relay-deploy

Production deploy artifacts for the relay JEPA contact-push puzzle — 26M-param ViT-Tiny + AR transformer predictor + DexWM joint state head, trained on actor-pattern-corrected data with a uniform-sampled head decoder. Plans via CEM in latent space.

Runs on the JEPA inference VPS at jepa.waweapps.win/relay/ behind a Caddy reverse proxy (jepa-vps-proxy).

What's here

A self-contained code subset of the private aura research repo — only what's needed to run the inference server. No training scripts you could use to reproduce the model (those stay in the private repo), no data, no notebooks.

• Dockerfile — CPU PyTorch + DINOv2 weights pre-downloaded at build time
• docker-compose.yml — joins the shared external web Docker network
• world_model/, scripts/, client/relay/ — minimum source for the FastAPI inference server + the browser client
• checkpoints/ — gitignored; populate via scp

Deploy on a fresh VPS

Prereqs: Caddy proxy already running on the web network with /relay/* → relay:8800 routing (see jepa-vps-proxy repo).

ssh jepa-vps
cd /srv/relay
git clone https://github.com/SotoAlt/relay-deploy.git .
docker compose up -d --build

Then from your Mac, scp the checkpoint over (one-time):

scp /path/to/relay_stage1_v9_trackE_ep02_uhead.pt \
    jepa-vps:/srv/relay/checkpoints/
ssh jepa-vps 'cd /srv/relay && docker compose restart'

Verify: open https://jepa.waweapps.win/relay/.

Updating

ssh jepa-vps 'cd /srv/relay && git pull && docker compose up -d --build'

Caddy and other apps (lepong, …) are not touched. Independent-deploy-safe by design.