llama.cpp

fd123cfe - Vulkan: Default to 1GB allocations instead of 4GB to avoid fragmentation and driver issues (#12434)