Function PPQLinearQuant_toInt

ppq/quantization/qfunction/linear.py:218–238

PPQ 核心量化函数，没啥好说的了吧，这个玩意只做 quant 不做 dequant

(tensor: torch.Tensor, config: TensorQuantizationConfig)

            config.quant_min, config.quant_max, config.rounding)

def PPQLinearQuant_toInt(tensor: torch.Tensor, config: TensorQuantizationConfig) -> torch.Tensor:
    """PPQ 核心量化函数，没啥好说的了吧，这个玩意只做 quant 不做 dequant"""
    if not config.policy.has_property(QuantizationProperty.LINEAR):
        raise ValueError('Critical Quantization Error! Non-linear config detected.')
    if config.policy.has_property(QuantizationProperty.PER_CHANNEL):
        shape = [1 if axis != config.channel_axis else -1 for axis in range(tensor.ndim)]
        scale, offset = config.scale.view(shape), config.offset.view(shape)
        tensor = ppq_tensor_round((tensor / scale), config.rounding) + offset
        tensor = torch.clamp(tensor, config.quant_min, config.quant_max)
    elif config.policy.has_property(QuantizationProperty.PER_TENSOR):
        tensor = ppq_tensor_round((tensor / config.scale), config.rounding) + config.offset
        tensor = torch.clamp(tensor, config.quant_min, config.quant_max)

    if config.num_of_bits == 8:
        if config.policy.has_property(QuantizationProperty.SYMMETRICAL):
            return tensor.type(dtype=torch.int8)
        if config.policy.has_property(QuantizationProperty.ASYMMETRICAL):
            return tensor.type(dtype=torch.uint8)
    elif config.num_of_bits > 8:
        return tensor.type(dtype=torch.int32)
    else: raise Exception('Do not konw how to convert value into int. num of bits is unexpected.')