[Baiju Meswani (baijumeswani)]

Describe your changes

Preserve native GenAI embedding construction for ordinary source tensors. Olive's ModelBuilder patch now delegates those tensors to the original builder while keeping the legacy Olive-quantized path.

The wrapper is idempotent, so repeated passes do not accumulate patches. This preserves native INT8 embedding selection and component metadata without adding model-specific behavior.

Validation: all 29 Model Builder pass tests and configured lintrunner checks, including Pylint, pass. A real reduced Ling component export also preserves INT8 CPU embedding storage and the native rank-two lookup interface.

Checklist before requesting a review

• Add unit tests for this change.
• Make sure all tests can pass.
• Update documents if necessary. No documentation change is needed.
• Lint and apply fixes to your code by running lintrunner -a. Configured lint passes without remaining fixes.
• Is this a user-facing change? Release note: ModelBuilder preserves native GenAI embedding export for ordinary source tensors.

[azure-pipelines]

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

[Copilot]

🟢 Approval recommended

No active review issues block approval of the embedding-routing change.

No open issues 🎉

What changed in this PR

Preserves native GenAI embedding construction in Olive’s ModelBuilder for ordinary tensors while retaining the existing Olive-quantized embedding path.

Changes:

• Delegates ordinary tensors to the original GenAI builder and routes tensors with qweight through Olive’s implementation.
• Makes the embedding wrapper idempotent so repeated patch calls do not stack wrappers.
• Extends unit coverage for both routing paths and repeated patching.

File	Description
test/​passes/​onnx/​test_model_builder.py	Verifies native and quantized embedding routing and patch idempotency.
olive/​passes/​onnx/​model_builder.py	Adds an idempotent wrapper that preserves native embedding construction.

_{Review effort: Balanced}

[Yuri Khrustalev (ykhrustalev)]

#2692 makes the same make_embedding change and adds two things. First, a guard that keeps a separate embedding table where onnxruntime-genai 0.17.1 reads a tied LM head back wrong: with the default algorithm, is_symmetric=false and matmul_mixed_precision=last_matmul:int8 builds fail to load, and matmulnbits_weights_prepacked=1 on CUDA generates garbage (Qwen2.5-0.5B-Instruct). Second, an end-to-end test that builds a tied tiny Llama. Once this PR merges, I can cut #2692 down to the guard and the test, or close it if Olive will require a newer onnxruntime-genai. Which do you prefer?