Replace onnxruntime.tools.symbolic_shape_infer with onnx-shape-inference

[Copilot]

Describe your changes

Swaps onnxruntime.tools.symbolic_shape_infer.SymbolicShapeInference for the onnx-shape-inference package — a port and improvement of the onnxruntime symbolic shape inference engine (as used by mobius) that operates on ONNX IR without serialization overhead.

Following review feedback, the branch was brought up to date with main (#2558, which migrated OnnxDAG usage to onnx_ir) and the helper now operates directly on the IR.

• Shared helper — added run_symbolic_shape_inference(ir_model) in olive/passes/onnx/common.py, which takes and returns an ir.Model and calls infer_symbolic_shapes directly on the IR.

def run_symbolic_shape_inference(ir_model: ir.Model) -> ir.Model:
    from onnx_shape_inference import infer_symbolic_shapes

    return infer_symbolic_shapes(ir_model)

• Call sites — SplitModel (split.py) runs inference directly on the IR (no proto round-trip). _fix_output_shapes (common.py) and scripts/transform_dla_model.py remain proto-oriented and adapt at the proto boundary via ir.from_proto / ir.to_proto.
• Dependencies — added onnx-shape-inference; bumped onnx_ir to >=0.2.0 (its minimum).
• Out of scope — mixed_precision.py uses a different engine (onnxruntime.transformers.shape_infer_helper), left unchanged.

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

[justinchuby]

@copilot rebase from main then operate directly on the ir

[Copilot]

@copilot rebase from main then operate directly on the ir

Done in dbbd9f1. Merged latest main (#2558) and run_symbolic_shape_inference now takes/returns an ir.Model, calling infer_symbolic_shapes directly on the IR. SplitModel runs inference directly on the IR (no proto round-trip); _fix_output_shapes and scripts/transform_dla_model.py only adapt at the proto boundary.