{ "model_type": "ONNX", "npu_mode": "NPU3", "target_hardware": "AX650", "onnx_opt": { "enable_onnxsim": false }, "quant": { "input_configs": [ { "tensor_name": "hidden", "calibration_dataset": "/data/calib/l2_attn_fused/hidden.tar", "calibration_format": "Numpy", "calibration_size": -1 }, { "tensor_name": "mask", "calibration_dataset": "/data/calib/l2_attn_fused/mask.tar", "calibration_format": "Numpy", "calibration_size": -1 }, { "tensor_name": "indices", "calibration_dataset": "/data/calib/l2_attn_fused/indices.tar", "calibration_format": "Numpy", "calibration_size": -1 } ], "layer_configs": [ { "start_tensor_names": [ "DEFAULT" ], "end_tensor_names": [ "DEFAULT" ], "data_type": "U16" }, { "layer_names": [ "/operator_norm/Mul", "/operator_norm/ReduceMean", "/operator_norm/Add", "/operator_norm/Sqrt", "/operator_norm/Div", "/operator_norm/Mul_1", "/self_attn/q_layernorm/Mul", "/self_attn/q_layernorm/ReduceMean", "/self_attn/q_layernorm/Add", "/self_attn/q_layernorm/Sqrt", "/self_attn/q_layernorm/Div", "/self_attn/q_layernorm/Mul_1", "/self_attn/k_layernorm/Mul", "/self_attn/k_layernorm/ReduceMean", "/self_attn/k_layernorm/Add", "/self_attn/k_layernorm/Sqrt", "/self_attn/k_layernorm/Div", "/self_attn/k_layernorm/Mul_1", "/ffn_norm/Mul", "/ffn_norm/ReduceMean", "/ffn_norm/Add", "/ffn_norm/Sqrt", "/ffn_norm/Div", "/ffn_norm/Mul_1", "/Add", "/Add_1", "/self_attn/Add_2" ], "data_type": "FP32" } ], "calibration_method": "MinMax", "precision_analysis": true, "precision_analysis_method": "EndToEnd", "conv_bias_data_type": "FP32", "enable_smooth_quant": false }, "input_processors": [ { "tensor_name": "hidden", "src_dtype": "FP32" }, { "tensor_name": "mask", "src_dtype": "FP32" }, { "tensor_name": "indices", "src_dtype": "S32" } ], "compiler": { "check": 0 } }