POWER8 VSX + CUDA TQ3 KV cache compression (Google TurboQuant ICLR 2026). 9.1x compression, bit-exact cross-platform. - View it on GitHub