LLM-Labs-Local/ansible/roles/llama-cpp/tasks/main.yml

---
# llama.cpp installation and build

- name: Check if running on WSL
  ansible.builtin.command: grep -qi microsoft /proc/version
  register: wsl_check
  changed_when: false
  failed_when: false

- name: Set WSL fact
  ansible.builtin.set_fact:
    is_wsl: "{{ wsl_check.rc == 0 }}"

- name: Detect GPU on Linux/WSL
  ansible.builtin.command: nvidia-smi
  register: nvidia_smi_output
  changed_when: false
  failed_when: false
  when: ansible_os_family == "Debian" or is_wsl | default(false)

- name: Set GPU type for WSL/Linux
  ansible.builtin.set_fact:
    gpu_type: "{{ 'nvidia' if nvidia_smi_output.rc == 0 else 'none' }}"
  when: is_wsl | default(false) or ansible_os_family == "Debian"

- name: Display detected GPU type
  ansible.builtin.debug:
    msg: "llama.cpp GPU type: {{ gpu_type | default('none') }}"

- name: Check if llama.cpp already exists
  ansible.builtin.stat:
    path: "{{ llmlab_base }}/lab2/llama.cpp"
  register: llama_cpp_stat

- name: Check existing build config
  ansible.builtin.command:
    cmd: grep -q "^GGML_CUDA:BOOL=ON" "{{ llmlab_base }}/lab2/llama.cpp/build/CMakeCache.txt" 2>/dev/null && echo "cuda" || echo "none"
  register: existing_gpu_check
  changed_when: false
  failed_when: false
  when: llama_cpp_stat.stat.exists

- name: Determine if rebuild needed
  ansible.builtin.set_fact:
    needs_rebuild: >-
      {{
        not llama_cpp_stat.stat.exists or
        (gpu_type == 'nvidia' and existing_gpu_check.stdout != 'cuda') or
        (gpu_type == 'amd' and existing_gpu_check.stdout != 'amd')
      }}

- name: Clean build directory for rebuild
  ansible.builtin.file:
    path: "{{ llmlab_base }}/lab2/llama.cpp/build"
    state: absent
  become: no
  when: needs_rebuild | default(false)

- name: Clone llama.cpp repository
  ansible.builtin.git:
    repo: https://github.com/ggerganov/llama.cpp
    dest: "{{ llmlab_base }}/lab2/llama.cpp"
    version: master
    update: no
  become: no
  when: not llama_cpp_stat.stat.exists

- name: Create build directory
  ansible.builtin.file:
    path: "{{ llmlab_base }}/lab2/llama.cpp/build"
    state: directory
    mode: '0755'
  become: no

- name: Configure llama.cpp only if not already configured
  ansible.builtin.command:
    cmd: test -f CMakeCache.txt
  args:
    chdir: "{{ llmlab_base }}/lab2/llama.cpp/build"
  register: cmake_configured
  changed_when: false
  failed_when: false

- name: Configure llama.cpp with CUDA (NVIDIA GPU)
  ansible.builtin.command:
    argv:
      - cmake
      - ..
      - -G Ninja
      - -DCMAKE_BUILD_TYPE=Release
      - -DGGML_CUDA=on
  args:
    chdir: "{{ llmlab_base }}/lab2/llama.cpp/build"
  when: gpu_type == 'nvidia' and cmake_configured.rc != 0
  become: no

- name: Configure llama.cpp for AMD (ROCm)
  ansible.builtin.command:
    argv:
      - cmake
      - ..
      - -G Ninja
      - -DCMAKE_BUILD_TYPE=Release
      - -DGGML_ROCM=on
  args:
    chdir: "{{ llmlab_base }}/lab2/llama.cpp/build"
  when: gpu_type == 'amd' and cmake_configured.rc != 0
  become: no

- name: Configure llama.cpp for CPU only
  ansible.builtin.command:
    argv:
      - cmake
      - ..
      - -G Ninja
      - -DCMAKE_BUILD_TYPE=Release
  args:
    chdir: "{{ llmlab_base }}/lab2/llama.cpp/build"
  when: gpu_type | default('none') == 'none' and cmake_configured.rc != 0
  become: no

- name: Build llama.cpp
  ansible.builtin.command:
    argv:
      - ninja
  args:
    chdir: "{{ llmlab_base }}/lab2/llama.cpp/build"
  become: no
  register: build_output

- name: Display build output
  ansible.builtin.debug:
    msg: "{{ build_output.stdout_lines[-10:] }}"
  when: build_output.stdout_lines is defined

- name: Add llama.cpp to user PATH
  ansible.builtin.lineinfile:
    path: "{{ llmlab_base }}/.bashrc"
    line: 'export PATH="$HOME/lab2/llama.cpp/build/bin:$PATH"'
    state: present
    insertafter: EOF
  notify: Shell updated

- name: Display llama.cpp installation
  ansible.builtin.debug:
    msg: "llama.cpp installed to {{ llmlab_base }}/lab2/llama.cpp"