llama.cpp

90f17bba - Vulkan: Default to 1GB allocations instead of 4GB to avoid fragmentation and driver issues