Dockerfile

# F5-TTS Server — production-ready container.
# Two stages: builder pulls Python deps, runtime ships minimal CUDA + the app.
# For CPU-only or Apple Silicon, see docker-compose.cpu.yml or build with
# `--build-arg BASE_IMAGE=python:3.12-slim` (note: CPU inference RTF > 1 in tests).

ARG BASE_IMAGE=nvidia/cuda:12.8.0-runtime-ubuntu24.04
FROM ${BASE_IMAGE}

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

RUN apt-get update && apt-get install -y --no-install-recommends \
        python3.12 python3.12-venv python3-pip \
        libsndfile1 ffmpeg curl ca-certificates \
    && rm -rf /var/lib/apt/lists/*

RUN ln -sf /usr/bin/python3.12 /usr/bin/python3

WORKDIR /app

# Pin versions to match agoracosmica.org production. F5-TTS itself is pinned to
# 1.1.5 — that's where the F5TTS_Base config trap was verified safe.
COPY requirements.txt /app/requirements.txt
RUN pip3 install --no-cache-dir --break-system-packages -r /app/requirements.txt

COPY serve.py voices.json /app/
COPY refs/ /app/refs/

ENV MODELS_DIR=/models \
    REF_DIR=/app/refs \
    VOICES_FILE=/app/voices.json \
    PORT=8000 \
    HOST=0.0.0.0 \
    F5_DEFAULT_CKPT=hvoss \
    F5_DEFAULT_LANGUAGE=de \
    F5_SPEED=1.0 \
    OMP_NUM_THREADS=4 \
    MKL_NUM_THREADS=4 \
    CORS_ALLOW_ORIGINS=*

# Checkpoint cache. Mount a host volume here to persist across `docker compose down`.
VOLUME /models

EXPOSE 8000

HEALTHCHECK --interval=30s --timeout=10s --start-period=120s --retries=3 \
    CMD curl -sf http://localhost:8000/health || exit 1

CMD ["python3", "serve.py"]