Fork of llama.cpp with fused TurboQuant flash attention — the FA kernel reads raw TBQ4_0 K/V blocks directly from global memory and dequants via centroid lookup in the FWHT-rotated domain. No separate ...