mindspore_gs.ptq.AWQConfig

class mindspore_gs.ptq.AWQConfig(duo_scaling=True, smooth_alpha=[i/20 for i in range(20)], weight_clip_ratio=[1-i/20 for i in range(10)])

Config for awq quant algorithm.

Parameters

• duo_scaling (bool, optional) – Use activation and weight to compute scale. Default value: True.

• smooth_alpha (List[float], optional) – The hyper-parameter of smooth search. Default value: [i/20 for i in range(20)].

• weight_clip_ratio (List[float], optional) – The hyper-parameter of clip search. Default value: [1-i/20 for i in range(10)].

Raises

• TypeError – If duo_scaling is not type bool.

• TypeError – If smooth_alpha is not type float or list.

• TypeError – If weight_clip_ratio is not float or list.

• ValueError – If smooth_alpha is less than 0 or greater than 1.

• ValueError – If weight_clip_ratio is less than 0 or greater than 1.

Examples

>>> from mindspore_gs.ptq import AWQConfig
>>> AWQConfig(duo_scaling=True, smooth_alpha=[i/20 for i in range(20)], weight_clip_ratio=[1-i/20 for i in range(10)])