llama.cpp

llama-bench : use two tokens in the warmup run for prompt evals

#3059

Merged