llama: use f16 mask for FA to save VRAM by am17an · Pull Request #23764 · ggml-org/llama.cpp(reddit.com)