PaddleX/deploy/hps/server_env/Dockerfile

# 1. Should I use `syntax` parser directive?
# 2. Can I use `COPY --link` to optimize caching?
# 3. Create a non-root user for deployment?
# 4. Should I use APT repository mirrors?
# 5. Use cache mounts for `pip`?

ARG DEVICE_TYPE=gpu
ARG ENV_TYPE=prod
ARG PIP_INDEX_URL=https://pypi.org/simple


# GPU base
# nvcr.io/nvidia/cuda:11.8.0-devel-ubuntu20.04
# We use the `devel` version because the CUDA toolchain is needed for JIT.
# However, can we use a smaller base image by not including tools and
# libraries that are not required?
FROM nvcr.io/nvidia/cuda@sha256:03681bbd11fea044ff3e3a5d65e190a6f935af30c56e03e740b27e4563f61e0f AS gpu-base


# CPU base
# ubuntu:20.04
FROM ubuntu:20.04@sha256:8feb4d8ca5354def3d8fce243717141ce31e2c428701f6682bd2fafe15388214 AS cpu-base


# Base
FROM ${DEVICE_TYPE}-base AS base

ARG PIP_INDEX_URL

# Should I add `SHELL ["/bin/bash", "-c"]`?

ENV DEBIAN_FRONTEND=noninteractive

RUN mkdir /paddlex

RUN echo 'deb http://archive.ubuntu.com/ubuntu jammy main universe' > /etc/apt/sources.list.d/jammy-temp.list \
    && apt-get update \
    && apt-get install -y --no-install-recommends python3.10 python3.10-venv \
    && python3.10 -m venv /paddlex/py310 \
    && rm -rf /var/lib/apt/lists/*
ENV PATH="/paddlex/py310/bin:${PATH}"

ENV PIP_DISABLE_PIP_VERSION_CHECK=1
ENV PIP_NO_CACHE_DIR=0
ENV PYTHONUNBUFFERED=1
ENV PYTHONDONTWRITEBYTECODE=1
ENV PIP_INDEX_URL=${PIP_INDEX_URL}


# Requirement collection
FROM base AS rc
RUN python -m pip install pip-tools==7.4.1

# `ENTRYPOINT` and `CMD`?


# Build stage 1
FROM base AS build1

RUN apt-get update \
    && apt-get install -y --no-install-recommends wget unzip \
    && rm -rf /var/lib/apt/lists/*

RUN mkdir -p /paddlex/libs 


# Build stage 2 for GPU
FROM build1 AS gpu-build2

RUN mkdir -p /paddlex/tensorrt

RUN --mount=type=bind,source=deploy/hps/server_env/cudnn-linux-x86_64-8.9.7.29_cuda11-archive.tar.xz,target=/tmp/cudnn-linux-x86_64-8.9.7.29_cuda11-archive.tar.xz \
    tar -xf /tmp/cudnn-linux-x86_64-8.9.7.29_cuda11-archive.tar.xz -C /paddlex \
    && mv /paddlex/cudnn-linux-x86_64-8.9.7.29_cuda11-archive/lib/* /paddlex/libs/

RUN --mount=type=bind,source=deploy/hps/server_env/TensorRT-8.6.1.6.Linux.x86_64-gnu.cuda-11.8.tar.gz,target=/tmp/TensorRT-8.6.1.6.Linux.x86_64-gnu.cuda-11.8.tar.gz \
    tar -xf /tmp/TensorRT-8.6.1.6.Linux.x86_64-gnu.cuda-11.8.tar.gz -C /paddlex \
    && mv /paddlex/TensorRT-8.6.1.6 /paddlex/tensorrt \
    && rm -rf /paddlex/tensorrt/data /paddlex/tensorrt/doc /paddlex/tensorrt/python /paddlex/tensorrt/samples


# Build Stage 2 for CPU
FROM build1 AS cpu-build2


# Build Stage 3
FROM ${DEVICE_TYPE}-build2 AS build3
ARG DEVICE_TYPE
ARG ENV_TYPE

RUN wget -P /paddlex "https://paddle-model-ecology.bj.bcebos.com/paddlex/PaddleX3.0/deploy/deps/tritonserver/tritonserver-2.15.0-${DEVICE_TYPE}.zip" \
    && unzip "/paddlex/tritonserver-2.15.0-${DEVICE_TYPE}.zip" -d /paddlex \
    && mv "/paddlex/tritonserver-2.15.0-${DEVICE_TYPE}" /paddlex/tritonserver \
    && rm "/paddlex/tritonserver-2.15.0-${DEVICE_TYPE}.zip"

RUN --mount=type=bind,source=deploy/hps/server_env/requirements/${DEVICE_TYPE}.txt,target=/tmp/requirements.txt \
    --mount=type=bind,source=deploy/hps/server_env/requirements/${DEVICE_TYPE}_hpi.txt,target=/tmp/hpi_requirements.txt \
    --mount=type=bind,source=deploy/hps/server_env/requirements/${DEVICE_TYPE}_dev.txt,target=/tmp/dev_requirements.txt \
    python -m pip install --requirement /tmp/requirements.txt --requirement /tmp/hpi_requirements.txt \
    && if [ "${ENV_TYPE}" = 'dev' ]; then \
        python -m pip install --requirement /tmp/dev_requirements.txt; \
    fi \
    && python -m pip install https://paddle-whl.bj.bcebos.com/nightly/cu126/safetensors/safetensors-0.6.2.dev0-cp38-abi3-linux_x86_64.whl

RUN --mount=type=bind,source=.,target=/tmp/PaddleX,rw \
    python -m pip install --no-deps /tmp/PaddleX
RUN --mount=type=bind,source=deploy/hps/server_env/paddlex-hps-server,target=/tmp/paddlex-hps-server,rw \
    python -m pip install --no-deps /tmp/paddlex-hps-server


# GPU Deployment
FROM base AS gpu-deployment

RUN apt-get update \
    && apt-get install -y --no-install-recommends datacenter-gpu-manager \
    && rm -rf /var/lib/apt/lists/*

COPY --from=build3 --chown=paddlex:paddlex /paddlex/tensorrt /paddlex/tensorrt
ENV LD_LIBRARY_PATH="/paddlex/tensorrt/lib:${LD_LIBRARY_PATH}"

RUN mkdir /paddlex/libs
RUN ln -s /usr/local/cuda-11.8/targets/x86_64-linux/lib/libcublas.so.11 /paddlex/libs/libcublas.so \
    && ln -s /usr/local/cuda-11.8/targets/x86_64-linux/lib/libcufft.so.10 /paddlex/libs/libcufft.so \
    && ln -s /usr/local/cuda-11.8/targets/x86_64-linux/lib/libcurand.so.10 /paddlex/libs/libcurand.so \
    && ln -s /usr/local/cuda-11.8/targets/x86_64-linux/lib/libcusolver.so.11 /paddlex/libs/libcusolver.so \
    && ln -s /usr/local/cuda-11.8/targets/x86_64-linux/lib/libcusparse.so.11 /paddlex/libs/libcusparse.so


# CPU Deployment
FROM base AS cpu-deployment


# Deployment
FROM ${DEVICE_TYPE}-deployment AS deployment

# Should we use an environment variable to store the version?
RUN groupadd -g 1000 paddlex \
    && useradd -m -s /bin/bash -u 1000 -g 1000 paddlex

# Not sure if all these deps are necessary
RUN apt-get update \
    && apt-get install -y --no-install-recommends libre2-5 libssl1.1 libb64-0d libnuma1 libarchive13 python3.10-dev libgl1-mesa-glx libglib2.0-0 libgomp1 ccache binutils build-essential texlive texlive-latex-base texlive-latex-extra \
    && rm -rf /var/lib/apt/lists/*

COPY --from=build3 --chown=paddlex:paddlex /paddlex/libs /paddlex/libs

# Should I use `ldconfig`?
ENV LD_LIBRARY_PATH="/paddlex/libs:${LD_LIBRARY_PATH}"

COPY --from=build3 --chown=paddlex:paddlex /paddlex/tritonserver /opt/tritonserver
ENV PATH="/opt/tritonserver/bin:${PATH}"

# Is the merging safe?
COPY --from=build3 --chown=paddlex:paddlex /paddlex/py310 /paddlex/py310

RUN mkdir /paddlex/var \
    && chown paddlex:paddlex /paddlex/var

USER paddlex

USER root

CMD ["/bin/bash"]