llama.cpp

b617cfd2 - ggml-alloc : fix leak when reusing a tensor with a larger size (#16679)