Function lower_to_executorch

examples/models/sortformer/export_sortformer.py:300–353

(programs, metadata=None, backend="portable")

def lower_to_executorch(programs, metadata=None, backend="portable"):
    if backend == "xnnpack":
        from executorch.backends.xnnpack.partition.xnnpack_partitioner import (
            XnnpackPartitioner,
        )

        print("\nLowering to ExecuTorch with XNNPACK...")
        partitioner = {}
        for key in programs.keys():
            if key == "preprocessor":
                partitioner[key] = []
            else:
                partitioner[key] = [XnnpackPartitioner()]
    elif backend in ("cuda", "cuda-windows"):
        from executorch.backends.cuda.cuda_backend import CudaBackend
        from executorch.backends.cuda.cuda_partitioner import CudaPartitioner
        from executorch.exir.backend.compile_spec_schema import CompileSpec

        print(
            f"\nLowering to ExecuTorch with CUDA{' (Windows)' if backend == 'cuda-windows' else ''}..."
        )
        partitioner = {}
        for key in programs.keys():
            if key == "preprocessor":
                partitioner[key] = []
                continue
            compile_specs = [CudaBackend.generate_method_name_compile_spec(key)]
            if backend == "cuda-windows":
                compile_specs.append(CompileSpec("platform", b"windows"))
            partitioner[key] = [CudaPartitioner(compile_specs)]
    else:
        print("\nLowering to ExecuTorch...")
        partitioner = []

    constant_methods = {}
    if metadata:
        for key, value in metadata.items():
            constant_methods[key] = value

    et_prog = to_edge_transform_and_lower(
        programs,
        partitioner=partitioner,
        compile_config=EdgeCompileConfig(
            _check_ir_validity=False,
            _skip_dim_order=True,
        ),
        constant_methods=constant_methods if constant_methods else None,
    )
    return et_prog.to_executorch(
        config=ExecutorchBackendConfig(
            extract_delegate_segments=True,
            memory_planning_pass=MemoryPlanningPass(alloc_graph_input=False),
        ),
    )


def main():