Function PPQDyamicLinearQuantFunction

ppq/quantization/qfunction/linear.py:175–198

Dynamic Linear Quantization Function(PPQ 动态量化函数). When calling this method, we firstly solve a scale & offset setting by min-max observer. Then we applys ordinary Linear Quantization Function with solved setting. If there is a pre-defined scale & offset within given c

(tensor: torch.Tensor, config: TensorQuantizationConfig)

        return dy, None

def PPQDyamicLinearQuantFunction(tensor: torch.Tensor, config: TensorQuantizationConfig) -> torch.Tensor:
    """
    Dynamic Linear Quantization Function(PPQ 动态量化函数).
    
    When calling this method, we firstly solve a scale & offset setting by min-max observer.
    
    Then we applys ordinary Linear Quantization Function with solved setting.
    
    If there is a pre-defined scale & offset within given config, they will be dropped without warning.
    
    动态量化函数将在执行量化之前统计出 tensor 的 min - max, 而后计算出 scale & offset 并完成量化
    
    此时 TQC 中的 scale 与 offset 将被忽略
    """
    if not QuantizationStates.is_activated(config.state): return tensor
    if not config.policy.has_property(QuantizationProperty.LINEAR):
        raise ValueError('Critical Quantization Error! Non-linear config detected.')
    if not config.policy.has_property(QuantizationProperty.DYNAMIC):
        raise ValueError('Quantization Policy Do Not Have Dynamic Flag!')

    if config.policy.has_property(QuantizationProperty.PER_CHANNEL):
        return ChannelwiseDynamicLinearQuantImpl.apply(tensor, config)
    elif config.policy.has_property(QuantizationProperty.PER_TENSOR):
        return TensorwiseDynamicLinearQuantImpl.apply(tensor, config)

def PPQLinearQuantFunction(
    tensor: torch.Tensor, config: TensorQuantizationConfig) -> torch.Tensor: