llama.cpp

spec: enable backend sampling for both dflash & dspark

#26958

Merged