llama.cpp

a9b5fe98 - fix: Fix logic for initializing inputs and attn layers for hybrid caches