llama.cpp

eeae28b6 - ggml-cpu/ops: vectorize flash-attention V-cache F16 to F32 conversion (#26947)