llama.cpp

kv-cache : support attention rotation for heterogeneous iSWA

#21513

Merged