Dockerfile

FROM nvidia/cuda:12.8.1-cudnn-runtime-ubuntu24.04

ENV DEBIAN_FRONTEND=noninteractive \
    PYTHONUNBUFFERED=1 \
    PIP_DISABLE_PIP_VERSION_CHECK=1

RUN apt-get update && apt-get install -y --no-install-recommends \
    ca-certificates \
    git \
    libportaudio2 \
    libsndfile1 \
    espeak-ng \
    ffmpeg \
    python3-pip \
    curl \
    build-essential \
    && rm -rf /var/lib/apt/lists/*

RUN pip install --break-system-packages --no-cache-dir uv

# Ubuntu 24.04 ya trae un usuario 'ubuntu' con UID 1000 — hay que sacarlo antes.
RUN userdel -r ubuntu 2>/dev/null || true \
    && useradd -m -u 1000 user
USER user
ENV HOME=/home/user \
    HF_HOME=/home/user/hf_cache

WORKDIR /home/user/app

# venv con Python 3.11 (numba/llvmlite viejos no compilan en 3.12).
RUN uv venv /home/user/venv --python 3.11
ENV VIRTUAL_ENV=/home/user/venv \
    PATH=/home/user/venv/bin:/home/user/.local/bin:$PATH

# Torch pinneado a 2.10.0 desde el índice CUDA 12.8. La 2.11 tiene un bug
# conocido en _dynamo/polyfills/sys.py (Duplicate dispatch rule para sys.intern)
# que rompe el import de transformers.flex_attention.
RUN uv pip install \
    torch==2.10.0 torchaudio==2.10.0 \
    --index-url https://download.pytorch.org/whl/cu128

# qwentts-cpp-python (extra ggml de faster-qwen3-tts) — no está en PyPI para
# nuestra CUDA, se instala del wheelhouse de HF.
RUN uv pip install \
    "https://huggingface.co/datasets/andito/qwentts-cpp-python-wheels/resolve/main/whl/cu128/qwentts_cpp_python-0.3.0+cu128-py3-none-manylinux_2_35_x86_64.whl"

# speech-to-speech desde main.
RUN uv pip install \
    "speech-to-speech @ git+https://github.com/huggingface/speech-to-speech.git@main"

RUN uv pip install num2words

COPY --chown=user:user entrypoint.sh /home/user/app/entrypoint.sh
RUN chmod +x /home/user/app/entrypoint.sh

EXPOSE 8765

CMD ["/home/user/app/entrypoint.sh"]