Google researchers have published a new quantization technique called TurboQuant that compresses the key-value (KV) cache in large language models to 3.5 bits per channel, cutting memory consumption ...