A configuration used as input to QNN ExecuTorch’s lowering API. This config initialization currently supports: 1. Provide command-line arguments paired with setup_common_args_and_variables. 2. Provide a json file that stores desired config. Attributes: backend (str): Th
| 65 | |
| 66 | @dataclass |
| 67 | class QnnConfig: |
| 68 | """ |
| 69 | A configuration used as input to QNN ExecuTorch’s lowering API. |
| 70 | This config initialization currently supports: |
| 71 | 1. Provide command-line arguments paired with setup_common_args_and_variables. |
| 72 | 2. Provide a json file that stores desired config. |
| 73 | |
| 74 | Attributes: |
| 75 | backend (str): The target backend, such as htp, gpu, etc. QnnConfig will then parse this to type QnnExecuTorchBackendType. |
| 76 | soc_model (QcomChipset): The target Qualcomm System on Chip (SoC) model. |
| 77 | build_folder (str): Path to cmake binary directory for target platform, e.g., /path/to/build-android. |
| 78 | direct_build_folder (str): Path to cmake binary directory for direct_mode. E.g., path/to/build-direct. |
| 79 | target (str): Target platform for deployment. |
| 80 | online_prepare (bool): Compose QNN graph on device if set to True. |
| 81 | shared_buffer (bool): Enables usage of shared buffer(zero-copy mechanism) between application and backend for graph I/O during runtime. |
| 82 | dump_intermediate_outputs (bool): Enables dumping model intermediate outputs. |
| 83 | profile_level (int): Level of profiling in runtime. |
| 84 | enable_x86_64: Enable x86_64 simulator execution. |
| 85 | host (str): Hostname where android device is connected. |
| 86 | device (str): Serial number for android device communicated via ADB. |
| 87 | port (int): IPC port for delivering execution result |
| 88 | ip (str): IPC address for delivering execution result. |
| 89 | skip_delegate_node_ids (str): If specified, skip delegation for the specified node based on node ids. Node ids should be separated by comma. e.g., aten_relu_default_10,aten_relu_default_2 |
| 90 | skip_delegate_node_ops (str): If specified, skip delegation for the specified op. Node ops should be separated by comma. e.g., aten.add.Tensor,aten.relu.default |
| 91 | compile_only (bool): If specified, only compile the model. |
| 92 | pre_gen_pte (str): Run the pre-generated pte in the given directory. |
| 93 | skip_push: If specified, skip pushing files to device. Assumes all required files are on device already. |
| 94 | ci (bool): This flag is for Continuous Integration(CI) purpose and is NOT recommended to turn on for typical use cases. It will use random inputs instead of real inputs. |
| 95 | seed (int): Set the seed for generating random numbers in both torch and random. |
| 96 | htp_performance_mode (QnnExecuTorchHtpPerformanceMode, optional): Option to set the performance mode for htp backend. |
| 97 | """ |
| 98 | |
| 99 | soc_model: str |
| 100 | build_folder: str |
| 101 | direct_build_folder: Optional[str] = None |
| 102 | backend: str = "htp" |
| 103 | target: str = "aarch64-android" |
| 104 | online_prepare: Optional[bool] = False |
| 105 | shared_buffer: Optional[bool] = False |
| 106 | dump_intermediate_outputs: Optional[bool] = False |
| 107 | profile_level: Optional[int] = 0 |
| 108 | enable_x86_64: Optional[bool] = False |
| 109 | host: Optional[str] = None |
| 110 | device: Optional[str] = None |
| 111 | port: Optional[str] = -1 |
| 112 | ip: Optional[str] = "" |
| 113 | skip_delegate_node_ids: Optional[str] = None |
| 114 | skip_delegate_node_ops: Optional[str] = None |
| 115 | compile_only: Optional[bool] = False |
| 116 | pre_gen_pte: Optional[str] = None |
| 117 | skip_push: Optional[bool] = False |
| 118 | ci: Optional[bool] = False |
| 119 | seed: Optional[int] = None |
| 120 | htp_performance_mode: QnnExecuTorchHtpPerformanceMode = ( |
| 121 | QnnExecuTorchHtpPerformanceMode.kHtpBurst |
| 122 | ) |
| 123 | |
| 124 | def __post_init__(self): |
no outgoing calls
no test coverage detected