llama.cpp

server: improve speed of speculative decoding

#17808

Merged