llama.cpp

0467fede - also strip models memory margin from child processes