I shipped Google's TurboQuant as a vLLM plugin 72 hours after the paper — here's what nobody else tested

First TurboQuant implementation validated on vision-language models. pip install, one flag, 3.76x KV cache compression.