INNER CODE UNIT · Python

run

onnxsim/onnxsim · onnxsim/rpc/server.py:74

    def run(self, inputs: Dict[str, np.ndarray]) -> Dict[str, np.ndarray]:
        if self.session is not None:
            names = [o.name for o in self.session.get_outputs()]
            return dict(zip(names, self.session.run(None, inputs)))
        from onnxsim import backend

        return dict(backend.run_model(model_bytes_to_proto(self.model_bytes), inputs))


class _TinygradRunner:
    """Runs a model with tinygrad's ONNX frontend on a chosen tinygrad device.

    Exists so tinygrad's code generation can be benchmarked on the server's hardware through the
    same session API. ``options`` are tinygrad codegen knobs (an allow-list, e.g. ``{"BEAM": 2}``)
    applied around compilation and execution. The frontend caches Python constants between calls,
    so this suits static-shape models; a graph with data-dependent shapes needs a fresh load per
    input.
    """

View source record →

📰 Research Paper
Loading…
⏳ Fetching content…