Quantized models are slower than float models on GPUs

[sakura]

thanks, I will do it !