cuda.Dockerfile

FROM nvidia/cuda:12.8.1-cudnn-devel-ubuntu24.04
ENV CUDA_PATH=/usr/local/cuda \
    CUDA_HOME=/usr/local/cuda \
    CUDA_TOOLKIT_ROOT_DIR=/usr/local/cuda \
    LD_LIBRARY_PATH=/opt/venv/lib/python3.12/site-packages/nvidia/cudnn/lib:/usr/local/cuda/lib64:$LD_LIBRARY_PATH \
    VIRTUAL_ENV=/opt/venv \
    PATH="/opt/venv/bin:/root/.local/bin:$PATH"
RUN apt-get update --fix-missing && \
    apt-get upgrade -y && \
    apt-get install -y --no-install-recommends \
       git build-essential cmake gcc g++ ninja-build \
       portaudio19-dev ffmpeg libportaudio2 libasound-dev \
       wget ocl-icd-opencl-dev opencl-headers \
       clinfo libclblast-dev libopenblas-dev unzip curl && \
    mkdir -p /etc/OpenCL/vendors && \
    echo "libnvidia-opencl.so.1" > /etc/OpenCL/vendors/nvidia.icd && \
    apt-get clean && \
    rm -rf /var/lib/apt/lists/* /var/cache/apt/* /tmp/* /var/tmp/*
# Install uv and create venv with Python 3.12
RUN curl -LsSf https://astral.sh/uv/install.sh | sh && \
    /root/.local/bin/uv venv /opt/venv --python 3.12
WORKDIR /app
# Use PyTorch 2.9.1 which is built against cuDNN 9.10.2
# Install nvidia-cudnn-cu12==9.10.2.21 to get matching cuDNN libraries (overrides system cuDNN 9.8.0)
RUN uv pip install torch==2.9.1+cu128 torchaudio==2.9.1+cu128 --index-url https://download.pytorch.org/whl/cu128 && \
    uv pip install nvidia-cudnn-cu12==9.10.2.21
# Install numpy and Cython for pkuseg (required by chatterbox-tts)
# numpy>=1.26.0 required for Python 3.12 compatibility
RUN uv pip install "numpy>=1.26.0" Cython "setuptools>=78.1.1"
# Install pkuseg separately (required by chatterbox-tts)
RUN uv pip install pkuseg==0.0.25 --no-build-isolation
COPY cuda-requirements.txt .
RUN uv pip install -r cuda-requirements.txt
# Install chatterbox-tts with --no-deps to bypass transformers==4.46.3 pin
# This allows us to use transformers>=4.53.0 for security fixes
RUN uv pip install chatterbox-tts --no-deps
ENV HOST=0.0.0.0 \
    CUDA_DOCKER_ARCH=all \
    CUDAVER=12.8.1
# Install xllamacpp with CUDA 12.8 support (compatible with CUDA 12.9)
RUN uv pip install xllamacpp==0.2.12 --reinstall --index-url https://xorbitsai.github.io/xllamacpp/whl/cu128
COPY . .
EXPOSE 8091
# Use start.py which runs precache once, then starts uvicorn workers
CMD ["python", "start.py"]