llama.cpp

llama : set n_outputs to 1 to avoid 0 outputs mean-pooling

#15791

Merged