FROM python:3.11-slim

WORKDIR /app

# Prevent Python from buffering stdout/stderr
ENV PYTHONUNBUFFERED=1

# Set HuggingFace cache to persistent storage
ENV HF_HOME=/data/.huggingface
ENV HF_HUB_CACHE=/data/.huggingface/hub

# Install system dependencies
RUN apt-get update && apt-get install -y \
    ffmpeg \
    libopenblas0 \
    && rm -rf /var/lib/apt/lists/*

# Install AVX512 + OpenBLAS optimized llama-cpp-python wheel (MAXIMUM SPEED!)
RUN pip install --no-cache-dir \
    https://huggingface.co/datasets/AIencoder/llama-cpp-wheels/resolve/main/llama_cpp_python-0.3.16%2Bcpuavx512_openblas-cp311-cp311-manylinux_2_31_x86_64.whl \
    || pip install --no-cache-dir llama-cpp-python

# Install Python dependencies
RUN pip install --no-cache-dir \
    gradio>=5.0.0 \
    faster-whisper \
    huggingface_hub

# Copy application
COPY app.py /app/app.py

# Expose port
EXPOSE 7860

# Run the app directly
CMD ["python", "app.py"]