NF4 re-host on op (.54) — lazy loading + fleet re-point (2026-09-13)

Moved the default TTS endpoint off mfwin-WSL (op-2, research workstation) to op (4790, research workstation):12439 with lazy loading, per request. The op-2 service keeps running as fallback (mac still points at it until it wakes).

Layout on .54 (WORKSTATION/breeze-tts2-nf4)

.venv/         torch 2.9.1+cu126, transformers 4.57.3, qwen-tts 0.1.1,
               bitsandbytes 0.50.2, accelerate 1.12.0, fastapi/uvicorn
merged-400/    staged from .51 (sha256-verified, see receipt below)
breeze_infer/ models/ training/   code rsynced from breeze-tts2-finetuning
server.py      deploy/server_lazy.py (lazy variant, see below)
breeze-tts.service  systemd unit (same as deploy/, After=tailscaled)

Source of the model: .51:WORKSTATION/merged-400 (op-2 has no sshd by design; mfwin:22 closed — .51 had the merge output).

Lazy loading (deploy/server_lazy.py)

Same generation path as deploy/server.py (NF4 backbone+depth, temp 0.7, seed 42, silence early-stop + trim, wav/mp3/pcm). Differences:

• lifespan does NOT load the model; service starts instantly with 0 VRAM.
• First POST /v1/audio/speech loads under the request lock (double-checked; concurrent callers wait), then it stays resident.
• GET /health reports status: idle | loading | ready plus lazy, loads.
• Knobs (env): BREEZE_LAZY=0 restores eager startup; BREEZE_MODEL_DIR; BREEZE_IDLE_UNLOAD_SEC>0 optional idle unload (default off).

Rationale: .54 shares the RTX 3060 with the MemPalace vLLM (Nemotron AWQ, ~2.5 GB) — don’t grab 5.5 GB at boot or on every restart.

Verification receipts

• Model transfer: sha256sum model-*.safetensors identical .51 vs .54 — 457773cd…ac2ed64, 1aff4236…8d76590. (First attempt nested the dir one level; flattened — tokenizer_class lookup failure was the symptom.)
• /health: {"status":"idle","lazy":true,"loads":0} at boot → ready with vram_allocated_gb: 5.442, nf4_modules: 282 after first request. vLLM EngineCore (2 548 MiB) untouched throughout.
• Speech: wav 4.24 s audio (seed 42) and mp3 128 kbps both 200 OK; first request ≈ 46 s incl. load, warm RTF ≈ 7.8 (GPU shared with vLLM; mfwin dedicated was ≈ 6.9).
• Fleet Hermes breeze_latam URL → private workstation service on .51, .54, .55 (sed + backup config.yaml.bak-breeze54-20260913; Hermes reloads on config mtime). Wrapper synthesis verified on each (deterministic wav, 105 644 B for the shared test line).
• Pending: mac (op-1) asleep/unreachable — still on op-2:12439 (working fallback); flip when awake (same one-line sed).

Rollback

Per machine: restore config.yaml.bak-breeze54-20260913. On .54: sudo systemctl disable --now breeze-tts (frees 5.4 GB VRAM; vLLM was never touched). op-2 endpoint stays up regardless.