# onnx_qnn

Python API: `depthai_nodes.runtime.onnx_qnn`

Create ONNX Runtime sessions on the OAK4 Hexagon DSP.

## Functions

### onnx_qnn_session

```python
def onnx_qnn_session(model_path, *, fp16=True, cache_context=True, performance_mode='burst', fallback_to_cpu=True, runtime_fallback='ort', device_wait_s=60, ep_options=None, session_options=None, verbose=False):
```

Create an ONNX Runtime session running on the OAK4 DSP (HTP).

This helper is intended for the `onnxruntime` variant of `oakapp-base`. That image provides the FastRPC setup, QNN plugin, and
required device nodes.

Parameters

 * `model_path`: path to a .onnx model. Inputs must have static shapes.
 * `fp16`: run fp32 graphs in fp16 on the HTP (no-op for QDQ int8 models).
 * `cache_context`: cache the compiled QNN graph (EPContext) next to the model so subsequent session creations skip HTP graph
   compilation.
 * `performance_mode`: QNN HTP performance mode (for example, `burst`).
 * `fallback_to_cpu`: if False, raise when the DSP is unavailable during session creation or any node cannot be placed on it. If
   True, return a CPU session when no QNN device is available.
 * `runtime_fallback`: behavior after a QNN execution error. `"ort"` keeps ONNX Runtime's automatic fallback behavior; `"raise"`
   disables it so the caller can handle the error explicitly.
 * `device_wait_s`: seconds to wait for a QNN device to appear during startup. Set to 0 to check once.
 * `ep_options`: extra QNN EP provider options (dict), merged last.
 * `session_options`: pre-configured ort.SessionOptions to extend.
 * `verbose`: enable verbose ORT logging.

Returns

 * onnxruntime.InferenceSession

## Attributes

### logger