Gemma3 CPU export: replace fused SkipSimplifiedLayerNormalization with explicit Add + RMSNorm path

[Copilot]

Gemma3 (Gemma3ForCausalLM) ONNX exports on CPU can diverge from PyTorch for prompts that trigger large early-layer activations (notably in sliding-attention layers), producing degenerate first-token behavior. The divergence was traced to layer-0 residual stream behavior around fused normalization.

• Root-cause-targeted export change (Gemma2/Gemma3 on CPU)

  • In Gemma2Model.make_layernorm_op, CPU handling for SkipSimplifiedLayerNormalization is now emitted as:
    • Add (residual combine, output_3)
    • SimplifiedLayerNormalization (normalized path, output_0)
  • Non-CPU paths continue to use the fused com.microsoft::SkipSimplifiedLayerNormalization op unchanged.

• Scope and compatibility

  • Change is isolated to Gemma builder logic (src/python/py/models/builders/gemma.py).
  • No behavior change for non-Gemma architectures or non-CPU execution providers.

• Regression coverage

  • Added focused builder unit tests (test/python/builder/test_gemma.py) to lock in:
    • CPU emits expanded Add + SimplifiedLayerNormalization
    • non-CPU still emits fused SkipSimplifiedLayerNormalization

# CPU Gemma export path for skip RMSNorm now expands to:
Add(root_input, skip_input) -> output_3
SimplifiedLayerNormalization(output_3, weight) -> output_0

[azure-pipelines]

Azure Pipelines:
There may be pipelines that require an authorized user to comment /azp run to run.