llama.cpp

llama-bench : use random tokens to improve accuracy with mixtral

#6069

Merged