A from-scratch PyTorch implementation of TurboQuant (ICLR 2026), Google's two-stage vector quantization algorithm for compressing LLM key-value caches — enhanced ...