llama.cpp

c1596f63 - llama : fix kv cache heuristic when context is less than 32