Dockerfile.server

# CUDA 13.1 + cuDNN 9 — matches node driver 590.48.01 on gx16 (A30, compute 8.0)
FROM pytorch/pytorch:2.7.0-cuda13.1-cudnn9-devel

ARG DEBIAN_FRONTEND=noninteractive

# Pinned together — torchaudio must share the same PyTorch ABI.
# Update both atomically whenever the base image is bumped.
ARG TORCH_VERSION=2.7.0
ARG CUDA_TAG=cu131

RUN apt-get update && \
    apt-get install -y --no-install-recommends \
        git \
        ffmpeg \
        build-essential \
    && rm -rf /var/lib/apt/lists/* && \
    apt-get clean

WORKDIR /app

ARG GIT_REPOSITORY="https://github.com/aihpi/F5-TTS"
# Hacky workaround, as bigvgan is not properly setup to be installed as a package:
# https://github.com/aihpi/F5-TTS?tab=readme-ov-file#2-local-editable-if-also-do-training-finetuning
ARG BIGVGAN_PATCH="import os\nimport sys\nsys.path.append(os.path.dirname(os.path.abspath(__file__)))\n"

RUN git clone $GIT_REPOSITORY \
    && cd F5-TTS \
    && git submodule update --init --recursive \
    && printf "$BIGVGAN_PATCH" | cat - src/third_party/BigVGAN/bigvgan.py > temp && mv temp src/third_party/BigVGAN/bigvgan.py \
    && pip install -e .[eval] \
    && pip install -r src/third_party/BigVGAN/requirements.txt

# Force-reinstall torch + torchaudio from the official CUDA 13.1 wheel index so they
# share the same ABI regardless of what pip resolved above.  This must be a separate
# RUN so Docker cache invalidation is scoped to version bumps, not source changes.
RUN pip install --no-cache-dir --force-reinstall \
    torch==${TORCH_VERSION} \
    torchaudio==${TORCH_VERSION} \
    --index-url https://download.pytorch.org/whl/${CUDA_TAG}

# for prototyping
COPY src/f5_tts/demo/ /app/F5-TTS/src/f5_tts/demo/

CMD ["python3", "-m", "uvicorn", "f5_tts.demo.server:app", "--host", "0.0.0.0", "--port", "8000"]