Function wrapper

deepspeed/inference/quantization/utils.py:242–249

(input: Tensor, weight: nn.Parameter, *args, **kwargs)

    @functools.wraps(f)
    def wrapper(input: Tensor, weight: nn.Parameter, *args, **kwargs) -> Tensor:
        if hasattr(weight, 'weight_quantized') and getattr(weight, 'weight_quantized'):
            quantized_weight, quant_scale, quant_min = weight.deconcat(weight)
            temp_dequantized_weight = weight.dequantizer.dequantize(quantized_weight.view(torch.uint8), quant_scale,
                                                                    quant_min)
            return f(input, temp_dequantized_weight, *args, **kwargs)
        else:
            return f(input, weight, *args, **kwargs)

    return wrapper