llama.cpp

97a20c01 - opencl: use `max_alloc_size` in backend ctx instead of querying again (#12705)