llama.cpp

bfd8ca34 - use memory margin instead of total size limit, apply to each device separately