llama.cpp

1b74b9d7 - ggml : extend support for n_seq for soft_max and fattn