TurboQuant KV Cache Compression for llama.cpp — 5.2x memory reduction with near-lossless quality | Implementation of Google DeepMind's TurboQuant (ICLR 2026) - View it on GitHub