lemonade: XDNA2 NPU LLM serving via nix-amd-ai

configuration.nix

@@ -18,6 +18,7 @@
     ./modules/power.nix
     ./modules/xray.nix
     ./modules/ollama.nix
+    ./modules/lemonade.nix
     ./modules/gaming.nix
     ./modules/litecoin.nix
     # ./modules/hostel-wifi.nix  # disabled 2026-05-08 — re-enable when needed

flake.lock

@@ -52,6 +52,40 @@
         "type": "github"
       }
     },
+    "flake-compat_2": {
+      "flake": false,
+      "locked": {
+        "lastModified": 1767039857,
+        "narHash": "sha256-vNpUSpF5Nuw8xvDLj2KCwwksIbjua2LZCqhV1LNRDns=",
+        "owner": "NixOS",
+        "repo": "flake-compat",
+        "rev": "5edf11c44bc78a0d334f6334cdaf7d60d732daab",
+        "type": "github"
+      },
+      "original": {
+        "owner": "NixOS",
+        "repo": "flake-compat",
+        "type": "github"
+      }
+    },
+    "flake-parts": {
+      "inputs": {
+        "nixpkgs-lib": "nixpkgs-lib"
+      },
+      "locked": {
+        "lastModified": 1775087534,
+        "narHash": "sha256-91qqW8lhL7TLwgQWijoGBbiD4t7/q75KTi8NxjVmSmA=",
+        "owner": "hercules-ci",
+        "repo": "flake-parts",
+        "rev": "3107b77cd68437b9a76194f0f7f9c55f2329ca5b",
+        "type": "github"
+      },
+      "original": {
+        "owner": "hercules-ci",
+        "repo": "flake-parts",
+        "type": "github"
+      }
+    },
     "gitignore": {
       "inputs": {
         "nixpkgs": [
@@ -98,6 +132,48 @@
         "type": "github"
       }
     },
+    "nix-amd-ai": {
+      "inputs": {
+        "flake-compat": "flake-compat_2",
+        "flake-parts": "flake-parts",
+        "nix-darwin": "nix-darwin",
+        "nixpkgs": "nixpkgs"
+      },
+      "locked": {
+        "lastModified": 1783698851,
+        "narHash": "sha256-cxQ4PCAdX/yN+WRQiR+9dhvYJlpS7kJOGuNZUafliAk=",
+        "owner": "noamsto",
+        "repo": "nix-amd-ai",
+        "rev": "a4bf7884d964b327b65ac456abf4c29fd7ac0e8b",
+        "type": "github"
+      },
+      "original": {
+        "owner": "noamsto",
+        "repo": "nix-amd-ai",
+        "type": "github"
+      }
+    },
+    "nix-darwin": {
+      "inputs": {
+        "nixpkgs": [
+          "nix-amd-ai",
+          "nixpkgs"
+        ]
+      },
+      "locked": {
+        "lastModified": 1781242433,
+        "narHash": "sha256-bchLZZ3sRn740zyvD2icZSnNoTaanN0nw7l6fjVXO+E=",
+        "owner": "nix-darwin",
+        "repo": "nix-darwin",
+        "rev": "aabb2037edfc0f210723b72cd5f528aab5dd3f0b",
+        "type": "github"
+      },
+      "original": {
+        "owner": "nix-darwin",
+        "repo": "nix-darwin",
+        "type": "github"
+      }
+    },
     "nix-flatpak": {
       "locked": {
         "lastModified": 1767983141,
@@ -115,6 +191,37 @@
       }
     },
     "nixpkgs": {
+      "locked": {
+        "lastModified": 1783224372,
+        "narHash": "sha256-8i/87eeoqiGE4yOTjwSA3Eh/ziJRQEmd/unYU+K27sk=",
+        "owner": "NixOS",
+        "repo": "nixpkgs",
+        "rev": "d407951447dcd00442e97087bf374aad70c04cea",
+        "type": "github"
+      },
+      "original": {
+        "owner": "NixOS",
+        "ref": "nixos-unstable",
+        "repo": "nixpkgs",
+        "type": "github"
+      }
+    },
+    "nixpkgs-lib": {
+      "locked": {
+        "lastModified": 1774748309,
+        "narHash": "sha256-+U7gF3qxzwD5TZuANzZPeJTZRHS29OFQgkQ2kiTJBIQ=",
+        "owner": "nix-community",
+        "repo": "nixpkgs.lib",
+        "rev": "333c4e0545a6da976206c74db8773a1645b5870a",
+        "type": "github"
+      },
+      "original": {
+        "owner": "nix-community",
+        "repo": "nixpkgs.lib",
+        "type": "github"
+      }
+    },
+    "nixpkgs_2": {
       "locked": {
         "lastModified": 1781577229,
         "narHash": "sha256-lrp67w8AulE9Ks53n27I45ADSzbOCn4H+CNW1Ck8B+8=",
@@ -157,8 +264,9 @@
       "inputs": {
         "anotherim": "anotherim",
         "lanzaboote": "lanzaboote",
+        "nix-amd-ai": "nix-amd-ai",
         "nix-flatpak": "nix-flatpak",
-        "nixpkgs": "nixpkgs",
+        "nixpkgs": "nixpkgs_2",
         "sops-nix": "sops-nix"
       }
     },

flake.nix

@@ -16,6 +16,9 @@
       inputs.nixpkgs.follows = "nixpkgs";
     };
     nix-flatpak.url = "github:gmodena/nix-flatpak/?ref=latest";
+    # No nixpkgs.follows on purpose: the flake builds against its own pinned
+    # nixpkgs so its Cachix binary cache hits (README warns follows breaks it).
+    nix-amd-ai.url = "github:noamsto/nix-amd-ai";
   };
 
   outputs = { self, nixpkgs, ... }@inputs: {

modules/lemonade.nix

@@ -0,0 +1,50 @@
+{ inputs, pkgs, ... }:
+
+{
+  imports = [ inputs.nix-amd-ai.nixosModules.default ];
+
+  # Prebuilt XRT/FLM/lemonade from the flake's Cachix; without it the whole
+  # NPU stack builds from source. Must live in nix.settings, not flake
+  # nixConfig (non-trusted users silently ignore the latter).
+  nix.settings = {
+    substituters = [ "https://nix-amd-ai.cachix.org" ];
+    trusted-public-keys = [ "nix-amd-ai.cachix.org-1:F4OU4vw/lV2oiG6SBHZ+nqjl4EFJuqI4X9A7pvaBmhQ=" ];
+  };
+
+  # XDNA2 NPU LLM serving (FLM runtime via lemonade, OpenAI-compatible API on
+  # 127.0.0.1:13305). iGPU/GGUF inference stays with ollama-rocm.
+  hardware.amd-npu = {
+    enable = true;
+    enableImageGen = false;
+    lemonade = {
+      user = "lemonade";
+      host = "127.0.0.1";
+      desktopApp.enable = false; # headless server only; skips the Tauri build path
+    };
+  };
+
+  # Same rationale as ollama: model store on plain-ext4 /mnt/data — model
+  # blobs are public data, no need to occupy LUKS root or pay dm-crypt.
+  # FLM/lemonade derive their model dirs from $HOME.
+  users.users.lemonade = {
+    isSystemUser = true;
+    group = "lemonade";
+    home = "/mnt/data/lemonade";
+    createHome = true;
+    extraGroups = [ "video" "render" ]; # /dev/accel0 is group video per module udev rule
+  };
+  users.groups.lemonade = { };
+
+  systemd.services.lemond.unitConfig.RequiresMountsFor = [ "/mnt/data/lemonade" ];
+
+  # lemonade only looks for flm on PATH when flm.prefer_system=true; default is
+  # false and nix-amd-ai's defaults seed doesn't cover it, so the FLM/NPU backend
+  # shows "not installed" and all -FLM models vanish from the registry. Patch the
+  # runtime config.json (cached copy overrides defaults) until fixed upstream.
+  systemd.services.lemond.preStart = ''
+    cfg="$HOME/.cache/lemonade/config.json"
+    if [ -f "$cfg" ]; then
+      ${pkgs.jq}/bin/jq '.flm.prefer_system = true' "$cfg" > "$cfg.tmp" && mv "$cfg.tmp" "$cfg"
+    fi
+  '';
+}