Dockerfile

Hi ,
is there anyway to find latest dockerfile or can NVIDIA open Dockerfile source ?
This can help us to run a same environment prior run it on AWS, since once VM created , it start to charge money, and i’d like to verify same docker environment with local build of Dockerfile firstly.

Chester

The NGC containers can be pulled and run on workstations and servers with Pascal and Volta GPUs, assuming you have Docker and the nvidia-docker utility installed on those systems. However, support for these systems is best effort only.

If you have access to one of these systems locally, you can pull the containers from NGC and test your application before deploying to AWS.

Please keep in mind that since the containers pulled from NGC include licensed software, they cannot be redistrubuted.

No one is going to redistribute NVIDIA software, but just reuse socket file so we can optimization on docker if needs.

Also from docker hub image of CUDA, CUDNN, it have docker file release already1.

https://gitlab.com/nvidia/cuda/blob/ubuntu16.04/9.0/base/Dockerfile

The images we provide on Dockerhub are different than those on NGC.

We have no plans to publish the Dockerfiles used to build NGC images - they would be of limited utility outside NVIDIA since they reference internal package repositories that could not be accessed from outside.

To modify the NGC images, NVIDIA recommends building on top of them (using the FROM directive in your Dockerfile) rather than rebuilding from scratch; there are example derivative Dockerfiles inside each image in the /workspace/docker-examples directory. This not only avoids a full rebuild but it also allows customizations to be now easily carried forward from one month’s release to the next.

Thanks for your interest in NGC!

We are not using docker but singularity in our lab. As as sysadmin, I am allowed to convert the NGC docker into singularity squashfs format for our research group usage or does each user need to to it by themself?

The NGC containers require the nvidia-docker utility to run them. Just converting the contents of NGC containers will not result in a functional alternative.

A description of nvidia-docker, along with how to install and use it, can be found at nvidia-docker/README.md at master · NVIDIA/nvidia-docker · GitHub.

If singularity can perfectly:

singularity shell --nv docker://tensorflow/tensorflow:latest-gpu-py3

I image I can just try to use the NGC registry as it is provided.