def qweight()

in optimum/quanto/nn/qmodule.py [0:0]

• 21 lines of code
• 4 McCabe index (conditional complexity)

    def qweight(self):
        """Return the module quantized weight

        When the module is frozen or does not quantize its weight parameter, it simply
        returns the weight.
        When the module is not frozen, this property is required to add the dynamic quantization
        of the weight parameter to the graph and allow gradients to be propagated to the
        underlying weight float values.
        """
        if self.weight_qtype is None:
            # QModule that does not quantize its weights
            return None
        if isinstance(self.weight, QTensor):
            # Frozen QModule
            return self.weight
        # Quantize dynamically the weights per-axis
        if isinstance(self.optimizer, SymmetricOptimizer):
            scale = self.optimizer(self.weight, qtype=self.weight_qtype, axis=0)
            shift = None
        else:
            scale, shift = self.optimizer(
                self.weight, qtype=self.weight_qtype, axis=0, group_size=self.weight_group_size
            )
        return quantize_weight(
            self.weight,
            qtype=self.weight_qtype,
            axis=0,
            scale=scale,
            shift=shift,
            group_size=self.weight_group_size,
            activation_qtype=self.activation_qtype,
        )