Fix ort genai model builder compatible issue

[Xiaoyu Z (xiaoyu-work)]

Describe your changes

Fix ort genai model builder compatible issue

Checklist before requesting a review

• Add unit tests for this change.
• Make sure all tests can pass.
• Update documents if necessary.
• Lint and apply fixes to your code by running lintrunner -a
• Is this a user-facing change? If yes, give a description of this change to be included in the release notes.

(Optional) Issue link

[Copilot]

🟡 Changes recommended

Two unresolved critical findings remain in olive/passes/onnx/model_builder.py.

Get a fresh assessment by requesting another Copilot review.

Pull request overview

This pull request updates ORT GenAI ModelBuilder compatibility for newer loader layouts and quantized models, with expanded regression coverage.

Changes:

• Updates quantization loader integration and group-size handling.
• Adds loader-patching and quantization tests.
• Aligns Qwen3 token configuration.
• Cleans up test requirements formatting.

File summaries

File	Summary
test/requirements-test-cpu.txt	Removes extraneous formatting.
test/passes/onnx/test_model_builder.py	Adds loader and quantization regression coverage.
test/cli/test_cli_test_model_smoke.py	Validates matching Qwen3 token IDs.
olive/passes/onnx/model_builder.py	Updates loader patching and quantized tensor handling. Critical (3 votes, line 493): unconditional newer loader imports break older ORT GenAI versions. Critical (1 vote, line 585): group_size=-1 must be normalized to the input dimension.

Review details

Suppressed comments (1)

olive/passes/onnx/model_builder.py:585

• Changing the assignment from _group_size to group_size is only valid for the new loader class, which added a writable setter. The legacy QuantizedTensorModule exposes group_size as a read-only property backed by _group_size, so a legacy fallback would raise AttributeError while loading every quantized tensor; use a compatibility helper or assign the appropriate backing field for each loader version.

                for q_attr, q_value in [("bits", local_bits), ("group_size", local_group_size)]:

• Files reviewed: 4/4 changed files
• Comments generated: 2
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.