gpu-free

name	gpu-free
description	Use to pause / resume Jarvis voice services (Whisper STT, Chatterbox TTS, Kokoro TTS docker) on <INFERENCE_HOST> to free RTX 4090 VRAM for testing larger Ollama models. Voice will not work while paused — explicit cost. Pairs with the <VOICE_RUNTIME> env-var toggles for durable on/off. Trigger phrases — "/gpu-free", "free up the gpu", "pause jarvis", "stop voice services so I can test a model", "resume jarvis".
category	execution
runtimes	["claude"]
pii_safe	true

/gpu-free — Pause/Resume Jarvis on <INFERENCE_HOST> to free GPU

<INFERENCE_HOST>'s RTX 4090 (24 GB) is shared by Ollama and the <VOICE_RUNTIME> voice stack (Whisper STT + Chatterbox TTS + Kokoro TTS docker). When testing a 17–24 GB model, the voice services hog ~5–9 GB of VRAM and can starve the model.

This skill stops them temporarily so you can test, then restarts them when done. For durable off-by-default mode (e.g., on-the-go text-only Jarvis), use the <VOICE_RUNTIME> env-var toggles instead — see "Durable mode" below.

⚠️ Voice goes silent while paused — no STT, no TTS, no Jarvis voice replies. Type to interact. Resume when finished.

Durable mode (preferred for "off-by-default")

When <USER_NAME>wants the voice stack permanently off (e.g., on-the-go), set these env vars in <LOCAL_PATH>/dev/<VOICE_RUNTIME>-voice/.env:

JARVIS_STT_ENABLED=false
JARVIS_TTS_CHATTERBOX_ENABLED=false
JARVIS_TTS_KOKORO_ENABLED=false

Then systemctl --user restart <VOICE_RUNTIME>-voice.service. The startup hook (src/service-control.js, merged from feature/voice-service-toggles on 2026-05-06) automatically stops the matching systemd units and Docker container on every <VOICE_RUNTIME>-voice startup. Survives reboots without ritual.

/gpu-free is for AD-HOC, "I'm running a benchmark for the next 10 minutes" usage. Don't pile both mechanisms on top of each other for the same outcome — pick one per use case.

What gets paused

Service	VRAM	Notes
whisper-service.service (system)	~1.4 GB	Whisper STT large-v3-turbo — <USER_NAME>s preferred STT
jarvis-chatterbox-tts.service (user)	~3.9 GB	<USER_NAME>voice clone TTS
Kokoro TTS docker (port 8880)	~1.1 GB	Backup TTS

Plus any Ollama models currently warm — those are evicted via keep_alive=0 so the next model load gets a clean GPU.

Subcommands

gpu-free pause       # stop services + evict ollama models, show VRAM
gpu-free resume      # restart services, show VRAM + compute-apps
gpu-free status      # show current state without changing anything
gpu-free help

Script location

~/.local/bin/gpu-free on <WORKSTATION_HOST> (drives generic via SSH). Distribute to other machines as needed via scp ~/.local/bin/gpu-free <host>:~/.local/bin/.

Behavior notes

• Idempotent — pause when already paused is harmless; same for resume.
• Uses ssh <INFERENCE_HOST> — needs working SSH. Override target: GENERIC_HOST=<other> gpu-free pause.
• The jarvis-whisper-stt.service (user-level large-v3) is intentionally NOT in the pause list because <USER_NAME>disabled it as a duplicate on 2026-05-06 (kept the system-level large-v3-turbo). If it gets re-enabled later, add it back to USER_SVCS in the script.
• Kokoro is a USER-level systemd unit (kokoro-tts.service under systemctl --user), not a system unit. The script stops via docker stop <container> which kills both the unit and the container in one shot. As of 2026-05-06, kokoro-tts.service is also disabled at user level so it doesn't auto-restart on next login.
• Kokoro restart is best-effort: gpu-free resume tries docker start on the existing container. If you've fully removed/disabled it (current state), gpu-free resume won't bring it back — you must manually systemctl --user enable --now kokoro-tts.service.
• Coexistence with env-var toggles: if JARVIS_*_ENABLED=false is set in <VOICE_RUNTIME>-voice's .env, the <VOICE_RUNTIME> startup hook will re-stop services on every <VOICE_RUNTIME>-voice restart. So if you gpu-free resume AND <VOICE_RUNTIME>-voice restarts, services come up only briefly before being stopped again. Either flip env vars to true first, or use gpu-free only in AD-HOC bursts where <VOICE_RUNTIME>-voice isn't restarting.

After resume — verify

gpu-free status
ssh <INFERENCE_HOST> 'curl -s http://localhost:8000/health'   # whisper
ssh <INFERENCE_HOST> 'curl -s http://localhost:8880/health'   # kokoro

If a service didn't come back, journalctl --user -u <service> -n 20 (or sudo journalctl -u <service>) will show why.

Why this exists

Created 2026-05-06 after testing qwen3.6:27b on <INFERENCE_HOST> and discovering 8.9 GB of VRAM was held by voice services, leaving only 715 MB free. Pausing them yielded full 23 GB free for the model.