llama.cpp

Add self‑speculative decoding (no draft model required)

#18471

Merged