自定义CUDA加速OpenCV的Pip Wheel构建失效问题排查与优化请求

我碰到了一个棘手的问题：我编写了一个Dockerfile，从源码编译带有CUDA加速的OpenCV，编译过程本身是成功的，但pip就是不识别这个自定义安装的版本。结果后来我安装依赖OpenCV的Python包时，pip直接拉取了预编译的opencv-python wheel，直接把我源码编译的版本给覆盖（或者说屏蔽）了。

我看到有建议说，正确的解决方法是把自定义编译的OpenCV打包成wheel再安装，这样pip就会认为依赖已经满足。我照着这个方法试了，wheel构建脚本跑完没有报错，但安装后的wheel完全不能用。能不能帮我检查一下这个wheel构建脚本或者Dockerfile，告诉我哪些地方需要调整或者优化？

我的Dockerfile内容 #

# syntax=docker/dockerfile:1
# Requires Docker BuildKit for cache mounts (build with DOCKER_BUILDKIT=1)
FROM nvidia/cuda:12.1.0-cudnn8-devel-ubuntu22.04

# Environment setup
ENV DEBIAN_FRONTEND=noninteractive \
    LANG=en_US.UTF-8 \
    LC_ALL=en_US.UTF-8 \
    PYTHONIOENCODING=UTF-8 \
    NVIDIA_DRIVER_CAPABILITIES=all \
    CUDA_HOME=/usr/local/cuda \
    PATH=/usr/local/cuda/bin:$PATH \
    LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH \
    CCACHE_DIR=/ccache

# Create ccache directory
RUN mkdir -p /ccache

# Set locale
RUN apt-get update && apt-get install -y --no-install-recommends \
    locales \
    && locale-gen en_US.UTF-8 \
    && update-locale LANG=en_US.UTF-8

# Essential system dependencies
RUN apt-get update && apt-get install -y --no-install-recommends \
    # Build tools
    build-essential \
    cmake \
    pkg-config \
    ninja-build \
    ccache \
    # GCC-10 for CUDA 12.x compatibility
    gcc-10 \
    g++-10 \
    # Version control
    git \
    # Utilities
    wget \
    curl \
    unzip \
    vim \
    nano \
    htop \
    software-properties-common \
    && add-apt-repository ppa:deadsnakes/ppa -y \
    && apt-get update \
    && apt-get install -y --no-install-recommends \
    python3.9 \
    python3.9-dev \
    python3.9-venv \
    python3.9-distutils \
    python3-pip \
    # OpenCV dependencies
    libgtk-3-dev \
    libgtk2.0-dev \
    libavcodec-dev \
    libavformat-dev \
    libswscale-dev \
    libswresample-dev \
    libtbb-dev \
    libjpeg-dev \
    libpng-dev \
    libtiff-dev \
    libwebp-dev \
    libopenexr-dev \
    libdc1394-dev \
    libgstreamer1.0-dev \
    libgstreamer-plugins-base1.0-dev \
    libv4l-dev \
    libxvidcore-dev \
    libx264-dev \
    libfdk-aac-dev \
    libmp3lame-dev \
    libtheora-dev \
    libvorbis-dev \
    libxine2-dev \
    libopencore-amrnb-dev \
    libopencore-amrwb-dev \
    # Math libraries
    libopenblas-dev \
    liblapack-dev \
    libatlas-base-dev \
    libeigen3-dev \
    libhdf5-dev \
    # OpenGL support
    libgl1-mesa-glx \
    libglu1-mesa-dev \
    libglew-dev \
    # Qt5 for OpenCV GUI
    qtbase5-dev \
    qtchooser \
    qt5-qmake \
    qtbase5-dev-tools \
    # Additional utilities
    libprotobuf-dev \
    protobuf-compiler \
    && apt-get clean \
    && rm -rf /var/lib/apt/lists/*

# Set Python 3.9 as default
RUN update-alternatives --install /usr/bin/python3 python3 /usr/bin/python3.9 1 \
    && update-alternatives --install /usr/bin/python python /usr/bin/python3.9 1 \
    && python3.9 -m pip install --upgrade pip setuptools wheel

# Install numpy first (required for OpenCV Python bindings)
# Also install wheel building tools
RUN python3 -m pip install --no-cache-dir \
    numpy==1.26.4 \
    wheel \
    setuptools \
    build \
    auditwheel \
    patchelf

# Remove any pre-existing OpenCV packages that might conflict
RUN pip3 uninstall -y opencv-python opencv-python-headless opencv-contrib-python opencv-contrib-python-headless || true

# Set GCC-10 as default for OpenCV build
RUN update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-10 100 \
    && update-alternatives --install /usr/bin/g++ g++ /usr/bin/g++-10 100

# Configure ccache
RUN ccache --set-config=cache_dir=/ccache \
    && ccache --set-config=max_size=15G \
    && ccache --set-config=compression=true \
    && ccache --set-config=compression_level=6

# OpenCV 4.8.0 with CUDA support and NONFREE modules
ARG OPENCV_VERSION=4.8.0
ARG CUDA_ARCH_BIN="7.5;8.0;8.6;8.9"

# opencv and opencv-contrib :
# including NONFREE code -could be used or not-
# Use BuildKit cache mount for ccache to speed up rebuilds
RUN --mount=type=cache,target=/ccache \
    cd /opt/ &&\
    wget https://github.com/opencv/opencv/archive/${OPENCV_VERSION}.zip -O opencv.zip &&\
    unzip -qq opencv.zip &&\
    rm opencv.zip &&\
    wget https://github.com/opencv/opencv_contrib/archive/${OPENCV_VERSION}.zip -O opencv-co.zip &&\
    unzip -qq opencv-co.zip &&\
    rm opencv-co.zip &&\
    mkdir /opt/opencv-${OPENCV_VERSION}/build && cd /opt/opencv-${OPENCV_VERSION}/build &&\
    # Configure OpenCV with proper CUDA linking:\
    # - Use stubs for linking but set RPATH to real CUDA libs for runtime\
    # - This prevents runtime failures from stub library references\
    cmake \
        -D BUILD_opencv_java=OFF \
        -D WITH_CUDA=ON \
        -D BUILD_opencv_dnn=ON \
        -D CUDA_ARCH_BIN="${CUDA_ARCH_BIN}" \
        -D WITH_CUBLAS=ON \
        -D WITH_CUDNN=ON \
        -D OPENCV_DNN_CUDA=ON \
        -D ENABLE_FA

内容来源于stack exchange