update

flake.lock

@@ -200,11 +200,11 @@
         ]
       },
       "locked": {
-        "lastModified": 1743424621,
+        "lastModified": 1743439256,
-        "narHash": "sha256-M4dHt10aGASKjoRtafFMJfewANh/7+O6t+ITb9oPsNY=",
+        "narHash": "sha256-eVxyPbpJuQV3qShmoerj6xzjUsB/UmNGbqP0cNBv42g=",
         "owner": "ggml-org",
         "repo": "llama.cpp",
-        "rev": "a8a1f3356786cbf8bcc3422e3c8737fc33b453e7",
+        "rev": "c80a7759dab10657b9b6c3e87eef988a133b9b6a",
         "type": "github"
       },
       "original": {
@@ -253,11 +253,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1743398199,
+        "lastModified": 1743445058,
-        "narHash": "sha256-Zy9o4AiBVjmswfXtw5l0YTUSEp676YKAqC9Z2d6MvI0=",
+        "narHash": "sha256-jayDP/dth0BdbFGM2C0Ny7uRHwwA0ppbB+2s23QG9Vk=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "e9549075069ef5e68ddf29cb07c0e38e2a915242",
+        "rev": "e4a74fd10b211a6d8f5575a85154aac0209a8e55",
         "type": "github"
       },
       "original": {

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.${pkgs.system}.default [
+      optimizeWithFlags
+        (inputs.llamacpp.packages.${pkgs.system}.default.overrideAttrs (old: {
+          cmakeFlags = old.cmakeFlags ++ [
+            "-DGGML_AVX2=ON"
+          ];
+        }))
+        [
           "-O3"
           "-march=znver2"
           "-mtune=znver2"
         ]
     );
     extraFlags = [
-
+      "--flash-attn"
     ];
   };