_sum output; ORT >=1.19 ignores the optional bias for the simplified variant -- matched empirically). Validated bit-close vs onnxruntime/numpy across shapes; no node-suite regression; clippy+fmt clean. Part of com.microsoft contrib-op coverage for ORT-exported LLMs.

…LayerNormalization

SimplifiedLayerNormalization is RMSNorm and reuses the existing
RMSNormalization handler. SkipSimplifiedLayerNormalization adds the
residual (input + skip), RMS-normalizes over the last axis and scales by
gamma, optionally emitting input_skip_bias_sum; ORT (>=1.19) does not
apply the optional bias for the simplified variant, so it is ignored.

Validated against onnxruntime (output) and numpy (RMSNorm); bit-close.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
@czoli1976 czoli1976 force-pushed the feature/onnx-rmsnorm-contrib branch from 89f21c3 to ab1bd22 Compare May 26, 2026 12:35
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant