DEPLOY.md

Deploy LLM Cinema to a Hugging Face Space (under the hackathon org)

Everything below runs from the claudemovies/ folder. You need to be logged into HF (hf auth login — paste a token with write access).

1. Create the Space under the org

hf auth login
hf repo create build-small-hackathon/llm-cinema --repo-type space --space_sdk gradio

(or create it in the web UI: huggingface.co/new-space → owner = build-small-hackathon, SDK = Gradio, name = llm-cinema.)

2. Upload the app (only what the Space needs)

hf upload build-small-hackathon/llm-cinema . . --repo-type space \
  --exclude "saved/*" "*.log" "*.jsonl" "*.bak" ".endpoint_key" \
            "drawn_cache.json" "__pycache__/*" "*.pyc"

This pushes stage.py (the app), the engine modules, sprites.txt, the showcase/ gallery films, README.md, and requirements.txt.

3. Set the Space secrets

Space → Settings → Variables and secrets → add (point these at your own OpenAI-compatible endpoint; the key never lives in the repo):

Name	Value
CLAUDEMOVIES_LLM_URL	https://<your-endpoint>/v1
CLAUDEMOVIES_LLM_MODEL	cinema
CLAUDEMOVIES_LLM_KEY	(your bearer key)

4. Verify (the judge's-eye check)

Open the Space in a fresh browser:

• the knight film auto-plays within a couple seconds (no model needed),
• Gallery films replay instantly,
• Make a film → type a concept → Cinema writes + plays it live,
• Adventure → Begin → A/B/C choices appear,
• Download produces a GIF.

Warm both endpoints right before judging / recording the demo (open the Space, run one Make and one Adventure) so the first judge click isn't a cold start.

Keeping the endpoints awake during judging

Both Modal apps scale to zero after 15 min idle. They wake on the first request (~1–2 min cold start for the teacher, ~2–4 min for Cinema loading MiniCPM). The Gallery covers that gap (always instant). If you want them always-warm during judging, redeploy with a larger scaledown_window — but that spends more credits.