llama.cpp

ggml-cuda : use graph allocator

#2684

Merged