llama.cpp

3470a5c8 - ggml-alloc : make gallocr prefer chunks that allow memory reuse (#16788)