Skip to content

Launcher

aivrar edited this page Nov 18, 2026 · 1 revision

Launcher

launcher.bat opens a small desktop app that starts LocalSoundsAPI servers, downloads models and shows every log in one window. On a fresh clone it runs the one-click installer first.

Instances & Log

Launcher: instances and log

  1. Pick a Device (every NVIDIA GPU found, plus CPU) and a Port (starts at 5006).
  2. Click Add Instance, then Start.
  3. Double-click the row, or click Open in Browser, to open the web UI.
Button What it does
Start / Stop Start or stop the selected instance
Start All / Stop All Every instance at once
Remove Stop and delete the selected instance
Open in Browser Open http://127.0.0.1:<port>

Status is checked every 15 s: Starting → Running, or Crashed if the process exited (the log says why, usually a busy port or out of memory).

The log shows every instance's output with a [GPU1:5006] prefix. Filter by instance, clear it, and it keeps the last 5,000 lines.

One instance or several?

Setup When to use it
One instance, engines spread over GPUs The usual choice. Each panel has its own device picker, and different engines run at the same time (Concurrency & Progress)
One instance per GPU Two copies of the same engine generating at once, e.g. two XTTS jobs. Each instance only sees its own GPU (CUDA_VISIBLE_DEVICES)

Models & Tools

Launcher: models and tools

  • Portable Python: shows the bundled Python; Download / Reinstall repairs it.
  • Portable tools: FFmpeg, RubberBand and eSpeak-ng, with a green dot when installed; one button reinstalls them all.
  • Hugging Face token: only needed for gated models (Fish Speech). Create one at huggingface.co/settings/tokens.
  • AI models: pick one and click Download Selected Model to fetch it ahead of time. Models also download automatically the first time you click Load in the web UI.
Model Folder Size
XTTS-v2 models/XTTS-v2/ ~6 GB
Fish Speech models/fish-speech/ ~8 GB
Kokoro-82M models/kokoro-82m/ ~0.5 GB
Whisper medium models/medium.en.pt ~1.5 GB
Stable Audio Open models/stable-audio-open-1.0/ ~10 GB
ACE-Step models/ace_step/ ~7 GB
CLAP models/clap-htsat-unfused/ ~0.6 GB

Closing the launcher stops every instance it started (it asks first). Prefer the command line? python main.py --port 5006 still works.

Clone this wiki locally