ConvertCustomConfig

class torch.ao.quantization.fx.custom_config.ConvertCustomConfig

Custom configuration for convert_fx().

Example usage:

convert_custom_config = ConvertCustomConfig()             .set_observed_to_quantized_mapping(ObservedCustomModule, QuantizedCustomModule)             .set_preserved_attributes(["attr1", "attr2"])

classmethod from_dict(convert_custom_config_dict)

Create a ConvertCustomConfig from a dictionary with the following items:

“observed_to_quantized_custom_module_class”: a nested dictionary mapping from quantization mode to an inner mapping from observed module classes to quantized module classes, e.g.:: { “static”: {FloatCustomModule: ObservedCustomModule}, “dynamic”: {FloatCustomModule: ObservedCustomModule}, “weight_only”: {FloatCustomModule: ObservedCustomModule} } “preserved_attributes”: a list of attributes that persist even if they are not used in forward

This function is primarily for backward compatibility and may be removed in the future.

Return type:

ConvertCustomConfig

set_observed_to_quantized_mapping(observed_class, quantized_class, quant_type=QuantType.STATIC)

Set the mapping from a custom observed module class to a custom quantized module class.

The quantized module class must have a from_observed class method that converts the observed module class to the quantized module class.

Return type:

ConvertCustomConfig

set_preserved_attributes(attributes)

Set the names of the attributes that will persist in the graph module even if they are not used in the model’s forward method.

Return type:

ConvertCustomConfig

to_dict()

Convert this ConvertCustomConfig to a dictionary with the items described in from_dict().

Return type:

dict[str, Any]