llama.cpp

llama : fix FA when KV cache is not used (i.e. embeddings)

#12825

Merged