(Optional) In-depth Convergence

In-depth hardware convergence helps reduce DDR access and improve the energy efficiency ratio (EER) during model inference. Currently, only the following scenarios are supported:

• When the shape is variable before compilation, call SetTuningStrategy to set the model optimization strategy to ON_DEVICE_TUNING.
• When in-depth convergence is enabled through dynamic update of the NPU operator library, call SetTuningStrategy to set the model optimization strategy to ON_DEVICE_PREPROCESS_TUNING.