Function make_quantizer

backends/qualcomm/export_utils.py:652–680

(
    quant_dtype: Optional[QuantDtype] = QuantDtype.use_8a8w,
    custom_annotations=(),
    per_channel_conv=True,
    per_channel_linear=False,
    per_channel_embedding=False,
    act_observer=MovingAverageMinMaxObserver,
    act_symmetric=False,
    is_qat=False,
    submodule_qconfig_list: Optional[List[Tuple[Callable, ModuleQConfig]]] = None,
    backend=QnnExecuTorchBackendType.kHtpBackend,
    soc_model="SM8750",
    eps=None,
)

def make_quantizer(
    quant_dtype: Optional[QuantDtype] = QuantDtype.use_8a8w,
    custom_annotations=(),
    per_channel_conv=True,
    per_channel_linear=False,
    per_channel_embedding=False,
    act_observer=MovingAverageMinMaxObserver,
    act_symmetric=False,
    is_qat=False,
    submodule_qconfig_list: Optional[List[Tuple[Callable, ModuleQConfig]]] = None,
    backend=QnnExecuTorchBackendType.kHtpBackend,
    soc_model="SM8750",
    eps=None,
):
    quantizer = QnnQuantizer(backend=backend, soc_model=getattr(QcomChipset, soc_model))
    quantizer.add_custom_quant_annotations(custom_annotations)
    quantizer.set_default_quant_config(
        quant_dtype,
        is_qat=is_qat,
        is_conv_per_channel=per_channel_conv,
        is_linear_per_channel=per_channel_linear,
        is_embedding_per_channel=per_channel_embedding,
        act_observer=act_observer,
        act_symmetric=act_symmetric,
        eps=eps,
    )
    submodule_qconfig_list = submodule_qconfig_list or []
    quantizer.set_submodule_qconfig_list(submodule_qconfig_list)
    return quantizer


def get_lpai_target_env(qnn_config: QnnConfig):