llama.cpp

fe00a84b - tests: enable kv_unified to prevent cuda oom error on rtx 2060 (#20645)