[Copilot]

Skip RMS fusion can combine float16 residual inputs with a float32 scale, producing a graph ONNX Runtime rejects.

• Preserve safe fusion: Narrow constant float32 scales only when they round-trip exactly through float16 and the norm output exclusively feeds a cast back to float16. Replace that cast with the fused output.
• Reject unsafe fusion: Require matching input, skip, and optional bias dtypes; skip incompatible scales or output consumers.
• Regression coverage: Cover exact and inexact scales, dynamic scales, shared outputs, bias variants, and Gemma-style optimization.

• Fixes fuse_skip_rms_normalization fuses a float16 residual into a norm whose scale is float32, which ONNX Runtime rejects #3155

[Copilot]

🟡 Changes recommended

The fusion can silently replace an overridable graph-input scale with a fixed constant, ignoring runtime values, and abort when external scale data is unavailable.

2 open issues

• Preserve runtime-overridable scales during narrowing · New
• Handle missing external tensor data during scale matching · New

What changed in this PR

This PR addresses #3155 by guarding skip RMS normalization fusion against mixed dtypes that ONNX Runtime rejects, while preserving fusion for eligible float32 scales.

Changes:

• Require matching input, skip, and optional bias dtypes.
• Narrow exactly representable constant scales to float16 when normalization exclusively feeds a cast back to float16.
• Add regression coverage for safe and unsafe scales, shared outputs, bias variants, and Gemma-style optimization.

File	Description
onnxscript/​rewriter/​ort_fusions/​skip_normalization.py	Adds dtype guards and cast-back fusion rules with scale narrowing.
onnxscript/​rewriter/​ort_fusions/​skip_normalization_unit_test.py	Tests mixed-precision fusion eligibility and numerical equivalence.

_{Review effort: Balanced}

[codecov]

Codecov Report

❌ Patch coverage is 95.32710% with 5 lines in your changes missing coverage. Please review.
✅ Project coverage is 73.42%. Comparing base (5b210da) to head (9164e8d).
✅ All tests successful. No failed tests found.

Files with missing lines	Patch %	Lines
...writer/ort_fusions/skip_normalization_unit_test.py	93.97%	5 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #3163      +/-   ##
==========================================
+ Coverage   73.34%   73.42%   +0.07%     
==========================================
  Files         269      269              
  Lines       33564    33668     +104     
  Branches     3210     3226      +16     
==========================================
+ Hits        24618    24720     +102     
- Misses       7876     7879       +3     
+ Partials     1070     1069       -1     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.