llama.cpp

bef122e1 - speculative: use slot.cache_tokens.insert