Repository navigation
Commands
lw.py [--hosts HOSTS] COMMAND. --hosts names the registry file; the default is $LABWRIGHT_HOME/hosts.json. Every command only reads.
| Command | What it answers |
|---|---|
hosts |
the registry, validated |
status [--json] |
each host idle, busy or down, with reasons (queue, loaded models, GPU load) |
where MODEL |
which host has a checkpoint, LoRA or Ollama tag |
pick --role ROLE [--model M] [--vram GB] [--batch] |
the idle host to run on, or the busy host to queue behind |
place --size GB [--role ROLE] |
which model stores keep the floor after a download |
doctor |
status plus drift between the registry and the probes |
ROLE is one of image, video, music, llm, train.
python lw.py hosts
Prints the registry after checking it.
python lw.py status
python lw.py status --json
Probes each host and reports idle, busy or down with the reason. The registry is a dated snapshot; the probe is the truth, so run this before calling a host free. Held VRAM on an idle host is reported.
python lw.py where flux1-dev
Lists the hosts that have a model by that name.
python lw.py pick --role image
python lw.py pick --role video --vram 24
python lw.py pick --role llm --model qwen3:8b
python lw.py pick --role image --batch
The answer is one of three:
-
run: the idle host to use. With--batch, every idle host, so the work can be split. -
queue: no capable host is idle; it names the busy one to queue behind and wait on. -
none: nothing capable is up. Say so rather than switching to another source.
--vram adds a note when the job will not fit a host's card.
python lw.py place --size 24
python lw.py place --size 12 --role image
Lists every store with its free space, the free space after the download and ok when that stays at or above floor_gb. Remote stores use the dated free_gb from the registry; ok: null means unknown, so measure before downloading. It never downloads; check the model's licence and gating first.
python lw.py doctor
Runs status and compares the registry with what ComfyUI and nvidia-smi report. Run it on a new machine, after a hardware change, or when results look odd.
This wiki describes labwright 0.2.0 and was last updated on 2026-11-15. MIT licensed. Report problems in the issues.