mode

This module contains the mode descriptor for the quantization mode.

Classes

BaseCalibrateModeDescriptor

Base class for quantization calibration algorithm modes.
class BaseCalibrateModeDescriptor

Bases: ModeDescriptor

Base class for quantization calibration algorithm modes.

All calibration algorithm modes must be derived from this base class. In addition, the config_class for the mode must return a subclass of QuantizeAlgorithmConfig.

This base class also provides some convenient wrappers/utilities for calibration algorithms to be translated into ModelOpt mode.

It includes:
  1. A utility to convert the algorithm name to a mode name. This is useful since many algorithm names

    are trivial and not a good fit as a mode name. For example, "max" or None.

  2. Conversion of the algorithm and kwargs arguments of

    calibrate API to a mode config list compatible with apply_mode.

  3. Wrapper for the calibration functions in modelopt.torch.quantization.model_calib to be

    compatible with the ModelOpt convert entrypoint.

__init__(*args, **kwargs)

Initialize Base calibrate mode descriptor.

abstract property config_class: type[QuantizeAlgorithmConfig]

Specifies the config class for the mode.

property convert: Callable[[Module, ModeloptBaseConfig], tuple[Module, dict[str, Any]]] | Callable[[Module, ModeloptBaseConfig, Any], tuple[Module, dict[str, Any]]]

The calibrate algorithm mode’s entrypoint for converting a model.

This method is called by the ModelOpt framework when applying this calibration mode to a model. See wrapped_calib_func() for more details on the logic.

Note: Subclasses must specify the _calib_func class attribute with the appropriate calibration function to be used or override this method.

property name: str

Returns the value (str representation) of the mode.

property restore: Callable[[Module, ModeloptBaseConfig, dict[str, Any]], Module]

The mode’s entrypoint for restoring a model.

property update_for_new_mode: Callable[[Module, ModeloptBaseConfig, dict[str, Any]], None]

The mode’s entrypoint for updating the models state before new mode.

property update_for_save: Callable[[Module, ModeloptBaseConfig, dict[str, Any]], None]

The mode’s entrypoint for updating the models state before saving.