Method to_executorch

extension/llm/export/builder.py:477–529

Lower the model to executorch and get an ExecutorchProgram.

(
        self,
        passes: Optional[List[ExportPass]] = None,
        external_constants_tag: Optional[
            Callable[[torch.fx.Node], Optional[str]]
        ] = None,
        share_mutable_buffers: bool = False,
    )

        return self

    def to_executorch(
        self,
        passes: Optional[List[ExportPass]] = None,
        external_constants_tag: Optional[
            Callable[[torch.fx.Node], Optional[str]]
        ] = None,
        share_mutable_buffers: bool = False,
    ) -> "LLMEdgeManager":
        """
        Lower the model to executorch and get an ExecutorchProgram.
        """
        to_executorch_passes = []
        if passes:
            # pyre-fixme[6]: In call `list.extend`, for 1st positional argument,
            # expected `Iterable[Union[ConvertToLinearPass, QuantFusionPass]]` but
            # got `List[ExportPass]
            to_executorch_passes.extend(passes)

        assert self.edge_manager, "Need to run export_to_edge() first"

        # If there are Linear operations left in the graph, let's execute
        # them with the optimized op_linear rather than materializing a
        # transpose followed by a regular op_mm.
        # TODO: ConvertToLinearPass is not a sound pass and must be called before
        # const propagation.  It requires fixing:
        # https://github.com/pytorch/executorch/issues/10499
        self.edge_manager.transform([ConvertToLinearPass()])

        self.export_program = self.edge_manager.to_executorch(
            ExecutorchBackendConfig(
                extract_delegate_segments=True,
                # pyre-fixme[6]: In call `ExecutorchBackendConfig.__init__`, for
                # argument `passes`, expected `List[typing.Callable[[GraphModule],
                # Optional[PassResult]]]` but got `List[Union[ConvertToLinearPass,
                # QuantFusionPass]]`.
                passes=to_executorch_passes,
                do_quant_fusion_and_const_prop=True,
                memory_planning_pass=MemoryPlanningPass(
                    alloc_graph_input=False,
                    share_mutable_buffers=share_mutable_buffers,
                ),
                sym_shape_eval_pass=ConstraintBasedSymShapeEvalPass(),
                external_constants=external_constants_tag,
            )
        )
        logging.info(
            "Required memory for activation in bytes: {}".format(
                self.export_program._emitter_output.program.execution_plan[
                    0
                ].non_const_buffer_sizes
            ),
        )
        return self

    def save_to_pte(self, output_name: str) -> None:
        """