llama.cpp

a37696d4 - speculative : more robust tokenizer comparison