feat: initial commit — Ansible roles storage-01/gpu-01 opérationnels

ansible/roles/llama_server/defaults/main.yml

@@ -0,0 +1,15 @@
+---
+llama_server_port: 1234
+llama_server_host: "0.0.0.0"
+llama_amdgpu_targets: "gfx1030"
+rocm_path: "/opt/rocm"
+hsa_override_gfx_version: "10.3.0"
+
+# Modèle à charger au démarrage
+llama_model_path: ""
+llama_model_alias: ""
+llama_ctx_size: 32768
+llama_n_gpu_layers: 99
+llama_parallel: 1
+llama_embeddings: true
+llama_pooling: "mean"

ansible/roles/llama_server/handlers/main.yml

@@ -0,0 +1,9 @@
+---
+- name: Reload systemd
+  ansible.builtin.systemd:
+    daemon_reload: true
+
+- name: Restart llama-server
+  ansible.builtin.systemd:
+    name: llama-server
+    state: restarted

ansible/roles/llama_server/tasks/main.yml

@@ -0,0 +1,79 @@
+---
+- name: Install build dependencies
+  ansible.builtin.dnf:
+    name:
+      - cmake
+      - gcc-c++
+      - git
+      - patchelf
+      - hipblas-devel
+      - rocblas-devel
+      - hip-devel
+      - hipcc
+    state: present
+
+- name: Clone llama.cpp
+  ansible.builtin.git:
+    repo: https://github.com/ggerganov/llama.cpp
+    dest: /opt/llama.cpp
+    depth: 1
+    version: "{{ llama_server_commit | default('HEAD') }}"
+  register: llama_clone
+
+- name: Check if llama-server binary exists
+  ansible.builtin.stat:
+    path: /opt/llama.cpp/build/bin/llama-server
+  register: llama_server_binary_stat
+
+- name: Configure cmake build (ROCm HIP)
+  ansible.builtin.command:
+    cmd: >
+      cmake -B build
+      -DGGML_HIP=ON
+      -DAMDGPU_TARGETS={{ llama_amdgpu_targets }}
+      -DCMAKE_BUILD_TYPE=Release
+      -DROCM_PATH={{ rocm_path }}
+      -DCMAKE_PREFIX_PATH={{ rocm_path }}
+      -DCMAKE_HIP_COMPILER={{ rocm_path }}/llvm/bin/clang++
+    chdir: /opt/llama.cpp
+  environment:
+    PATH: "{{ rocm_path }}/bin:{{ ansible_env.PATH }}"
+  when: llama_clone is changed or not llama_server_binary_stat.stat.exists
+
+- name: Build llama-server
+  ansible.builtin.command:
+    cmd: cmake --build build --target llama-server -j{{ ansible_processor_nproc }}
+    chdir: /opt/llama.cpp
+  environment:
+    PATH: "{{ rocm_path }}/bin:{{ ansible_env.PATH }}"
+  when: llama_clone is changed or not llama_server_binary_stat.stat.exists
+
+- name: Deploy systemd service
+  ansible.builtin.template:
+    src: llama-server.service.j2
+    dest: /etc/systemd/system/llama-server.service
+    mode: '0644'
+  notify:
+    - Reload systemd
+    - Restart llama-server
+
+- name: Enable llama-server service
+  ansible.builtin.systemd:
+    name: llama-server
+    enabled: true
+    daemon_reload: true
+    state: started
+
+- name: Open port in firewall
+  ansible.posix.firewalld:
+    port: "{{ llama_server_port }}/tcp"
+    permanent: true
+    state: enabled
+    immediate: true
+
+- name: Disable lm-studio service (replaced by llama-server)
+  ansible.builtin.systemd:
+    name: lm-studio
+    enabled: false
+    state: stopped
+  ignore_errors: true

ansible/roles/llama_server/templates/llama-server.service.j2

@@ -0,0 +1,28 @@
+[Unit]
+Description=llama-server (llama.cpp OpenAI-compatible API)
+After=network-online.target
+Wants=network-online.target
+
+[Service]
+Type=simple
+User=root
+Environment=HSA_OVERRIDE_GFX_VERSION={{ hsa_override_gfx_version }}
+Environment=LD_LIBRARY_PATH={{ rocm_path }}/lib
+ExecStart=/opt/llama.cpp/build/bin/llama-server \
+    --model {{ llama_model_path }} \
+    --host {{ llama_server_host }} \
+    --port {{ llama_server_port }} \
+    --ctx-size {{ llama_ctx_size }} \
+    --n-gpu-layers {{ llama_n_gpu_layers }} \
+    --parallel {{ llama_parallel }} \
+    --alias {{ llama_model_alias }} \
+{% if llama_embeddings %}
+    --embeddings \
+    --pooling {{ llama_pooling }} \
+{% endif %}
+    --log-disable
+Restart=on-failure
+RestartSec=10
+
+[Install]
+WantedBy=multi-user.target