Method _quantize_int8

deepspeed/inference/quantization/utils.py:73–82

(self, tensor: Tensor)

        assert False, 'Unsupported quantization bits {}'.format(self.config['num_bits'])

    def _quantize_int8(self, tensor: Tensor) -> Tuple[Tensor, Tensor, Tensor]:
        q_range = 2**self.config['num_bits'] - 1
        min_value = tensor.amin(dim=self.config['group_dim'] + 1, keepdim=True)
        max_value = tensor.amax(dim=self.config['group_dim'] + 1, keepdim=True)

        scale = q_range / (max_value - min_value)

        tensor = tensor.sub_(min_value).mul_(scale)
        tensor = tensor_round(tensor_clamp(tensor, 0, q_range)).to(torch.uint8)
        return tensor, scale, min_value

    def _compress_uint8_to_uint4(self, tensor: Tensor) -> Tensor:
        assert tensor.shape[-1] % 2 == 0