- name: Setup llama.cpp - Vulkan
  hosts: llamacpp
  vars:
    podman_user: "{{ llama_user | default('llama') }}"
    podman_group: "{{ llama_group | default('llama') }}"
    podman_extra_groups: "users,video"
    podman_user_home: "{{ llama_home | default('/home/llama') }}"
    __llama_models__: "{{ llama_models | default([]) }}"
    __llama_router_dir__: "{{ podman_user_home }}/models/.router"
    __llama_managed_dir__: "{{ podman_user_home }}/models/managed"
    __llama_links_dir__: "{{ podman_user_home }}/models/managed-links"
    __llama_manifest_file__: "{{ __llama_router_dir__ }}/manifest.json"
    __llama_models_input_file__: "{{ __llama_router_dir__ }}/desired-models.json"
    __llama_preset_global__: "{{ llama_preset_global | default({}) }}"
    __llama_preset_global_input_file__: "{{ __llama_router_dir__ }}/preset-global.json"
    __llama_preset_file__: "{{ __llama_router_dir__ }}/models.ini"
    __llama_sync_script__: "{{ __llama_router_dir__ }}/llama_hf_sync.py"
    __llama_sync_venv__: "{{ llama_sync_venv | default(__llama_router_dir__ ~ '/.venv') }}"
    __llama_sync_python__: "{{ __llama_sync_venv__ }}/bin/python"
    __llama_models_max__: "{{ llama_models_max_loaded | default(1) }}"
    podman_user_folders:
      - models
      - .cache
      - .cache/llama.cpp
    pods:
      - name: llama.cpp
        repo:
          image: ghcr.io/ggml-org/llama.cpp:server-vulkan
        ports:
          - "0.0.0.0:{{ llama_port | default(8012) }}:8080/tcp"
        command:
          - "-t"
          - "{{ llama_cmd_threads | default('12') }}"
          - "-np"
          - "{{ llama_cmd_parallel | default('1') }}"
          - "-b"
          - "{{ llama_cmd_batch_size | default('1024') }}"
          - "-ub"
          - "{{ llama_cmd_ubatch_size | default('512') }}"
          - "-fa"
          - "{{ llama_cmd_flash_attn | default('on') }}"
          - "-ctk"
          - "{{ llama_cmd_cache_type_k | default('q4_0') }}"
          - "-ctv"
          - "{{ llama_cmd_cache_type_v | default('q4_0') }}"
          - "-cram"
          - "{{ llama_cmd_cache_reuse | default('-1') }}"
          - "-sm"
          - "{{ llama_cmd_split_mode | default('layer') }}"
          - "-dev"
          - "{{ llama_cmd_devices | default('Vulkan1,Vulkan2') }}"
          - "-ts"
          - "{{ llama_cmd_tensor_split | default('8,12') }}"
          - "-fit"
          - "{{ llama_cmd_fit | default('off') }}"
          - "-mg"
          - "{{ llama_cmd_main_gpu | default('1') }}"
          - "--mmap"
          - "--models-preset"
          - "{{ llama_cmd_models_preset_path | default('/models/.router/models.ini') }}"
          - "--models-max"
          - "{{ llama_cmd_models_max | default(__llama_models_max__) }}"
        volumes:
          - "{{ podman_user_home }}/models:/models:Z"
          - "{{ podman_user_home }}/.cache:/app/.cache:Z"
        device:
          - "nvidia.com/gpu=all"
          - "/dev/kfd:/dev/kfd:rw"
          - "/dev/dri:/dev/dri:rw"
  tasks:
    - name: Sync llama models from Hugging Face
      ansible.builtin.include_tasks:
        file: tasks/llama_models_sync.yml
      when: llama_models_sync_enabled | default(true)

    - name: Setup Pods
      ansible.builtin.include_tasks:
        file: tasks/podman.yml

    - name: Query llama router model list
      ansible.builtin.uri:
        url: "http://127.0.0.1:{{ llama_port | default(8012) }}/models?reload=1"
        method: GET
        status_code: 200
      register: __llama_router_models__
      changed_when: false
      when:
        - llama_models_sync_enabled | default(true)
        - not (llama_models_dry_run | default(false))

    - name: Validate expected model aliases are available in router
      ansible.builtin.assert:
        that:
          - item.name in (__llama_router_models__.json.data | map(attribute='id') | list)
        fail_msg: "Model name '{{ item.name }}' not present in llama router /models output"
      loop: "{{ __llama_sync_result__.models | default([]) }}"
      loop_control:
        label: "{{ item.name }}"
      when:
        - llama_models_sync_enabled | default(true)
        - not (llama_models_dry_run | default(false))