readme_docker.md

The Docker Base Image and Dockerfile

For an initial test, you can build FROM python:3.8 and have torch as a requirement.

However, please note that should you encounter other errors in production, see below for information on other images you might try to use.

While pytorch has a docker hub, and nvidia does too, nvidia also hosts pytorch- and cuda-ready images here (available via e.g., docker pull nvcr.io/nvidia/pytorch:21.05-py3).

If your project requires something more advanced, you'll want to choose the latter, Nvidia's Pytorch image, for the following reasons/errors/issues:

• Using FROM python:3.8 or 3.9 or similar runs on AWS instances (including the p and g families with GPUs), but may be missing CUDA-specific infrastructure for advanced pytorch calculations
• Using only the pytorch image in a Dockerfile via FROM pytorch/pytorch:latest won't have the necessary nvidia infrastructure, so if you need something more advanced than the base python we recommend choosing the nvidia image if pairing with g4dn.
• Using Nvidia's base image e.g., FROM nvidia/cuda:11.3.0-cudnn8-runtime-centos8 doesn't match the AWS GPU instances
• Of note, the nvidia GPU instance g4dn we chose from here requires environment variables for CUDA to work. It turns out that some of these come with the image and some don't.

TL;DR, make note of your preferred values for NVIDIA_REQUIRE_CUDA and NVIDIA_DRIVER_CAPABILITES environment variables.