llama.cpp

1bded5a3 - kv-cache : better estimate of n_kv for multi-sequence batches (#15610)