feat(export): normalize exported ONNX in-place via optimize_onnx

[vortex-captain]

Fixes #677.

Summary

• Adds a post-export normalization step inside export_pytorch(), the single function used by every export path (the winml export CLI, build_hf_model, the winml build command, and direct Python API callers), so normalization runs automatically for all of them.
• New helper _normalize_exported_model(output_path) -> bool in src/winml/modelkit/export/pytorch.py:
  • logs Normalizing model
  • runs optimize_onnx() into a fresh tempfile.mkdtemp()
  • on success, replaces the original .onnx (and .data sidecar, if any) via copy_onnx_model() — confirmed to overwrite both files correctly by new tests in tests/unit/onnx/test_external_data.py
  • on failure, logs a warning and leaves the original export untouched
  • cleans up the temp directory in a finally clause either way
  • records the outcome in stats["model_normalization_succeeded"]

Tests

• tests/unit/export/test_pytorch_export.py — success path populates value_info; mocked-failure path leaves the model un-shape-inferenced and reports model_normalization_succeeded=False.
• tests/unit/onnx/test_external_data.py — two new overwrite tests covering copy_onnx_model against a pre-existing dst with and without external data sidecars.

[timenick]

Six findings — four substantive (perf, API contract, error handling, atomicity claim) and two test gaps. None are strict blockers, but #1 and #4 in particular are worth addressing before this lands as the default for every export path.

[DingmaomaoBJTU]

Code review

Found 2 issues:

1. _normalize_exported_model is called outside the warnings.catch_warnings() block, leaking optimizer/shape-inference warnings to callers. A previous PR (Centralize ONNX export warning suppression in export_pytorch() #460) explicitly centralized all warning suppression into this block to prevent stderr noise; the new call undoes that.

winml-cli/src/winml/modelkit/export/pytorch.py

Lines 96 to 100 in 39e73f7

	)
	stats["model_normalization_succeeded"] = _normalize_exported_model(output_path)
	return stats

2. Normalization is unconditional — there is no parameter (e.g. normalize: bool = True) to skip it. export_pytorch is the single shared entry point for all export paths, so callers that need the raw torch-exporter output (debugging, downstream pipelines, tests asserting exporter semantics) have no escape hatch.

winml-cli/src/winml/modelkit/export/pytorch.py

Lines 87 to 100 in 39e73f7

	with warnings.catch_warnings():
	warnings.filterwarnings("ignore")
	stats = exporter.export(
	model=model,
	output_path=str(output_path),
	export_config=export_config,
	model_name_or_path=model_name_or_path,
	task=task,
	**kwargs,
	)
	stats["model_normalization_succeeded"] = _normalize_exported_model(output_path)
	return stats

🤖 Generated with Claude Code

_{- If this code review was useful, please react with 👍. Otherwise, react with 👎.}