Dockerfile

# thecollabagepatch/magenta:latest - optimized
FROM nvidia/cuda:12.6.2-cudnn-runtime-ubuntu22.04

# CUDA libs present + on loader path
RUN apt-get update && apt-get install -y --no-install-recommends \
    cuda-libraries-12-4 && rm -rf /var/lib/apt/lists/*
ENV LD_LIBRARY_PATH=/usr/local/cuda/lib64:/usr/local/cuda-12.4/lib64:/usr/local/cuda-12.4/compat:/usr/local/cuda/targets/x86_64-linux/lib:${LD_LIBRARY_PATH}
RUN ln -sf /usr/local/cuda/targets/x86_64-linux/lib /usr/local/cuda/lib64 || true

# Ensure the NVIDIA repo key is present (non-interactive) and install cuDNN 9.8
RUN set -eux; \
  apt-get update && apt-get install -y --no-install-recommends gnupg ca-certificates curl; \
  install -d -m 0755 /usr/share/keyrings; \
  curl -fsSL https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/3bf863cc.pub \
    | gpg --batch --yes --dearmor -o /usr/share/keyrings/cuda-archive-keyring.gpg; \
  apt-get update; \
  apt-mark unhold libcudnn9-cuda-12 || true; \
  apt-get install -y --no-install-recommends \
      'libcudnn9-cuda-12=9.8.*' \
      'libcudnn9-dev-cuda-12=9.8.*' \
      --allow-downgrades --allow-change-held-packages; \
  apt-mark hold libcudnn9-cuda-12 || true; \
  ldconfig; \
  rm -rf /var/lib/apt/lists/*

# Optional preload workaround if still needed
ENV LD_PRELOAD=/usr/local/cuda/lib64/libcusparse.so.12:/usr/local/cuda/lib64/libcublas.so.12:/usr/local/cuda/lib64/libcublasLt.so.12:/usr/local/cuda/lib64/libcufft.so.11:/usr/local/cuda/lib64/libcusolver.so.11

# Better allocator (less fragmentation than BFC during XLA autotune)
ENV TF_GPU_ALLOCATOR=cuda_malloc_async

# Let cuBLAS use TF32 fast path on Ada (L40S) for big GEMMs
ENV TF_ENABLE_CUBLAS_TF32=1 NVIDIA_TF32_OVERRIDE=1

ENV DEBIAN_FRONTEND=noninteractive \
    PYTHONUNBUFFERED=1 \
    PIP_NO_CACHE_DIR=1 \
    TF_FORCE_GPU_ALLOW_GROWTH=true \
    XLA_PYTHON_CLIENT_PREALLOCATE=false

ENV JAX_PLATFORMS=""

# --- OS dependencies ---
RUN apt-get update && apt-get install -y --no-install-recommends \
    software-properties-common curl ca-certificates git \
    libsndfile1 ffmpeg \
    build-essential pkg-config \
    && add-apt-repository ppa:deadsnakes/ppa -y \
    && apt-get update && apt-get install -y --no-install-recommends \
    python3.11 python3.11-venv python3.11-distutils python3-pip \
    && rm -rf /var/lib/apt/lists/*

# Make python3 => 3.11 for convenience
RUN ln -sf /usr/bin/python3.11 /usr/bin/python && python -m pip install --upgrade pip

# --- Python dependencies (install order matters!) ---

# 1) Install JAX first
RUN python -m pip install "jax[cuda12]==0.7.1" "jaxlib==0.7.1"

# 2) Install seqio dependencies manually (excluding tfds-nightly and tensorflow-text)
RUN python -m pip install \
    absl-py clu sentencepiece

# 3) Install seqio without dependencies
RUN python -m pip install --no-deps "seqio==0.0.11"

# # 4) Create a patch file and apply it to seqio
# RUN echo 'Patching seqio/vocabularies.py...' && \
#     VOCAB_FILE="/usr/local/lib/python3.11/dist-packages/seqio/vocabularies.py" && \
#     cp "$VOCAB_FILE" "${VOCAB_FILE}.bak" && \
#     cat "$VOCAB_FILE" | sed 's/^import tensorflow_text as tf_text$/# import tensorflow_text as tf_text  # Patched for Magenta RT/' > "${VOCAB_FILE}.tmp" && \
#     mv "${VOCAB_FILE}.tmp" "$VOCAB_FILE" && \
#     echo "Verifying patch..." && \
#     grep "# import tensorflow_text" "$VOCAB_FILE" && \
#     echo "Patch successful!" || (echo "Patch failed!" && exit 1) Patch seqio to remove tensorflow-text dependency


# 5) Install correct protobuf version for stable TensorFlow
RUN python -m pip install "protobuf>=3.20.3,<6.0.0"

# 6) Install stable TensorFlow (not nightly)
RUN python -m pip install --no-deps \
    "tensorflow[and-cuda]>=2.18.0" \
    "tensorflow-hub==0.16.1"

# 6) Install remaining packages (consolidated, no duplicates)
RUN python -m pip install \
    tf2jax gin-config librosa resampy soundfile \
    fastapi uvicorn[standard] python-multipart pyloudnorm \
    google-cloud-storage "numpy==2.1.3" \
    huggingface_hub gradio soxr cached-property

# 7) Install airio
RUN python -m pip install 'git+https://github.com/google/airio.git@main'

RUN python -c "\
file_path = '/usr/local/lib/python3.11/dist-packages/seqio/vocabularies.py'; \
content = open(file_path, 'r').read(); \
content = content.replace('import tensorflow_text as tf_text', '# import tensorflow_text as tf_text'); \
open(file_path, 'w').write(content); \
print('Patched seqio/vocabularies.py')"

# 8) Install t5x and flaxformer without dependencies
RUN python -m pip install --no-deps \
    "t5x @ git+https://github.com/google-research/t5x.git@92c5b46" \
    "flaxformer @ git+https://github.com/google/flaxformer@399ea3a"

RUN python -m pip install aqtp fiddle
    

# 9) Install magenta-rt without dependencies
RUN python -m pip install --no-deps \
    'git+https://github.com/magenta/magenta-realtime#egg=magenta_rt[gpu]'

# 10) Remove conflicting stable tensorflow packages
RUN python -m pip uninstall -y tensorflow tensorflow-cpu tensorflow-text || true

# Install stable TensorFlow instead of nightly (following KEHANG's recommendation)
RUN python -m pip install --upgrade \
    "tensorflow[and-cuda]>=2.18.0"

# 11) Verify TensorFlow installation
RUN python -c "import tensorflow as tf; print('='*50); print('TensorFlow version:', tf.__version__); print('='*50)"

# Switch to non-root user for security
RUN useradd -m -u 1000 appuser
WORKDIR /home/appuser/app

# Copy application files
COPY --chown=appuser:appuser app.py /home/appuser/app/app.py
COPY --chown=appuser:appuser utils.py /home/appuser/app/utils.py
COPY --chown=appuser:appuser jam_worker.py /home/appuser/app/jam_worker.py
COPY --chown=appuser:appuser one_shot_generation.py /home/appuser/app/one_shot_generation.py
COPY --chown=appuser:appuser model_management.py /home/appuser/app/model_management.py
COPY --chown=appuser:appuser documentation.html /home/appuser/app/documentation.html
COPY --chown=appuser:appuser lil_demo_540p.mp4 /home/appuser/app/lil_demo_540p.mp4
COPY --chown=appuser:appuser magentaRT_rt_tester.html /home/appuser/app/magentaRT_rt_tester.html
COPY --chown=appuser:appuser magenta_prompts.js /home/appuser/app/magenta_prompts.js
COPY --chown=appuser:appuser docs/ /home/appuser/app/docs/

USER appuser

EXPOSE 7860
CMD ["bash", "-lc", "python -m uvicorn app:app --host 0.0.0.0 --port ${PORT:-7860}"]