llama.cpp

finetune : keep allocs alive until all allocations are done

#4486

Merged