llama.cpp

1b598b30 - vulkan: use smaller combined allocations to avoid fragmentation (#11551)