llama.cpp

model : avoid ggml_cont_3d for fused QKV weights

#15662

Merged