torch.onnx.testing

Created On: Sep 10, 2025 | Last Updated On: Sep 10, 2025

Utilities to aid in testing exported ONNX models.

torch.onnx.testing.assert_onnx_program(program, *, rtol=None, atol=None, args=None, kwargs=None, strategy='TorchExportNonStrictStrategy', backend='onnxruntime')

Assert that the ONNX model produces the same output as the PyTorch ExportedProgram.

Parameters:

• program (_onnx_program.ONNXProgram) – The ONNXProgram to verify.

• rtol (float | None) – Relative tolerance.

• atol (float | None) – Absolute tolerance.

• args (tuple[Any, ...] | None) – The positional arguments to pass to the program. If None, the default example inputs in the ExportedProgram will be used.

• kwargs (dict[str, Any] | None) – The keyword arguments to pass to the program. If None, the default example inputs in the ExportedProgram will be used.

• strategy (str | None) – Assert the capture strategy used to export the program. Values can be class names like “TorchExportNonStrictStrategy”. If None, the strategy is not asserted.

• backend (Literal['onnxruntime', 'reference']) – The backend to use for evaluating the ONNX program. Supported values are “onnxruntime” and “reference”.