llama.cpp

026bb1b1 - batched-bench : add readme + n_kv_max is now configurable