feat: local ai

flake.lock

@@ -78,6 +78,22 @@
         "type": "github"
       }
     },
+    "llama-cpp-3-nixpkgs": {
+      "locked": {
+        "lastModified": 1788881743,
+        "narHash": "sha256-2V9GZGvPfrNzxFozhI9dcqV+c3QdA8YZrvAAzqEB+dI=",
+        "owner": "nixos",
+        "repo": "nixpkgs",
+        "rev": "d6524aaca2ff07876657ae2b323f24be4874944b",
+        "type": "github"
+      },
+      "original": {
+        "owner": "nixos",
+        "repo": "nixpkgs",
+        "rev": "d6524aaca2ff07876657ae2b323f24be4874944b",
+        "type": "github"
+      }
+    },
     "matugen": {
       "inputs": {
         "nixpkgs": [
@@ -158,6 +174,7 @@
       "inputs": {
         "elephant": "elephant",
         "home-manager": "home-manager",
+        "llama-cpp-3-nixpkgs": "llama-cpp-3-nixpkgs",
         "matugen": "matugen",
         "nix-flatpak": "nix-flatpak",
         "nixpkgs": "nixpkgs",

flake.nix

@@ -1,6 +1,7 @@
 {
   inputs = {
     nixpkgs.url = "github:nixos/nixpkgs/nixos-unstable";
+    llama-cpp-3-nixpkgs.url = "github:nixos/nixpkgs/d6524aaca2ff07876657ae2b323f24be4874944b";
     home-manager = {
       url = "github:nix-community/home-manager";
       inputs.nixpkgs.follows = "nixpkgs";
@@ -40,6 +41,7 @@
       nix-flatpak,
       tidalcycles,
       walker,
+      llama-cpp-3-nixpkgs,
       ...
     }@inputs:
     let
@@ -49,6 +51,9 @@
         config:
         (eachSystem (
           system:
+          let
+            llama-cpp-3-pkgs = import llama-cpp-3-nixpkgs { inherit system; };
+          in
           import nixpkgs {
             inherit system;
             config = config // {
@@ -59,6 +64,7 @@
             };
             overlays = [
               (final: prev: {
+                llama-cpp-3-rocm = llama-cpp-3-pkgs.llama-cpp-rocm;
                 cubyz-bin = prev.callPackage ./overlays/cubyz-bin { };
                 matugen = matugen.packages.${prev.system}.default;
                 gccdiag = prev.callPackage ./overlays/gccdiag { };

modules/nixos/usecases/localai.nix

@@ -16,32 +16,148 @@ in
     enable = mkEnableOption "Enable local LLM services";
   };
 
-  config = mkIf cfg.enable {
+  config =
-    home-manager.users.${username}.home.packages = [ pkgs.llama-cpp-rocm ];
+    let
-    services = {
+      llama-cpp-turboquant = pkgs.llama-cpp-3-rocm;
-      open-webui = {
+      /*
-        enable = false;
+        llama-cpp-turboquant = pkgs.llama-cpp-rocm.overrideAttrs (
-        port = 57461;
+          finalAttrs: prevAttrs: {
-        environment = {
+            version = "tqp-v0.3.0";
-          ANONYMIZED_TELEMETRY = "False";
+            src = pkgs.fetchFromGitHub {
-          DO_NOT_TRACK = "True";
+              owner = "TheTom";
-          SCARF_NO_ANALYTICS = "True";
+              repo = "llama-cpp-turboquant";
-          WEBUI_AUTH = "False";
+              tag = finalAttrs.version;
+              hash = "sha256-prdy3m7i+zhSJTzoZNk3RU+GqyExYTSK4uznSJsWLPQ=";
+            };
+            npmDepsHash = "sha256-TU4Gv+dd48WDpswhfVtm79IVIOwoCXz1fZ/DI/z40Wg=";
+          }
+        );
+      */
+    in
+    mkIf cfg.enable {
+      home-manager.users.${username}.home.packages = [
+        llama-cpp-turboquant
+        pkgs.opencode
+        pkgs.github-copilot-cli
+      ];
+      services = {
+        llama-cpp = {
+          enable = true;
+          package = llama-cpp-turboquant;
+          settings = {
+            host = "127.0.0.1";
+            port = 9931;
+            models-max = 1;
+            models-preset = pkgs.writeText "presets.ini" (
+              pkgs.lib.generators.toINI { } {
+                "Qwen3.6-35B-A3B" = {
+                  #model = "/mnt/llms/Qwen3.6-35B-A3B/Qwen3.6-35B-A3B-UD-Q6_K.gguf";
+                  model = "/mnt/llms/Qwen3.6-35B-A3B/Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf";
+                  threads = 6; # 1 per real core
+                  ctx-size = 131072; # 262144;
+                  n-cpu-moe = 34;
+                  n-gpu-layers = 999;
+                  load-mode = "mlock";
+                  #no-mmap = true;
+                  #mlock = true;
+                  jinja = true;
+                  cache-type-k = "q8_0";
+                  cache-type-v = "q8_0";
+                  #cache-type-k = "turbo4";
+                  #cache-type-v = "turbo3";
+                  #temp = 1.0;
+                  #top-p = 0.95;
+                  #top-k = 20;
+                };
+                "Qwen3.8-Flash" = {
+                  model = "/mnt/llms/Qwen3.8-Flash/Qwen3.8-Flash-Next-UD-IQ3_XXS-00001-of-00003.gguf";
+                  threads = 6; # 1 per real core
+                  n-gpu-layers = 99;
+                  n-cpu-moe = 99;
+                  load-mode = "mmap";
+                  flash-attn = "on";
+                  cache-type-k = "q8_0";
+                  cache-type-v = "q8_0";
+                  ctx-size = 65536; # 32768;
+                  batch-size = 2048;
+                  ubatch-size = 512;
+                  jinja = true;
+                };
+              }
+            );
+          };
+          /*
+            // (
+              let
+                model-dir = "/mnt/llms";
+                model = ;
+              in
+              {
+                model = "${model-dir}/${model}/${model}-UD-Q4_K_XL.gguf";
+                mmproj = "${model-dir}/${model}/mmproj-F16.gguf";
+                alias = model;
+                threads = 6; # 1 per real core
+                ctx-size = 65536; # 262144;
+                n-cpu-moe = 35;
+                n-gpu-layers = 99;
+                no-mmap = true;
+                mlock = true;
+                jinja = true;
+                #flash-attn = "on";
+                cache-type-k = "turbo4"; # "q8_0";
+                cache-type-v = "turbo3"; # "q8_0";
+                temp = 1.0;
+                top-p = 0.95;
+                top-k = 20;
+                #presence-penalty = 1.5;
+              }
+            );
+          */
+          /*
+            settings = {
+              hf-repo = "yuxinlu1/gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2-GGUF:Q4_K_M";
+              #hf-repo = "yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF:Q4_K_M";
+              # offline = true;
+              ctx-size = 32768;
+              #sleep-idle-seconds = 600;
+              n-gpu-layers = 99;
+              no-mmap = true;
+              flash-attn = "on";
+              jinja = true;
+              cache-type-k = "q8_0";
+              cache-type-v = "q8_0";
+              temp = 1.0;
+              top-p = 0.95;
+              top-k = 64;
+              repeat-penalty = 1.1;
+            };
+          */
         };
+        /*
+          open-webui = {
+            enable = true;
+            port = 57461;
+            environment = {
+              ANONYMIZED_TELEMETRY = "False";
+              DO_NOT_TRACK = "True";
+              SCARF_NO_ANALYTICS = "True";
+              WEBUI_AUTH = "False";
+            };
+          };
+        */
       };
+      /*
+        virtualisation.oci-containers.containers.open-terminal =
+        let
+          xdg = config.home-manager.users.${username}.xdg;
+        in
+        {
+          podman.user = "${username}";
+          image = "ghcr.io/open-webui/open-terminal";
+          volumes = [ "${xdg.dataHome}/open-terminal:/home/user" ];
+          ports = [ "54183:8000" ];
+          extraOptions = [ "--env-file=${xdg.configHome}/open-terminal.env" ];
+        };
+      */
     };
-    /*
-      virtualisation.oci-containers.containers.open-terminal =
-      let
-        xdg = config.home-manager.users.${username}.xdg;
-      in
-      {
-        podman.user = "${username}";
-        image = "ghcr.io/open-webui/open-terminal";
-        volumes = [ "${xdg.dataHome}/open-terminal:/home/user" ];
-        ports = [ "54183:8000" ];
-        extraOptions = [ "--env-file=${xdg.configHome}/open-terminal.env" ];
-      };
-    */
-  };
 }