{
  pkgs,
  service_configs,
  config,
  inputs,
  lib,
  ...
}:
{
  services.llama-cpp = {
    enable = true;
    model = builtins.toString (
      pkgs.fetchurl {
        url = "https://huggingface.co/ibm-granite/granite-4.0-h-tiny-GGUF/resolve/main/granite-4.0-h-tiny-Q4_0.gguf";
        sha256 = "7ced406c5ee7ae653b772d56aa79ca32997d9e3da69d97b2c8a66dd5e7d5f77d";
      }
    );
    port = service_configs.ports.llama_cpp;
    host = "0.0.0.0";
    # vulkan broken: https://github.com/ggml-org/llama.cpp/issues/13801
    package = (
      lib.optimizePackage (
        inputs.llamacpp.packages.${pkgs.system}.default.overrideAttrs (old: {
          postPatch = "";
        })
      )
    );
    extraFlags = [
      # "-ngl"
      # "9999"
    ];
  };

  # have to do this in order to get vulkan to work
  systemd.services.llama-cpp.serviceConfig.DynamicUser = lib.mkForce false;

  services.caddy.virtualHosts."llm.${service_configs.https.domain}".extraConfig = ''
    ${builtins.readFile ../secrets/caddy_auth}
    reverse_proxy :${builtins.toString config.services.llama-cpp.port}
  '';
}