check_model in _framework_apis is a no-op — torch.onnx.export(verify=True) reports success without validating anything

[gabrielfruet]

check_model in onnxscript/_framework_apis/torch_2_5.py is:

def check_model(model: ir.Model) -> None:
    """Check the model."""

    del model  # Unused yet

torch_2_6.py through torch_2_11.py all re-export this same function unchanged (confirmed on current main).

torch.onnx.export(..., dynamo=True, verify=True) calls this via onnxscript_apis.check_model(onnx_program.model) in torch/onnx/_internal/exporter/_core.py, prints "Check the ONNX model... ✅", and sets export_status.onnx_checker = True — regardless of whether the model is actually valid. A model that would fail onnx.checker.check_model(..., full_check=True) still gets reported as checked and passing.

Repro:

import torch

class M(torch.nn.Module):
    def forward(self, x):
        return x + 1

torch.onnx.export(M(), (torch.randn(2, 2),), "m.onnx", dynamo=True, verify=True)
# Prints "Check the ONNX model... ✅" no matter what — the check never runs.

Is real validation (e.g. wiring in onnx.checker.check_model) planned, or should the verify=True reporting reflect that this step is currently a stub?

[justinchuby]

Thanks for noticing this. No, validation is not planned. It will remain a no-op.

[gabrielfruet]

Just to understand the motivation.

Why do we have a check argument if it is a no-op? Should we add a warning then?

[justinchuby]

Initially it was intended that we ensure the model to be a "valid" model before returning. Later we realized (1) the checker may produce false positives and (2) it adds overhead for the export logic. So we decided to remove this step. But the api is maintained for compatibility.

[gabrielfruet]

Thank you for the explanation, very clarifying.