neural_compressor.conf.pythonic_config

Module Contents

Classes

QuantizationConfig

Basic class for quantization config. Inherited by PostTrainingQuantConfig and QuantizationAwareTrainingConfig.

class neural_compressor.conf.pythonic_config.QuantizationConfig(inputs=[], outputs=[], backend='default', device='cpu', approach='post_training_static_quant', calibration_sampling_size=[100], op_type_dict=None, op_name_dict=None, strategy='basic', strategy_kwargs=None, objective='performance', timeout=0, max_trials=100, performance_only=False, reduce_range=None, use_bf16=True, quant_level='auto', accuracy_criterion=accuracy_criterion, use_distributed_tuning=False)

Basic class for quantization config. Inherited by PostTrainingQuantConfig and QuantizationAwareTrainingConfig.

Parameters:

• inputs – Inputs of model, only required in tensorflow.

• outputs – Outputs of model, only required in tensorflow.

• backend – Backend for model execution. Support ‘default’, ‘itex’, ‘ipex’, ‘onnxrt_trt_ep’, ‘onnxrt_cuda_ep’

• domain – Model domain. Support ‘auto’, ‘cv’, ‘object_detection’, ‘nlp’ and ‘recommendation_system’. Adaptor will use specific quantization settings for different domains automatically, and explicitly specified quantization settings will override the automatic setting. If users set domain as auto, automatic detection for domain will be executed.

• recipes –

  Recipes for quantiztaion, support list is as below. ‘smooth_quant’: whether do smooth quant ‘smooth_quant_args’: parameters for smooth_quant ‘fast_bias_correction’: whether do fast bias correction ‘weight_correction’: whether do weight correction ‘gemm_to_matmul’: whether convert gemm to matmul and add, only valid for onnx models ‘graph_optimization_level’: support ‘DISABLE_ALL’, ‘ENABLE_BASIC’, ‘ENABLE_EXTENDED’, ‘ENABLE_ALL’

  only valid for onnx models

  ’first_conv_or_matmul_quantization’: whether quantize the first conv or matmul ‘last_conv_or_matmul_quantization’: whether quantize the last conv or matmul ‘pre_post_process_quantization’: whether quantize the ops in preprocess and postprocess ‘add_qdq_pair_to_weight’: whether add QDQ pair for weights, only vaild for onnxrt_trt_ep ‘optypes_to_exclude_output_quant’: don’t quantize output of specified optypes ‘dedicated_qdq_pair’: whether dedicate QDQ pair, only vaild for onnxrt_trt_ep

• quant_format – Support ‘default’, ‘QDQ’ and ‘QOperator’, only required in ONNXRuntime.

• device – Support ‘cpu’ and ‘gpu’.

• calibration_sampling_size – Number of calibration sample.

• op_type_dict –

  Tuning constraints on optype-wise for advance user to reduce tuning space. User can specify the quantization config by op type: example: {

  ’Conv’: {

  ‘weight’: {

  ‘dtype’: [‘fp32’]

  }, ‘activation’: {

  ’dtype’: [‘fp32’]

  }

  }

  }

• op_name_dict –

  Tuning constraints on op-wise for advance user to reduce tuning space. User can specify the quantization config by op name: example: {

  ”layer1.0.conv1”: {

  “activation”: {

  “dtype”: [“fp32”]

  }, “weight”: {

  ”dtype”: [“fp32”]

  }

  },

  }

• strategy – Strategy name used in tuning, Please refer to docs/source/tuning_strategies.md.

• strategy_kwargs – Parameters for strategy, Please refer to docs/source/tuning_strategies.md.

• objective – Objective with accuracy constraint guaranteed, support ‘performance’, ‘modelsize’, ‘footprint’. Please refer to docs/source/objective.md. Default value is ‘performance’.

• timeout – Tuning timeout (seconds). default value is 0 which means early stop

• max_trials – Max tune times. default value is 100. Combine with timeout field to decide when to exit

• performance_only – Whether do evaluation

• reduce_range – Whether use 7 bit to quantization.

• example_inputs – Used to trace PyTorch model with torch.jit/torch.fx.

• excluded_precisions – Precisions to be excluded, Default value is empty list. Neural compressor enable the mixed precision with fp32 + bf16 + int8 by default. If you want to disable bf16 data type, you can specify excluded_precisions = [‘bf16].

• quant_level – Support auto, 0 and 1, 0 is conservative strategy, 1 is basic or user-specified strategy, auto (default) is the combination of 0 and 1.

• accuracy_criterion – Accuracy constraint settings.

• use_distributed_tuning – Whether use distributed tuning or not.