flake.nix

@@ -67,6 +67,7 @@
     lib = nixpkgs.lib;
   in {
     nixosConfigurations = lib.genAttrs [
+      "dixie"
       "imladris"
       "ono-sendai"
       "orion"

hw/dixie/authorized_keys.pub

@@ -0,0 +1,4 @@
+ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAIDXu7ME91tLfU+Q3fNFe8VyIouVvoktoYy8xVNpObxuv kyle@titan.local
+ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAIGea83yMIdCi0QUUPgmhRgIrii7lS1dYxZ6LSxSsDOph kyle@europa
+ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAIHukZAxnlMSoTlRtRF14mLbC0x+rb1zlxnSLVLg+6Xbh kyle@vade
+ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAILw318jrtFJw1Cn3WnOk17Gkr8bo7NOACQwMQ7S4hvoL kyle@straylight

hw/dixie/default.nix

@@ -0,0 +1,186 @@
+{ inputs, pkgs, lib, config, ... }:
+let
+  # Unstable instantiated with unfree allowed: CUDA is unfree, and legacyPackages does not
+  # inherit the host's nixpkgs.config (straylight's Vulkan build never hit this).
+  pkgsUnstable = import inputs.nixpkgs-unstable {
+    system = pkgs.stdenv.hostPlatform.system;
+    config = { allowUnfree = true; cudaSupport = true; };
+  };
+  # llama.cpp with the CUDA backend, from unstable to match straylight's build (same model
+  # files, same router semantics). Built once on straylight and shipped with dixie-push
+  # --closure, or pulled from the cache when Dixie's link allows.
+  llama-cpp = pkgsUnstable.llama-cpp.override { cudaSupport = true; };
+  modelsDir = "/var/lib/llama-server/models";
+
+  # Dixie is the fleet's helper tier: everything that is latency-sensitive, small, and was
+  # starving on straylight's CPU (2026-09-21). One RTX 3060 (12 GB) today; card 2 joins
+  # when its fans are replaced.
+  #
+  # Router: the three models that must never be swapped out — approval guardian / Honcho
+  # deriver (9B), Honcho dialectic (4B, non-thinking), Honcho embeddings — resident together:
+  # 5.6 + 2.5 + 0.7 GB weights, ~3 GB left for KV. Vision does NOT go here: 5 GB + 1.2 GB
+  # mmproj does not fit beside them, so it gets its own on-demand process below.
+  llamaModelsIni = pkgs.writeText "llama-models.ini" ''
+    version = 1
+    [*]
+    jinja = true
+    flash-attn = on
+    cache-type-k = q8_0
+    cache-type-v = q8_0
+    n-gpu-layers = 999
+    ; three resident models share 12 GB: keep contexts modest
+    ctx-size = 16384
+    parallel = 2
+    cache-ram = 512
+
+    ; Ornith 1.5 9B, refusal-ablated. Hermes auxiliary model (approval guardian, title,
+    ; web_extract) for all agents and Honcho's deriver/summary/dream model.
+    [ornith-1.5-9b-uncensored]
+    model = ${modelsDir}/ornith-1.5-9b-uncensored/Ornith-1.5-9B-uncensored.Q4_K_M.gguf
+    ; Runaway guard carried over from straylight: helper tasks need short answers.
+    reasoning-budget = 2048
+    n-predict = 4096
+
+    ; Honcho dialectic: non-thinking instruct model so answers are short and fast.
+    [honcho-dialectic]
+    model = ${modelsDir}/qwen3-4b-instruct-2507/Qwen3-4B-Instruct-2507-Q4_K_M.gguf
+    n-predict = 1024
+
+    ; Honcho embeddings (1024-dim; EMBEDDING_VECTOR_DIMENSIONS=1024 on rift).
+    [honcho-embed]
+    model = ${modelsDir}/qwen3-embedding-0.6b/Qwen3-Embedding-0.6B-Q8_0.gguf
+    embedding = true
+    pooling = cls
+    ubatch-size = 8192
+    ctx-size = 8192
+    parallel = 4
+  '';
+
+  # Vision (auxiliary.vision for all agents): its own server on :11433, loaded on first
+  # request and dropped again after an idle hour so the router's three keep their VRAM.
+  visionModelsIni = pkgs.writeText "llama-vision.ini" ''
+    version = 1
+    [*]
+    jinja = true
+    flash-attn = on
+    n-gpu-layers = 999
+    ctx-size = 16384
+    parallel = 1
+    sleep-idle-seconds = 3600
+
+    [qwen3-vl-8b-abliterated]
+    model = ${modelsDir}/qwen3-vl-8b-abliterated/Qwen3-VL-8B-Instruct-abliterated-v2.0.Q4_K_M.gguf
+    mmproj = ${modelsDir}/qwen3-vl-8b-abliterated/Qwen3-VL-8B-Instruct-abliterated-v2.0.mmproj-f16.gguf
+  '';
+
+  llamaService = { description, port, preset, modelsMax }: {
+    inherit description;
+    wantedBy = [ "multi-user.target" ];
+    after = [ "network-online.target" "var-lib-llama\\x2dserver-models.mount" ];
+    wants = [ "network-online.target" ];
+    environment.HOME = "/var/lib/llama-server";
+    serviceConfig = {
+      ExecStart = ''
+        ${llama-cpp}/bin/llama-server \
+          --host 0.0.0.0 \
+          --port ${toString port} \
+          --models-preset ${preset} \
+          --models-max ${toString modelsMax} \
+          --timeout 7200
+      '';
+      User = "llama-server";
+      Group = "llama-server";
+      StateDirectory = "llama-server";
+      WorkingDirectory = "/var/lib/llama-server";
+      SupplementaryGroups = [ "video" ];
+      Restart = "on-failure";
+      RestartSec = "10s";
+      TimeoutStartSec = "600";
+    };
+  };
+in
+{
+  imports = [
+    ./hardware-configuration.nix
+    ./disk-config.nix
+  ];
+
+  config = {
+    # Headless. The base configuration.nix has no desktop; nothing GUI is added here.
+    boot.loader.systemd-boot.enable = true;
+    boot.loader.efi.canTouchEfiVariables = true;
+    # 6.12 LTS: mt7921u (the AX9L USB Wi-Fi) is in-tree from 5.18; avoids the 6.17/6.18
+    # mt7921 monitor-mode regression for good measure.
+    boot.kernelPackages = pkgs.linuxPackages_6_12;
+
+    # NVIDIA, compute only: no X, no modesetting needed for CUDA.
+    hardware.graphics.enable = true;
+    services.xserver.videoDrivers = [ "nvidia" ];
+    hardware.nvidia = {
+      open = false;
+      modesetting.enable = false;
+      nvidiaPersistenced = true;
+      package = config.boot.kernelPackages.nvidiaPackages.production;
+    };
+    hardware.enableRedistributableFirmware = true;
+
+    # Power cap for the 3060(s): thermal hour on 2026-09-21 held 75 °C at 140 W, case closed.
+    systemd.services.nvidia-power-cap = {
+      description = "Cap GPU power (thermals in the Aurora chassis)";
+      wantedBy = [ "multi-user.target" ];
+      after = [ "nvidia-persistenced.service" ];
+      serviceConfig = {
+        Type = "oneshot";
+        RemainAfterExit = true;
+        ExecStart = "${config.hardware.nvidia.package.bin}/bin/nvidia-smi -pl 140";
+      };
+    };
+
+    # Headless and Wi-Fi only: there is no console to add a key from after install, so the
+    # keys go in with the system. Same set the live ISO used (~/src/dixie-iso).
+    users.users.kyle.openssh.authorizedKeys.keyFiles = [ ./authorized_keys.pub ];
+    users.users.root.openssh.authorizedKeys.keyFiles = [ ./authorized_keys.pub ];
+
+    users.users.llama-server = { isSystemUser = true; group = "llama-server"; home = "/var/lib/llama-server"; };
+    users.groups.llama-server = { };
+
+    # Model weights live on the USB cache stick (ext4, label dixie-cache), populated from
+    # straylight with dixie-push. Both servers wait for the mount.
+    fileSystems."/mnt/cache" = {
+      device = "/dev/disk/by-label/dixie-cache";
+      fsType = "ext4";
+      options = [ "noatime" "nofail" "x-systemd.device-timeout=30s" ];
+    };
+    fileSystems."/var/lib/llama-server/models" = {
+      device = "/mnt/cache/models";
+      fsType = "none";
+      options = [ "bind" "nofail" "x-systemd.requires=/mnt/cache" ];
+    };
+
+    systemd.services.llama-server = llamaService {
+      description = "llama.cpp router: helper tier (9B guardian/deriver, dialectic, embeddings)";
+      port = 11434; preset = llamaModelsIni; modelsMax = 3;
+    };
+    systemd.services.llama-vision = llamaService {
+      description = "llama.cpp vision server (qwen3-vl-8b, on demand)";
+      port = 11433; preset = visionModelsIni; modelsMax = 1;
+    };
+
+    # Wi-Fi only host. NetworkManager comes from the base config; the AX9L (mt7921u) becomes
+    # the primary once plugged in and the internal CNVi radio stays as fallback. Connection
+    # profiles are copied from the live-USB session at install time (see the install notes),
+    # not stored in this repo.
+    networking.networkmanager.wifi.backend = "wpa_supplicant";
+
+    # Reachable only over the tailnet: 11434 (router) and 11433 (vision) on tailscale0, and
+    # SSH from the LAN (base config opens 22) for bring-up.
+    networking.firewall.interfaces.tailscale0.allowedTCPPorts = [ 11433 11434 ];
+
+    environment.systemPackages = with pkgs; [
+      llama-cpp
+      pciutils usbutils lm_sensors dmidecode nvme-cli smartmontools ethtool iw wpa_supplicant
+      htop tmux git jq curl rsync nvtopPackages.nvidia
+    ];
+
+  };
+}

hw/dixie/disk-config.nix

@@ -0,0 +1,38 @@
+## (dixie) disko config
+## Alienware Aurora R11: Samsung PM9A1 1 TB NVMe is the system disk. The 6 TB SATA HDD and the
+## USB cache stick are left alone here (the stick is mounted by label in default.nix).
+## Unencrypted on purpose: headless box that must come back from a power cut with nobody
+## at the console to unlock it.
+
+{ lib, ... }:
+
+{
+  disko.devices.disk.main = {
+    type = "disk";
+    device = "/dev/disk/by-id/nvme-eui.36355630521205220025384500000001";
+    content = {
+      type = "gpt";
+      partitions = {
+        ESP = {
+          size = "1G";
+          type = "EF00";
+          content = {
+            type = "filesystem";
+            format = "vfat";
+            mountpoint = "/boot";
+            mountOptions = [ "umask=0077" ];
+          };
+        };
+        root = {
+          size = "100%";
+          content = {
+            type = "filesystem";
+            format = "ext4";
+            mountpoint = "/";
+            mountOptions = [ "noatime" ];
+          };
+        };
+      };
+    };
+  };
+}

hw/dixie/hardware-configuration.nix

@@ -0,0 +1,18 @@
+# Do not modify this file!  It was generated by ‘nixos-generate-config’
+# and may be overwritten by future invocations.  Please make changes
+# to /etc/nixos/configuration.nix instead.
+{ config, lib, pkgs, modulesPath, ... }:
+
+{
+  imports =
+    [ (modulesPath + "/installer/scan/not-detected.nix")
+    ];
+
+  boot.initrd.availableKernelModules = [ "xhci_pci" "ahci" "nvme" "usbhid" "usb_storage" "sd_mod" ];
+  boot.initrd.kernelModules = [ ];
+  boot.kernelModules = [ "kvm-intel" ];
+  boot.extraModulePackages = [ ];
+
+  nixpkgs.hostPlatform = lib.mkDefault "x86_64-linux";
+  hardware.cpu.intel.updateMicrocode = lib.mkDefault config.hardware.enableRedistributableFirmware;
+}