llama.cpp

1fb033fd - ggml : ggml_rope now takes a vector with positions instead of n_past