llama.cpp

llama : preserve original batch order for speculative decoding layer inputs

#29019

Merged