llama.cpp

b0bc9f4a - llama-bench : use random tokens to improve accuracy with mixtral (#6069)