llama.cpp

llama: use FA + max. GPU layers by default

#15434

Merged