Files
infrastructure/llm-server/llm/ansible
spektr c6cabd98ef LLM-Server: Benchmark-Harness (lm-eval) aufgesetzt und per Smoke-Test verifiziert
Neue Rolle benchmark (GNAI-009 5.1/5.2): Egress-Netzwerk-Quadlet
gnai-testing-bench (dokumentierte Ausnahme nur fuer Datensatz-Downloads),
persistentes results/- und HF-Cache-Verzeichnis, gepinntes Image
gnai-bench:lm-eval-0.4.13 (python:3.13-slim per Digest, lm-eval[api]-Pin,
bewusst ohne Torch), Runner-Skript mit Network-Unit-Absicherung.

Live-Funde aus dem Smoke-Test (2 AIME-Samples gegen Swift 1.5):
- lokale Chat-Completions brauchen --apply_chat_template (Server wendet
  sein --jinja-Template an, kein lokaler Tokenizer noetig)
- Request-Timeout 1800 s statt lm-eval-Default 300 s (xhigh-Thinking bei
  ~18 t/s)
- Quadlet-Netzwerke materialisieren erst mit gestarteter Unit; Rolle und
  Runner starten sie explizit
- Ergebnis: exact_match 1.0, Per-Sample-Logs in results/; Tokenzaehlung
  erfolgt nachtraglich exakt ueber die /tokenize-Endpunkte der Engine

Generated by Mistral Vibe.
Co-Authored-By: Mistral Vibe <vibe@mistral.ai>
2026-11-02 05:58:58 +01:00
..

LLM serving stack – Ansible control project

Executes the LLM serving stack for exactly one environment on ghostnet-ki (GNAI-008 section 7.4). Host baseline, Podman runtime, workload accounts and AMD GPU access are managed by ../../host/ansible; this project never changes host state.

Layout

  • inventories/<environment>/ – one inventory per KI environment (currently testing); deployment uses the ansible-deploy account with become: true.
  • playbooks/ai.yml – applies model_store then inference_engine.
  • roles/model_store/ – shared, pinned model weights under /var/lib/ghostnet-ai/shared/models (GNAI-010 sections 3 and 4).
  • roles/inference_engine/ – rootless Quadlets, network, image provisioning and the exclusive profile switch for llama.cpp (GNAI-010 sections 5 and 6).

Control node setup

Use the repository-wide control venv built from ../../../shared/ansible/requirements-control.txt (ansible-core==2.21.4), for example:

python3 -m venv .venv
.venv/bin/pip install -r shared/ansible/requirements-control.txt
cd llm-server/llm/ansible
../../../.venv/bin/ansible-playbook -i inventories/testing playbooks/ai.yml --syntax-check

Usage (Testing)

All commands use the operator deploy key ~/.ssh/id_ed25519_ghostnet_ai_ansible_deploy (see AGENTS.md; no ssh-agent required):

# Preview only – no changes, no downloads, no service start:
../../../.venv/bin/ansible-playbook -i inventories/testing playbooks/ai.yml \
  --check --diff --private-key ~/.ssh/id_ed25519_ghostnet_ai_ansible_deploy

# Provision model weights only (tag models):
../../../.venv/bin/ansible-playbook -i inventories/testing playbooks/ai.yml \
  --tags models --private-key ~/.ssh/id_ed25519_ghostnet_ai_ansible_deploy

# Render quadlets and start/verify the selected profile (tag service):
../../../.venv/bin/ansible-playbook -i inventories/testing playbooks/ai.yml \
  --tags service --private-key ~/.ssh/id_ed25519_ghostnet_ai_ansible_deploy

inference_engine_active_profile selects which rendered quadlet is started; all other profiles stay rendered but stopped. Check mode downloads no model and starts no service. A second identical run must report changed=0 (GNAI-010 section 10.4).

The full technical acceptance (64k/128k start, VRAM distribution, prefill ladder, restart behaviour) is an operator-run protocol documented in GNAI-010; this role automates deployment and basic health only.