llama.cpp

eb425962 - llama : do not use KV cache for non-causal models