A Speed Odyssey for Deployable Quantization of LLMs