diff --git a/services/llama-cpp.nix b/services/llama-cpp.nix index 76918de..7a4f568 100644 --- a/services/llama-cpp.nix +++ b/services/llama-cpp.nix @@ -34,14 +34,20 @@ in port = service_configs.ports.llama_cpp; host = "0.0.0.0"; package = ( - optimizeWithFlags inputs.llamacpp.packages.${pkgs.system}.default [ - "-O3" - "-march=znver2" - "-mtune=znver2" - ] + optimizeWithFlags + (inputs.llamacpp.packages.${pkgs.system}.default.overrideAttrs (old: { + cmakeFlags = old.cmakeFlags ++ [ + "-DGGML_AVX2=ON" + ]; + })) + [ + "-O3" + "-march=znver2" + "-mtune=znver2" + ] ); extraFlags = [ - + "--flash-attn" ]; };