What does this PR do?

Adds an input_check: bool = False argument to LightningModule.to_onnx. When set, after torch.onnx.export returns we load the saved model and run onnx.checker.check_model on it, so the caller gets a ValidationError instead of finding out at deploy time that the exported protobuf is malformed.

Default behavior is unchanged: input_check=False.

Refs #7279.

Notes on scope

The original issue lists two things: the spec check (onnx.checker.check_model) and a PyTorch-vs-onnxruntime output comparison. I only did the first one here — the runtime comparison needs onnxruntime as a hard dep and a way to deal with multi-output / dict-output models, which feels like a separate PR. Happy to add it as a follow-up if you'd like.

Behavior

• file_path is a str/Path → loaded with onnx.load, then checked.
• file_path is BytesIO → loaded with onnx.load_model_from_string; the buffer's cursor position is restored so the caller sees the same state as without the flag.
• file_path is None → raises ValueError; there is nothing to load.
• dynamo=True → raises ValueError; that path returns an ONNXProgram rather than a standalone protobuf file we can hand to the checker.

Tests

Added to tests/tests_pytorch/models/test_onnx.py:

• test_input_check_runs_onnx_checker — happy path for both file and BytesIO.
• test_input_check_raises_without_file_path — ValueError when file_path=None.
• test_input_check_detects_invalid_model — monkeypatches onnx.checker.check_model to fail; verifies the error propagates.
• test_input_check_rejects_dynamo — ValueError when combined with dynamo=True.

All onnx tests pass locally on macOS (torch 2.12.0, onnx 1.21.0, onnxruntime 1.26.0):

13 passed, 5 skipped

Ruff lint + format pass.

Before submitting

• Was this discussed/agreed via a GitHub issue? — yes, Add checks for model spec and matching output values in to_onnx() method #7279
• Did you read the contributor guideline, Pull Request section?
• Did you make sure your PR does only one thing, instead of bundling different changes together?
• Did you make sure to update the documentation with your changes? (added a short note in docs/source-pytorch/deploy/production_advanced.rst)
• Did you write any new necessary tests?
• Did you verify new and existing tests pass locally with your changes?
• Did you list all the breaking changes introduced by this pull request? — none, new kwarg defaults to False.
• Did you update the CHANGELOG?

PR review

Anyone in the community is welcome to review the PR.

📚 Documentation preview 📚: https://pytorch-lightning--21716.org.readthedocs.build/en/21716/