Link to this sectionReference for ultralytics/nn/backends/qnn.py#

class QNNBackend(BaseBackend):

def forward(self, im: torch.Tensor) -> list:
    """Run inference on the Qualcomm QNN runtime.

    Args:
        im (torch.Tensor): Input image tensor in BCHW format, normalized to [0, 1].

    Returns:
        (list): Model predictions as a list of output arrays.
    """
    if self.nhwc:
        im = im.permute(0, 2, 3, 1)
    return self.session.run(self.output_names, {self.session.get_inputs()[0].name: im.cpu().numpy()})

def load_model(self, weight: str | Path) -> None:
    """Load a QNN context-binary model with ONNX Runtime's QNN Execution Provider plugin.

    Args:
        weight (str | Path): Path to the `*_qnn.onnx` file.

    Raises:
        OSError: If the QNN Execution Provider cannot be registered (e.g. not on Snapdragon hardware).
    """
    check_requirements("onnxruntime-qnn")
    import onnxruntime

    from ultralytics.utils.export.qnn import qnn_library_paths

    onnx_file = Path(weight)
    LOGGER.info(f"Loading {onnx_file} for Qualcomm QNN inference...")

    # Register the QNN EP (libraries resolved from the plugin helper or the onnxruntime/capi bundle) and select
    # it; ep_library is None when QNN is already built into ONNX Runtime and needs no plugin registration
    ep_name = "QNNExecutionProvider"
    ep_library, htp_backend = qnn_library_paths()
    ep_options = {"backend_path": htp_backend}
    options = onnxruntime.SessionOptions()
    if ep_library:
        onnxruntime.register_execution_provider_library(ep_name, ep_library)
        devices = [d for d in onnxruntime.get_ep_devices() if d.ep_name == ep_name]
        if not devices:
            raise OSError(
                "QNN Execution Provider registered but no QNN devices were found. Run on a Qualcomm Snapdragon "
                "device with 'onnxruntime-qnn' installed."
            )
        options.add_provider_for_devices(devices, ep_options)
        self.session = onnxruntime.InferenceSession(str(onnx_file), sess_options=options)
    else:
        self.session = onnxruntime.InferenceSession(
            str(onnx_file), sess_options=options, providers=[ep_name], provider_options=[ep_options]
        )
    self.output_names = [x.name for x in self.session.get_outputs()]
    shape = self.session.get_inputs()[0].shape  # channel-last exports take [N, H, W, C] input
    self.nhwc = len(shape) == 4 and shape[3] in {1, 3} and shape[1] not in {1, 3}

    metadata_map = self.session.get_modelmeta().custom_metadata_map
    if metadata_map:
        self.apply_metadata(dict(metadata_map))