Docker离线安装Nvidia-container-toolkit实现容器内GPU调用

背景

需求：实验室内通过Docker搭建隔离环境保证各用户数据安全和服务器环境安全，防止软件环境混杂造成莫名其妙的bug，容器内需要能够调用显卡资源。

预先准备

本文的内容基于以下软件版本：
Docker：Docker version 20.10.17, build 100c701
CUDA：NVIDIA-SMI 510.68.02 Driver Version: 510.68.02 CUDA Version: 11.6
系统：Ubuntu 20.04.4 LTS

Nvidia-container-toolkit架构

Nvidia官网架构概述，可以用Chrome自带网页翻译仔细阅读，本文仅简单介绍。

架构

The NVIDIA container stack is architected so that it can be targeted to support any container runtime in the ecosystem.
The components of the stack include:
	The nvidia-docker wrapper
	The NVIDIA Container Runtime (nvidia-container-runtime)
	The NVIDIA Container Runtime Hook (nvidia-container-toolkit / nvidia-container-runtime-hook)
	The NVIDIA Container Library and CLI (libnvidia-container1, nvidia-container-cli)
...
Note that as of version 3.6.0, the nvidia-container-runtime package is a meta package that 
only depends on the nvidia-container-toolkit package and does not provide any functionality of itself.

NVIDIA container主要组件包括nvidia-container-runtime, nvidia-container-toolkit, libnvidia-container和CUDA驱动；
在3.6.0版本后，runtime包成为一个只依赖于toolkit包（指container-toolkit而不是nvidia CUDA toolkit）的包，在官网中也指出&#