# Dockerfile for TinyMetatron SLM (CPU-only) — Hugging Face Docker Space. # Python 3.13-slim base per IMPLEMENTATION_CONTRACT.md section 2 / README.md. FROM python:3.13-slim WORKDIR /app # Environment: UTF-8 stdio, import path, runtime paths, deploy mode. # TMT_DEPLOY_MODE=demo -> read-only public Space (training/data-write disabled). # TMT_DB_PATH / TMT_CHECKPOINT_DIR point at image-layer /app paths so the demo # works with NO attached storage. Override both to a mounted /data volume for a # stateful private deployment. ENV PYTHONPATH=/app \ PYTHONUNBUFFERED=1 \ PYTHONIOENCODING=utf-8 \ PIP_NO_CACHE_DIR=1 \ TMT_DB_PATH=/app/metatron.db \ TMT_CHECKPOINT_DIR=/app/ckpt \ TMT_DEPLOY_MODE=demo # CPU-only PyTorch wheel (no CUDA runtime -> smaller image). RUN pip install --no-cache-dir torch==2.13.0 --index-url https://download.pytorch.org/whl/cpu \ && pip install --no-cache-dir fastapi==0.115.6 uvicorn[standard]==0.34.0 pydantic==2.10.4 # Copy the flat repo (no package) into /app. COPY . /app # Ensure runtime directories for data and checkpoints exist. RUN mkdir -p /app/data /app/ckpt # Build-time training: seed a small corpus and train a checkpoint so the Space # ships with a real active model in /app/ckpt + an active row in /app/metatron.db. # /generate finds the active checkpoint at runtime with no extra config. # (~20-40s on a CPU builder for a 6.35M model; no large file committed to the repo.) RUN python manage_data.py generate --domain cybersecurity --count 400 \ --db_path /app/metatron.db \ && python train_db.py --steps 300 --domain cybersecurity \ --min_quality 0.4 --db_path /app/metatron.db --checkpoint_dir /app/ckpt # Hugging Face Docker Spaces expect the app on port 7860 (app_port in README). EXPOSE 7860 # Serve the FastAPI app (module-level `app` in api.py) on all interfaces. CMD ["uvicorn", "api:app", "--host", "0.0.0.0", "--port", "7860"]