# SN59 babelbit ARENA miner — gist S2S (whisper-large-v3 → Qwen2.5-7B gist → Kokoro)
# Validated qualifying score 0.7397 (place #1). Models are DOWNLOADED AT STARTUP (not baked)
# to keep the image small enough for the arena registry size check (~8 GB vs 28 GB baked).
# Server self-warms (downloads + loads whisper+Qwen+Kokoro) at startup; serves /v1/predict + /predict + health.
FROM nvidia/cuda:12.6.0-cudnn-runtime-ubuntu22.04

ENV DEBIAN_FRONTEND=noninteractive PIP_NO_CACHE_DIR=1 \
    HF_HOME=/models/hf PIP_BREAK_SYSTEM_PACKAGES=1

RUN apt-get update && apt-get install -y --no-install-recommends \
    python3 python3-pip espeak-ng ffmpeg ca-certificates && \
    rm -rf /var/lib/apt/lists/*

# torch (CUDA 12.x) + serving + model deps (gist pipeline)
RUN pip3 install --no-cache-dir torch --index-url https://download.pytorch.org/whl/cu124 && \
    pip3 install --no-cache-dir \
      faster-whisper transformers accelerate bitsandbytes sentencepiece \
      kokoro "misaki[en]" espeakng-loader phonemizer-fork \
      soundfile numpy huggingface_hub fastapi "uvicorn[standard]"

WORKDIR /app

# Models are NOT baked (would make the image ~28 GB and fail the arena size check).
# whisper-large-v3 + Qwen2.5-7B + Kokoro-82M download at startup into HF_HOME on first warm.
RUN mkdir -p /models/hf

COPY server.py model.py gist_model.py /app/

# Winning config (validated). Managed/arena predict path is /predict; health is GET /health.
# Babelbit arena contract: port 8000, serves /v1/predict + /predict + /healthz + /health.
ENV BB_MINER_MODEL=gist \
    BB_ASR_MODEL=large-v3 BB_ASR_BEAMS=5 \
    BB_GIST_LLM=Qwen/Qwen2.5-7B-Instruct BB_GIST_SPEED=1.15 BB_GIST_VOICE=af_heart \
    PORT=8000

EXPOSE 8000
# Server self-warms (loads whisper+Qwen+Kokoro) on startup before accepting traffic.
CMD ["sh","-c","uvicorn server:app --host 0.0.0.0 --port ${PORT} --log-level info"]