Skip to content

Configuration

This page lists the Gradio account, the database, and the input limits. The starter does not take a provider key. /opt/sushi/docs/README.md on the instance is the package reference.

Where things live

Path What it is
sushi-gradio.service The application service. It starts on boot as a non-root user.
127.0.0.1:7860 Authenticated page and queue API. Open it through a tunnel.
/home/ubuntu/.config/sushi-gradio/access.json Username and password, mode 0600, created on first boot.
/var/lib/gradio/predictions.sqlite3 Saved predictions.
/var/lib/gradio Service state to include in a backup.
~/gradio-results JSON files written by sushi-gradio predict.
/opt/sushi/docs/demo.sh Walkthrough.
/opt/sushi/docs/README.md Image notes.

The service keeps the first identity and the saved predictions across restarts. Do not delete the access file or the identity files to rerun setup.

Sign-in account

Read access.json in the ubuntu session when you need the username and password. The helper reads the same file when you run sushi-gradio predict and sushi-gradio read. Do not copy the file into a command line, a log, or a ticket.

Input the model accepts

  • At most 512 characters.
  • At most 64 English words.
  • The text includes a word from the original examples.

The documented examples are helpful excellent service and awful poor service. The page and the terminal client use the same queue API. Pass a new --output path for every sushi-gradio predict call.

The model is a local teaching model. Do not use its class or probabilities for a consequential decision. The starter does not download weights and does not call a model provider.

Listen address

Forward workstation port 7860 to the instance's port 7860, and open http://127.0.0.1:7860 on the workstation. The published guide says to keep both ends bound to localhost and to leave the application port off the internet. It does not describe a public listener.

Last updated October 27, 2026 · Suggest an improvement