llama.cpp

51f0bd50 - Remove custom pre attention scaling and use computed value instead.