mirror of
https://github.com/Alkatrazz24/Funk-lab.git
synced 2026-07-12 00:14:41 +02:00
4da1589a6f
Remplace Qwen3 (chat réutilisé, dim 4096, peu discriminant) par un modèle d'embedding spécialisé nomic-embed-text (dim 768) pour la mémoire long-terme. Ansible (rôle llama_server) : - nouvelle instance optionnelle `llama-embed` (llama_embed_enabled) servant un modèle d'embedding dédié sur :1238, GPU ; télécharge le GGUF si absent. - activée sur gpu-01 (host_vars) : nomic-embed-text-v1.5 f16. STT-server : - STT_EMBED_URL → :1238, STT_EMBED_MODEL → nomic-embed-text (deployment + config). - _ensure_collection détecte le changement de dimension (4096→768) et recrée automatiquement la collection stt-memory (anciens vecteurs incomparables) — pas de drop manuel. Docs : llama_server README, rag.md, stt.md (5d ✅), CLAUDE.md. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013FmcxGsyXZXogiAHQLjnZT
125 lines
3.3 KiB
YAML
125 lines
3.3 KiB
YAML
---
|
|
- 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
|
|
|
|
# --- Instance dédiée embeddings (nomic-embed-text) — optionnelle ---------------
|
|
|
|
- name: Ensure embedding model directory exists
|
|
ansible.builtin.file:
|
|
path: "{{ llama_embed_model_path | dirname }}"
|
|
state: directory
|
|
mode: '0755'
|
|
when: llama_embed_enabled and llama_embed_model_path | length > 0
|
|
|
|
- name: Download embedding model (GGUF) if absent
|
|
ansible.builtin.get_url:
|
|
url: "{{ llama_embed_model_url }}"
|
|
dest: "{{ llama_embed_model_path }}"
|
|
mode: '0644'
|
|
timeout: 120
|
|
when:
|
|
- llama_embed_enabled
|
|
- llama_embed_model_url | length > 0
|
|
- llama_embed_model_path | length > 0
|
|
|
|
- name: Deploy embedding llama-server service
|
|
ansible.builtin.template:
|
|
src: llama-embed.service.j2
|
|
dest: /etc/systemd/system/llama-embed.service
|
|
mode: '0644'
|
|
notify:
|
|
- Reload systemd
|
|
- Restart llama-embed
|
|
when: llama_embed_enabled
|
|
|
|
- name: Enable embedding llama-server service
|
|
ansible.builtin.systemd:
|
|
name: llama-embed
|
|
enabled: true
|
|
daemon_reload: true
|
|
state: started
|
|
when: llama_embed_enabled
|
|
|
|
- name: Open embedding port in firewall
|
|
ansible.posix.firewalld:
|
|
port: "{{ llama_embed_port }}/tcp"
|
|
permanent: true
|
|
state: enabled
|
|
immediate: true
|
|
when: llama_embed_enabled
|