Google researchers have proposed TurboQuant, a method for compressing the key-value caches that large language models rely on ...