llama.cpp

vulkan: use smaller combined allocations to avoid fragmentation

#11551

Merged