Method to_executorch

exir/program/_program.py:1731–1844

Transforms the program to the ExecuTorch backend. Args: config: An optional argument used to provide greater control over the transformation to the ExecuTorch backend. Returns: ExecutorchProgramManager: A manager representing the sta

(  # noqa (FLAKE8) C901
        self,
        config: Optional[ExecutorchBackendConfig] = None,
    )

    @et_logger("to_executorch")
    def to_executorch(  # noqa (FLAKE8) C901
        self,
        config: Optional[ExecutorchBackendConfig] = None,
    ) -> "ExecutorchProgramManager":
        """
        Transforms the program to the ExecuTorch backend.

        Args:
            config: An optional argument used to provide greater control over
                the transformation to the ExecuTorch backend.

        Returns:
            ExecutorchProgramManager: A manager representing the state of the EdgeProgramManager
            after it has been transformed to the ExecuTorch backend.
        """
        config = config if config else ExecutorchBackendConfig()
        execution_programs: Dict[str, ExportedProgram] = {}
        for name, program in self._edge_programs.items():
            if config.do_quant_fusion_and_const_prop:
                if program.graph_signature.backward_signature is not None:
                    raise Exception(
                        "Cannot run do_quant_fusion_and_const_prop on a graph with a backward signature intended for on-device training."
                        " Please set do_quant_fusion_and_const_prop to False in the ExecutorchBackendConfig."
                    )
                program = quant_fusion_and_const_prop_pass(program)
            if config.run_reinplace_pass:
                program = reinplace_pass(program)
            program = weights_to_outputs_pass(program)
            program = unsafe_remove_auto_functionalized_pass(program)
            gm, new_signature = insert_write_back_for_buffers_pass(program)
            new_gm = program.graph_module
            for p in edge_to_executorch_passes(config, name):
                new_gm_res = p(new_gm)
                assert new_gm_res is not None
                new_gm = new_gm_res.graph_module
                if isinstance(p, SpecPropPass):
                    # Note that this is a hacky way to get around the fact that
                    # placeholder nodes corresponding to the parameters of the graph module
                    # shall not participate in memory planning. It increases runtime memory
                    # footprint.
                    # Proper way would be to have ExportPass work with ExportedProgram
                    # instead of GraphModule. This is because ExportPass should work
                    # on top of the export artifact of torch.export whichi s ExportedProgram.
                    # Working with GraphModule does not provide all the information contained
                    # in the ExportedProgram
                    # TODO(who?)
                    p.update_placeholder_tensor_specs(program, new_gm)

            # Tag constant weights.
            if (
                isinstance(config.external_constants, bool)
                and config.external_constants
            ):
                new_gm_res = external_constants_pass(new_gm)
                new_gm = new_gm_res.graph_module
            elif callable(config.external_constants):
                new_gm_res = external_constants_pass(new_gm, config.external_constants)
                new_gm = new_gm_res.graph_module