Function build_executorch_binary

backends/qualcomm/export_utils.py:502–649

A function to generate an ExecuTorch binary for Qualcomm platforms. Attributes: model (torch.nn.Module): The model to be converted into an ExecuTorch binary. qnn_config: (QnnConfig): A config class that saves qnn lowering and execution configuration. file_name (str)

(
    model: torch.nn.Module,  # noqa: B006
    qnn_config: QnnConfig,
    file_name: str,
    dataset: List[torch.Tensor] | Callable[[torch.fx.GraphModule], None],
    quant_dtype: Optional[QuantDtype] = None,
    custom_quantizer: Optional[QnnQuantizer] = None,
    metadata=None,
    qnn_intermediate_debugger: QNNIntermediateDebugger = None,
    passes_job=None,
    passes_dependency=None,
    qat_training_data=None,
    op_package_options: QnnExecuTorchOpPackageOptions = None,
)

def build_executorch_binary(
    model: torch.nn.Module,  # noqa: B006
    qnn_config: QnnConfig,
    file_name: str,
    dataset: List[torch.Tensor] | Callable[[torch.fx.GraphModule], None],
    quant_dtype: Optional[QuantDtype] = None,
    custom_quantizer: Optional[QnnQuantizer] = None,
    metadata=None,
    qnn_intermediate_debugger: QNNIntermediateDebugger = None,
    passes_job=None,
    passes_dependency=None,
    qat_training_data=None,
    op_package_options: QnnExecuTorchOpPackageOptions = None,
):
    """
    A function to generate an ExecuTorch binary for Qualcomm platforms.

    Attributes:
        model (torch.nn.Module): The model to be converted into an ExecuTorch binary.
        qnn_config: (QnnConfig): A config class that saves qnn lowering and execution configuration.
        file_name (str): Name for the output binary file (.pte).
        dataset (List[torch.Tensor] | Callable): A dataset for quantization calibration.
        quant_dtype (QuantDtype, optional): Data type for quantization.
        custom_quantizer (Callable, optional): Custom quantizer.
        metadata (dict, optional): An optional dictionary that maps each method name to a constant value in eager mode.
        passes_job (OrderedDict, optional): Custom passes job in to_edge_transform_and_lower, users can enable/disable specific passes or modify their attributes.
        passes_dependency (Dict, optional): A dictionary mapping each pass to its corresponding list of dependencies.
        qat_training_data (List[torch.Tensor], optional): A dataset for quantization aware training(QAT). Typically is a pair of tensors, such as [features, ground truth].
        op_package_options: Optional structure to specify op packages
            loaded and used by the backend.

    Returns:
        None: The function writes the output to a specified .pte file.
    """
    if qnn_config.pre_gen_pte:
        logging.info(
            f"Skip build_executorch_binary, using {file_name} under {qnn_config.pre_gen_pte}."
        )
        return

    sample_input = dataset[0]
    if (
        qnn_config.backend == QnnExecuTorchBackendType.kGpuBackend
        and not qnn_config.online_prepare
    ):
        raise RuntimeError(
            "Currently GPU backend only supports online_prepare. Please add --online_prepare flag."
        )
    if (
        qnn_config.backend == QnnExecuTorchBackendType.kLpaiBackend
        and qnn_config.online_prepare
    ):
        raise RuntimeError("Currently LPAI backend only supports offline_prepare.")
    backend_options = {
        QnnExecuTorchBackendType.kLpaiBackend: generate_lpai_compiler_spec(
            target_env=get_lpai_target_env(qnn_config)
        ),
        QnnExecuTorchBackendType.kGpuBackend: generate_gpu_compiler_spec(),