`pinned: true` model is freed after every request — idle-unload exclusion (#9305) ineffective on v4.7.1

[loryanstrant]

Follow-up to #9305 / #9309 (exclude models from idle unloading via pinned: true). On v4.7.1 a pinned: true model is not kept resident — its backend process is torn down after every request, so each call pays a full cold reload.

Environment

• LocalAI v4.7.1, image localai/localai:v4.7.1-gpu-nvidia-cuda-13, single NVIDIA 16 GB GPU, llama-cpp (cuda13) backend.
• Model YAML: backend: llama-cpp, pinned: true, context_size: 24576.
• Runtime settings: watchdog_enabled: false, single_backend: false, max_active_backends: 0 (docs: "0 = unlimited"), memory_reclaimer_enabled: false, size_aware_eviction: false, force_eviction_when_busy: false.

Expected

Per the docs ("a pinned model is never evicted") and #9305: with no watchdog, no LRU limit, and pinned: true, the model should stay loaded across requests.

Actual

The backend is freed immediately after each response completes (not after any idle timeout):

INFO  ... Predict request completed successfully
DEBUG Deleting process model="qwen3.5-9b"           (pkg/model/process.go:49)
DEBUG Calling Free() to release GPU resources model="qwen3.5-9b"
INFO  Backend process stopped id="qwen3.5-9b" exitCode="0"

This repeats for every single request across the log; the next request cold-loads (~10 s). Confirmed via nvidia-smi: the backend process disappears seconds after each response and reloads on the next call.

Repro

1. Load a pinned: true llama-cpp model on v4.7.1 with the runtime settings above.
2. Send two chat completions about a minute apart.
3. Observe (via nvidia-smi) the backend process disappear after the first completes and reload for the second.

Question

Is pinned: true expected to keep a model resident, or only to exempt it from eviction while some other path still frees it? If max_active_backends: 0 (documented as "unlimited") is implicated, the "0 = unlimited" behaviour and the #9305 exclusion don't appear to hold. Net effect: pinned currently provides no persistence benefit — every request pays a full cold model load.

[localai-org-maint-bot]

@mudler this is a valid regression report: pinned is intended to exclude a model from idle/LRU/memory-pressure eviction, and max_active_backends: 0 is documented and implemented as unlimited. The immediate post-request teardown shown here is therefore not expected. The distributed worker lifecycle is the important distinction from the existing watchdog pin tests; this needs a focused reproduction/test on the worker-supervisor path before a fix is proposed.