llama.cpp

f855d523 - Fix import of llama2.c models (only 42M model shares weights between wcls and token_embedding_table)