llama.cpp

kv-cache : better estimate of n_kv for multi-sequence batches

#15610

Merged