Self-hosting

Base is a Django API with Celery workers and a Next.js web app. Running it on your own machine takes seven processes, two Python environments and a few settings; training on Modal also needs Base to be reachable from the internet.

The repository's README.md has the complete list of settings; this page covers what a local Base needs.

Processes

Each runs in its own terminal, from backend/ unless noted:

runserver serves HTTP and WebSockets: Daphne is installed, so it runs the ASGI app. Keep the train worker's concurrency low: a launch holds its slot for minutes while Modal builds the image and uploads datasets.

For a quick look without workers, CELERY_TASK_ALWAYS_EAGER=1 python manage.py runserver 8000 runs every task inline.

Two Python environments

Django and solarian live in separate environments. Base runs the solarian CLI as a subprocess for dataset ops, deployment builds, training launches, relays and controller sessions.

# Base: Django, Celery, and the modal client that polls training jobs
conda create -n solarian_base python=3.10 -y && conda activate solarian_base
cd backend && pip install -e ".[dev]"

# solarian: the CLI Base runs, with the plugins training needs
conda create -n solarian python=3.10 -y && conda activate solarian
pip install solarian solarian-modal solarian-base

Base finds the CLI in CONTROLLER_SOLARIAN_BIN, then on PATH, then in a conda env named solarian next to the one Django runs in.

Configure

Settings come from backend/.env or the process environment, read when each process starts: restart the API and the workers after a change.

Then create the database:

python manage.py migrate
python manage.py createsuperuser   # optional, for /admin/

Training on Modal from a local Base

A training job reports to PUBLIC_API_URL. A Modal container cannot reach localhost, nor a Tailscale address, since it is not on your tailnet. Tailscale Funnel publishes the API at a public HTTPS address:

tailscale funnel --bg --set-path /api http://127.0.0.1:8000/api
tailscale funnel status          # https://<machine>.<tailnet>.ts.net/api

Only /api is published. Then:

1. Set PUBLIC_API_URL=https://<machine>.<tailnet>.ts.net and restart the train worker.

2. Log the CLI in through the same URL, so trainings started from your machine report there too:

   solarian base login --url https://<machine>.<tailnet>.ts.net
   

   CopyCopied!

3. Check from outside: curl -i https://<machine>.<tailnet>.ts.net/api/projects/ answers 401.

Check it

• Web app: http://localhost:3000
• API: http://localhost:8000/api/, with interactive docs at /api/docs/
• Admin: http://localhost:8000/admin/
• redis-cli ping answers PONG

See also

• Getting Started — connecting the CLI to your Base.
• Training & Runs — what the train worker launches.