Simplified PTQ Configuration File

Table 1 describes the fields in the calibration_config_pytorch.proto file. Find the file in /amct_pytorch/proto/calibration_config_pytorch.proto under the AMCT installation directory.

**Table 1** calibration_config_pytorch.proto
Parameter	Required/Optional	Type	Field	Description
AMCTConfig	-	-	-	Simplified PTQ configuration of AMCT.
	Optional	UInt32	batch_num	Batch number for quantization.
	Optional	Boolean	activation_offset	Whether to quantize activations with offset. It is a global configuration parameter. true: with offset. Activations are asymmetrically quantized. false: without offset. Activations are symmetrically quantized.
	Repeated	String	skip_layers	Layers to skip quantization.
	Repeated	String	skip_layer_types	Types of layers to skip quantization.
	Optional	NuqConfig	nuq_config	NUQ configuration.
	Optional	FakequantPrecisionMode	fakequant_precision_mode	scale_d value precision mode of the quantization custom operator in the fake-quantized model. FORCE_FP16_QUANT: converts the scale_d value from float32 to float16. Empty (default), that is, not configured. The precision of scale_d is float32.
	Optional	CalibrationConfig	common_config	Common quantization configuration, which is a global parameter. Use this configuration if a layer is not overridden by override_layer_types or override_layer_configs. Parameter priority: override_layer_configs > override_layer_types > common_config
	Repeated	OverrideLayerType	override_layer_types	Certain types of layers to override the quantization configurations. It is used to determine which layers are to be differentiatedly quantized. By using this parameter, you can perform differentiated quantization on some layers to change the quantization factor search step from 0.01 to 0.02. Parameter priority: override_layer_configs > override_layer_types > common_config
	Repeated	OverrideLayer	override_layer_configs	Layer to override the quantization configurations. It is used to determine which layers are to be differentiatedly quantized. By using this parameter, you can perform differentiated quantization on some layers to change the quantization factor search step from 0.01 to 0.02. Parameter priority: override_layer_configs > override_layer_types > common_config
	Optional	Boolean	do_fusion	BN fusion switch. Defaults to true, indicating BN fusion enabled.
	Repeated	String	skip_fusion_layers	Layers to skip BN fusion.
NuqConfig	-	-	-	NUQ configuration.
	Required	String	mapping_file	JSON file of the quantized model, which is obtained by converting the deployable model after uniform quantization into an offline model with ATC.
	Optional	NUQuantize	nuq_quantize	NUQ configuration.
OverrideLayerType	-	-	-	Quantization configuration to override by layer type.
	Required	String	layer_type	Quantizable layer type.
	Required	CalibrationConfig	calibration_config	Quantization configuration to override.
OverrideLayer	-	-	-	Quantization configuration to override by layer.
	Required	String	layer_name	Layers to override.
	Required	CalibrationConfig	calibration_config	Quantization configuration to override.
CalibrationConfig	-	-	-	Calibration-based quantization configuration.
	-	ARQuantize	arq_quantize	Weight quantization algorithm configuration. arq_quantize: ARQ algorithm configuration.
	-	NUQuantize	nuq_quantize	Weight quantization algorithm configuration. nuq_quantize: NUQ algorithm configuration.
	-	ADAquantize	ada_quantize	Weight quantization algorithm configuration. ada_quantize: Adaptive Rounding (AdaRound) algorithm configuration.
	-	FMRQuantize	ifmr_quantize	Activation quantization algorithm configuration. ifmr_quantize: IFMR algorithm configuration.
	-	HFMGQuantize	hfmg_quantize	Activation quantization algorithm configuration. hfmg_quantize: HFMG algorithm configuration.
	-	DMQBalancer	dmq_balancer	Balanced quantization algorithm configuration. dmq_balancer: DMQ Balancer configuration.
ARQuantize	-	-	-	ARQ algorithm configuration. For details about the algorithm, see ARQ Algorithm. This algorithm cannot be configured together with the ADAquantize and NUQuantize algorithms. If they are configured together, the quantization algorithm configured last in the configuration file is used.
	Optional	Boolean	channel_wise	Whether to use different quantization factors for each channel. true: Channels are separately quantized using different quantization factors. false: All channels are quantized altogether using the same quantization factors.
	Optional	UInt32	quant_bits	Weight quantization bit width. The value can be INT6, INT7, or INT8. INT8 quantization is used by default. This field can be set to INT6 or INT7 only for Conv2d operators. If quant_bits is set to INT6 or INT7 in common_config, the setting takes effect only for Conv2d operators. For other operators, the default value INT8 is used.
ADAquantize	-	-	-	Adaptive Rounding (AdaRound) algorithm. For details about the algorithm, see ADA Algorithm. For details about the quantization example, see AdaRound quantization calibration. This algorithm cannot be configured together with the ARQuantize and NUQuantize algorithms. If they are configured together, the quantization algorithm configured last in the configuration file is used. The quantization algorithm supports the following activation functions: ReLU, RReLU, LeakyReLU, PReLU, GELU, ReLU6, Sigmoid, and Tanh. The ONNX operators corresponding to the preceding activation functions are Relu, LeakyRelu, LeakyRelu, PRelu, Gelu, Clip, Sigmoid, and Tanh. AMCT obtains quantizable layers (torch.nn.Linear, torch.nn.Conv2d, and torch.nn.ConvTranspose2d) from the quantization configuration (for details about the restrictions, see Uniform Quantization). The quantizable modules are obtained based on the topology sequence in the exported ONNX model. If the preceding activation functions are used after the quantizable module, the quantizable module and activation function are used as a whole. Note: To export the preceding activation functions as ONNX operators, the torch.onnx.export API of torch 2.1 is required, the opset version of the Gelu operator must be v20, and the opset version of other operators must be v17.
	Optional	UInt32	num_iteration	Number of iterations. The value is greater than or equal to 0. The default value is 10000.
	Optional	Float	reg_param	Regularization parameter. The value range is (0,1). The default value is 0.01.
	Optional	Float	beta_range_start	Beta start parameter. The default value is 20. beta_range_start > beta_range_end > 0.
	Optional	Float	beta_range_end	Beta end parameter. The default value is 2.
	Optional	Float	warm_start	Warm-up factor. The value range is (0,1). The default value is 0.2.
	Optional	Boolean	channel_wise	Whether to use different quantization factors for each channel. true: Channels are separately quantized using different quantization factors. false: All channels are quantized altogether using the same quantization factors.
DMQBalancer	-	-	-	DMQ Balancer algorithm configuration. For details about the algorithm, see DMQ Balancer Algorithm.
DMQBalancer	Optional	Float	migration_strength	Migration strength, indicating the degree to which the quantization difficulty of activations is migrated to weights. The value range is [0.2, 0.8]. The default value is 0.5. Set the migration strength to a small value if there are many outliers in the activation distribution.
FMRQuantize	-	-	-	FMR algorithm for activation quantization. For details about the algorithm, see IFMR Algorithm. This algorithm cannot be configured together with the HFMGQuantize algorithm. If they are configured together, the quantization algorithm configured last in the configuration file is used.
	Optional	Float	search_range_start	Quantization factor search start.
	Optional	Float	search_range_end	Quantization factor search end.
	Optional	Float	search_step	Quantization factor search step.
	Optional	Float	max_percentile	Upper bound for searching for the largest.
	Optional	Float	min_percentile	Lower bound for searching for the smallest.
	Optional	Boolean	asymmetric	Whether to perform symmetric quantization. It is used to select the layer-wise quantization algorithm. true: asymmetric quantization false: symmetric quantization If this parameter is set for override_layer_configs, override_layer_types, and common_config, or if the activation_offset parameter is set, the priority is as follows: override_layer_configs > override_layer_types > common_config > activation_offset
	Optional	CalibrationDataType	dst_type	Quantization bit width for activation quantization, either INT8 (default) or INT16 quantization. The current version supports only INT8 quantization.
HFMGQuantize	-	-	-	HFMG algorithm for activation quantization. For details about the algorithm, see HFMG Algorithm. This algorithm cannot be configured together with the FMRQuantize algorithm. If they are configured together, the quantization algorithm configured last in the configuration file is used.
	Optional	UInt32	num_of_bins	Number of bins (the minimum unit in a histogram). Value range: {1024, 2048, 4096, 8192}. Defaults to 4096.
	Optional	Boolean	asymmetric	Whether to perform symmetric quantization. It is used to select the layer-wise quantization algorithm. true: asymmetric quantization false: symmetric quantization If this parameter is set for override_layer_configs, override_layer_types, and common_config, or if the activation_offset parameter is set, the priority is as follows: override_layer_configs > override_layer_types > common_config > activation_offset
	Optional	CalibrationDataType	dst_type	Quantization bit width for activation quantization, either INT8 (default) or INT16 quantization. The current version supports only INT8 quantization.
NUQuantize	-	-	-	NUQ configuration. For details about the algorithm, see NUQ Algorithm. This algorithm cannot be configured together with the ARQuantize and ADAquantize algorithms. If they are configured together, the quantization algorithm configured last in the configuration file is used.
	Optional	UInt32	num_steps	Number of steps for NUQ.
	Optional	UInt32	num_of_iteration	Number of iterations for NUQ optimization.

The following is an example of the simplified configuration file (quant.cfg) for uniform quantization:

# global quantize parameter
batch_num : 2
activation_offset : true
skip_layers : "Opname"
skip_layer_types:"Optype"
do_fusion: true
skip_fusion_layers : "Opname"
common_config : {
    arq_quantize : {
        channel_wise : true
        quant_bits : 7
    }
    ifmr_quantize : {
        search_range_start : 0.7
        search_range_end : 1.3
        search_step : 0.01
        max_percentile : 0.999999
        min_percentile : 0.999999
        asymmetric : true
    }
     dmq_balancer : {
         migration_strength : 0.5
    }
}
 
override_layer_types : {
    layer_type : "Conv2d"
    calibration_config : {
        arq_quantize : {
            channel_wise : false
            quant_bits : 6
        }
        ifmr_quantize : {
            search_range_start : 0.8
            search_range_end : 1.2
            search_step : 0.02
            max_percentile : 0.999999
            min_percentile : 0.999999
            asymmetric : false
        }
         dmq_balancer : {
              migration_strength : 0.5
         }
    }
}
 
override_layer_configs : {
    layer_name : "Opname"
    calibration_config : {
        arq_quantize : {
            channel_wise : true
        }
        ifmr_quantize : {
            search_range_start : 0.8
            search_range_end : 1.2
            search_step : 0.02
            max_percentile : 0.999999
            min_percentile : 0.999999
             asymmetric : false
        }
         dmq_balancer : {
              migration_strength : 0.5
         }
    }
}

If the HFMG algorithm is used for activation quantization, replace the lines in bold in the preceding configuration file with the following ones. (The following configuration file is only an example. Modify it as required.)

# global quantize parameter
activation_offset : true
batch_num : 2
...
common_config : {
    hfmg_quantize : {
        num_of_bins : 4096
         asymmetric : false
    }
...
}

The following is an example of the simplified configuration file (ada_round.cfg) for adaptive rounding:

common_config : {
    ada_quantize : {
        num_iteration : 10000
        warm_start : 0.2
        reg_param : 0.01
        beta_range_start : 20
        beta_range_end : 2
        channel_wise : false
    }
}

The following is an example of the simplified configuration file (dmq_balancer.cfg) for activation quantization balance preprocessing:

# global quantize parameter
batch_num : 2
activation_offset : true
skip_layers : "Opname"
skip_layer_types:"Optype"
do_fusion: true
skip_fusion_layers : "Opname"
common_config : {
    arq_quantize : {
        channel_wise : true
    }
    ifmr_quantize : {
        search_range_start : 0.7
        search_range_end : 1.3
        search_step : 0.01
        max_percentile : 0.999999
        min_percentile : 0.999999
        asymmetric : true
    }
     dmq_balancer : {
        migration_strength : 0.5
    }
}
 
override_layer_types : {
    layer_type : "Optype"
    calibration_config : {
        arq_quantize : {
            channel_wise : false
        }
        ifmr_quantize : {
            search_range_start : 0.8
            search_range_end : 1.2
            search_step : 0.02
            max_percentile : 0.999999
            min_percentile : 0.999999
            asymmetric : false
        }
       dmq_balancer : {
           migration_strength : 0.5
       }
    }
}
 
override_layer_configs : {
    layer_name : "Opname"
    calibration_config : {
        arq_quantize : {
            channel_wise : true
        }
        ifmr_quantize : {
            search_range_start : 0.8
            search_range_end : 1.2
            search_step : 0.02
            max_percentile : 0.999999
            min_percentile : 0.999999
            asymmetric : false
        }
       dmq_balancer : {
           migration_strength : 0.5
       }
    }
}

Parent topic: Reference