nvidia docker Cannot load libnvcuvid.so.1

问题#

使用docker对runtime进行封装，程序需要用到GPU且在host正常运行,在容器内则报错:

Cannot load libnvcuvid.so.1
[hevc_cuvid @ 0x559da3fbd80] Failed loading nvcuvid.
terminate called after throwing an instance of 'std::runtime_error'
  what(): failed to open avcodec
Aborted (core dumped)

排查#

第一反应ldd查看依赖，程序并不直接依赖libnvcuvid.so.1
host存在/usr/lib/x86_64-linux-gnu/libnvcuvid.so.1
容器内部却找不到库文件，已包含docker run --gpus all
容器执行nvidia-smi能够正常显示GPU调用CUDA也正常

分析#

既然CUDA正常排除container-toolkit问题，深挖应用节点依赖的avcodec需要用到显卡codec
默认不开启codec支持，Driver Capability列表如下

Driver Capability	Description
`compute`	required for CUDA and OpenCL applications.
`compat32`	required for running 32-bit applications.
`graphics`	required for running OpenGL and Vulkan applications.
`utility`	required for using nvidia-smi and NVML.
`video`	required for using the Video Codec SDK.
`display`	required for leveraging X11 display.
empty or unset	default driver capability: `utility`, `compute`

解决#

$ docker run --rm --runtime=nvidia -e NVIDIA_DRIVER_CAPABILITIES=compute,video,utility nvidia/cuda:12.0.1-cudnn8-runtime-ubuntu18.04 ldconfig -p | grep libnvcuvid
	libnvcuvid.so.1 (libc6,x86-64) => /usr/lib/x86_64-linux-gnu/libnvcuvid.so.1

$ docker run --rm --gpus 'all,"capabilities=compute,video,utility"' nvidia/cuda:12.0.1-cudnn8-runtime-ubuntu18.04 ldconfig -p | grep libnvcuvid
	libnvcuvid.so.1 (libc6,x86-64) => /usr/lib/x86_64-linux-gnu/libnvcuvid.so.1