llama.cpp

c4df49a4 - kleidiai: generalize compute_forward_kv_cache to compute_forward_fp16 (#15817)