llama.cpp

c5650ed4 - server : avoid context swaps by shifting the KV cache