Betreiberbeschluss 01.11.2026: Swift 1.5 Q6_K aufgenommen (GNAI-009 3.4, Artefakt gepinnt: Revision 14bfe4b, SHA-256 cd2eb65 aus HF-Blob-Metadaten, 22,2 GiB). Testmatrix um S-Reihe und AB3/AB4 erweitert (AB3 isoliert die Thinking-Optimierung Swift 1.0 vs 1.5, AB4 ist der Praxisvergleich). Benchmark-Design ergaenzt: lm-evaluation-harness gegen die interne Engine-API (AIME, GPQA Diamond, MATH-500, IFEval) plus Kernkennzahl Score je Median-Reasoning-Token aus der getrennten reasoning_content. model_store: drittes Artefakt; inference_engine: Profil swift15 (Unit gnai-testing-llama-swift15, Alias swift-1.5-qwen3.8-27b-q6k). Generated by Mistral Vibe. Co-Authored-By: Mistral Vibe <vibe@mistral.ai>
inference_engine
Deploys the inference engine for one KI environment as rootless Podman
Quadlets (GNAI-010 sections 5 and 6; first profile: llama.cpp). Runs after
model_store in the same play and reuses model_store_root.
Variables
inference_engine_image_ref– immutable image digest (ghcr.io/...@sha256:...).inference_engine_user/inference_engine_uid/inference_engine_user_home– workload account created by the hostworkload_accountsrole.inference_engine_unit_prefix– unit and container prefix of the environment (Testing:gnai-testing-llama).inference_engine_network_name/inference_engine_network_alias– internal-only Quadlet network (Internal=true), aliasllama-cppfor the later API policy layer.inference_engine_profiles– map of model profiles; each needsunit_suffix,model_directory,model_file,alias,ctx_size,n_predict,reasoning_effort, sampling values and anmtpblock.inference_engine_active_profile– the single profile to start.inference_engine_ctx_size_override– when notnull, replaces every profile'sctx_size(Testing bring-up: 65536; set tonullafter the 128k acceptance, GNAI-008 section 7.5).
Behaviour
- Validates declarations, requires
/dev/kfdplus exactly two render nodes, and requires every profile's model artifact (run--tags modelsfirst). - Pulls the pinned image by digest into the environment's rootless store only when missing.
- Renders
<network>.networkand one.containerper profile into/etc/containers/systemd/users/<uid>/; units are never enabled fordefault.target. - Units mutually exclude each other via
Conflicts=/Before=(GNAI-010 section 5.3). - Exercises the Quadlet generator with
--dryrunbefore any reload; quadlet changes notify adaemon-reloadof the workload user manager. Services are never restarted automatically. - Switch/start (only with the
servicetag, never in check mode): stop any other active engine unit, start the selected unit, wait for/health, and verify the served alias via/v1/models. The HTTP check uses whichever ofcurl/wgetexists in the pinned image (GNAI-010 section 8).
Tags
service– everything in this role. Model provisioning belongs to themodelstag of themodel_storerole.
Logs
conmon writes the container output to the system journal (journald driver) with the syslog identifier set to the container name. Read as root:
journalctl _COMM=conmon -t gnai-testing-llama-swift --no-pager
podman logs returns empty for the workload account (journal file
permissions); acceptance protocols are collected as root (GNAI-010 section
8.1).
Notes
- There is no
PublishPort=; the engine is reachable only inside the environment network. - The role does not start the engine in check mode and never downloads models.