feat(llama_server): heartbeat embed — garde chaud le slot nomic :1238 (#28)

Le grounding RAG du STT ratait par intermittence (docs=0 → réponses génériques)
parce que l'instance d'embedding :1238 partait à froid après une pause : la 1ʳᵉ
requête mettait 5-8 s et dépassait le timeout 4 s du STT-server (recall + RAG doc).

llama-embed-heartbeat : service systemd local sur gpu-01 qui envoie un petit embed
à :1238 toutes les 20 s → le slot reste résident/chaud → la 1ʳᵉ vraie requête est
rapide → docs>0 fiable. Pendant du llm-heartbeat (slot chat :1234), pour le slot embed.

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
ALI YESILKAYA 2026-06-19 22:17:23 +02:00 committed by GitHub
parent 11693e2abf
commit 1b85d4ee65
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
7 changed files with 140 additions and 0 deletions

View file

@ -124,6 +124,34 @@
immediate: true
when: llama_embed_enabled
# --- Heartbeat embed (garde chaud le slot :1238) — optionnel ------------------
- name: Deploy llama-embed-heartbeat script
ansible.builtin.template:
src: llama-embed-heartbeat.sh.j2
dest: /usr/local/bin/llama-embed-heartbeat
mode: '0755'
notify: Restart llama-embed-heartbeat
when: llama_embed_enabled and llama_embed_heartbeat_enabled
- name: Deploy llama-embed-heartbeat service
ansible.builtin.template:
src: llama-embed-heartbeat.service.j2
dest: /etc/systemd/system/llama-embed-heartbeat.service
mode: '0644'
notify:
- Reload systemd
- Restart llama-embed-heartbeat
when: llama_embed_enabled and llama_embed_heartbeat_enabled
- name: Enable llama-embed-heartbeat service
ansible.builtin.systemd:
name: llama-embed-heartbeat
enabled: true
daemon_reload: true
state: started
when: llama_embed_enabled and llama_embed_heartbeat_enabled
# --- Watchdog auto-réparation (wedge ROCm) — optionnel ------------------------
- name: Deploy llama-watchdog script