# deploy/tgi_space/Dockerfile.fallback
#
# Fallback: custom transformers FastAPI server for Spaces where TGI fails.
# Uses Qwen/Qwen2.5-0.5B-Instruct (smaller model, faster on CPU, ~1GB).
#
# Deploy by renaming to Dockerfile and uploading to the failing Space:
#   cp deploy/tgi_space/Dockerfile.fallback deploy/tgi_space/Dockerfile
#   # then re-upload via setup_spaces.py --space-name tgi-persona-synth

FROM python:3.11-slim

WORKDIR /app

# Install inference deps — two stages to keep --index-url unambiguous
RUN pip install --no-cache-dir fastapi "uvicorn[standard]" httpx transformers accelerate
RUN pip install --no-cache-dir torch --index-url https://download.pytorch.org/whl/cpu

COPY app.py .

ENV PORT=7860
# Use 0.5B model — smaller, faster on CPU, still generates valid JSON personas
ENV MODEL_ID=Qwen/Qwen2.5-0.5B-Instruct
ENV HF_HOME=/data
ENV HUGGINGFACE_HUB_CACHE=/data/hub
ENV MAX_NEW_TOKENS=600

# Expose port required by HF Spaces
EXPOSE 7860

CMD ["python", "app.py"]