o diY-@sddlZddlmZddlmZddlmZmZmZm Z ddl m Z ddl m Z e r2ddlmZnddlmZGd d d ZGd d d eeZGd ddeeZGdddeeZddZe eGdddZe eGdddZe eGdddZe eGdddZe eGdddZeGdddZeGdd d ZeGd!d"d"ZeGd#d$d$eZeGd%d&d&eeZe eGd'd(d(eeeZ dS))N)field)Enum)DictListOptionalUnion)is_pydantic_available)generate_doc_dataclass) dataclassc@s<eZdZgdZededefddZedefddZdS) APIFeaturesManager)text-classificationztoken-classificationzquestion-answeringzimage-classification model_typetaskc Csn|}|tjvrt|dttjd|d|tj|vr5t|d|dtj|d|ddS)N is not supported yet. Only ' are supported. If you want to support ) please propose a PR or open up an issue.z is not supported yet for model z. Only )lowerr _SUPPORTED_MODEL_TYPEKeyErrorlistkeys)rrr`/lsinfo/ai/hellotax_ai/llm_service/venv_embed/lib/python3.10/site-packages/optimum/utils/runs.pycheck_supported_model_task_pair"s$   z2APIFeaturesManager.check_supported_model_task_paircCs*|tjvrt|dtjd|ddS)Nrrr)r _SUPPORTED_TASKSrrrrrcheck_supported_task2s z'APIFeaturesManager.check_supported_taskN)__name__ __module__ __qualname__r staticmethodstrrrrrrrr s r c@seZdZdZdS) Frameworks onnxruntimeN)rrr r$rrrrr#<sr#c@seZdZdZdZdZdS)CalibrationMethodsminmax percentileentropyN)rrr r&r'r(rrrrr%@sr%c@seZdZdZdZdS)QuantizationApproachstaticdynamicN)rrr r*r+rrrrr)Fsr)cCs(tdkrdStdkrdStd)NLinuxlscpuDarwinzsysctl -a | grep machdep.cpuzOS not supported.)platformsystemNotImplementedErrorrrrrcpu_info_commandKs  r2c@seZdZUdZeddidZeed<eddidZe ed<eddd id Z e e ed <eddd id Z e eed <edddid Ze e ed<dS) CalibrationzBParameters for post-training calibration with static quantization. descriptionzDCalibration method used, either "minmax", "entropy" or "percentile".metadatamethodzVNumber of examples to use for the calibration step resulting from static quantization.num_calibration_samplesNz:The percentile used for the percentile calibration method.defaultr6 calibration_histogram_percentilezjWhether to compute the moving average of the minimum and maximum values for the minmax calibration method.calibration_moving_averagezConstant smoothing factor to use when computing the moving average of the minimum and maximum values. Effective only when the selected calibration method is minmax and `calibration_moving_average` is set to True.#calibration_moving_average_constant)rrr __doc__rr7r%__annotations__r8intr;rfloatr<boolr=rrrrr3Ts, r3c@sNeZdZUedddidZeeed<edddidZeeed<d d Z d S) FrameworkArgs r4z,ONNX opset version to export the model with.r9opsetrzONNX optimization level.optimization_levelcCs8|jdks Jd|j|jdvsJd|jdS)NzUnsupported OnnxRuntime opset: )rrcz,Unsupported OnnxRuntime optimization level: )rErFselfrrr __post_init__xs  zFrameworkArgs.__post_init__N) rrr rrErr@r?rFrLrrrrrCrs  rCc@seZdZUdZeddidZeed<eddidZeed<eddidZ e ee d effed <edd idZ e eed <ed dd idZeeed<ed ddidZeeed<d S) DatasetArgsz"Parameters related to the dataset.r4z9Path to the dataset, as in `datasets.load_dataset(path)`.r5pathz0Dataset split used for evaluation (e.g. "test"). eval_splitzZDataset columns used as input data. At most two, indicated with "primary" and "secondary".N data_keysz5Dataset column used for references during evaluation.ref_keysz?Name of the dataset, as in `datasets.load_dataset(path, name)`.r9namez2Dataset split used for calibration (e.g. "train").calibration_split)rrr r>rrNr"r?rOrPrrrQrrRrrSrrrrrMs rMc@s.eZdZUdZedddidZeeed<dS)TaskArgszTask-specific parameters.Nr4zVText classification specific. Set whether the task is regression (output = one float).r9 is_regression) rrr r>rrUrrBr?rrrrrTs rTc@sJeZdZUdZedddidZeeed<edddidZ eeed <d S) BenchmarkTimeArgs%Parameters related to time benchmark.r4z+Duration in seconds of the time evaluation.r9duration z?Number of warmup calls to forward() before the time evaluation. warmup_runsN) rrr r>rrYrr@r?r[rrrrrVs rVc@seZdZUeddidZeed<eddidZeed<eddidZe ed<edd idZ e ed <edd idZ e ed <edd idZeed<dS)_RunBaser4z7Name of the model hosted on the Hub to use for the run.r5model_name_or_pathzTask performed by the model.rz.Whether to use dynamic or static quantization.quantization_approachzDDataset to use. Several keys must be set on top of the dataset name.datasetz0Name of the framework used (e.g. "onnxruntime"). frameworkzFramework-specific arguments.framework_argsN)rrr rr]r"r?rr^r)r_rMr`r#rarCrrrrr\s r\c@seZdZUeddddidZeeeed<eddddidZ eeeed <ed dd id Z ee ed <edddid Z ee ed<edddid Zeeed<ed ddid Zee ed<edddid Zeeed<eeddid Zeeed<dS) _RunDefaultscCddgS)NAddMatMulrrrrrz_RunDefaults.r4zWOperators to quantize, doing no modifications to others (default: `["Add", "MatMul"]`).default_factoryr6operators_to_quantizecCsgdS)N) layernormgeluresidualgathersoftmaxrrrrrrfrgzsSpecific nodes to exclude from being quantized (default: `['layernorm', 'gelu', 'residual', 'gather', 'softmax']`).node_exclusionFz3Whether to quantize per channel (default: `False`).r9 per_channelNzrrxrr"r?rrrrrws rwc@sVeZdZUeddddidZeeeed<eddddidZ eeeed <d S) _RunConfigDefaultscCrc)Nrrrrrrfrgz_RunConfigDefaults.r4z:Batch sizes to include in the run to measure time metrics.rh batch_sizescCsdgS)NrrrrrrfszrL __classcell__rrrrrsr)!r/ dataclassesrenumrtypingrrrrr docr pydantic.dataclassesr r r"r#r%r)r2r3rCrMrTrVr\rbrwryrrrrrrsT        %