The most recent Tao Toolkit Quick Start Guide (TAO Toolkit Quick Start Guide - NVIDIA Docs) mentions a requirement for Python >=3.6.9<3.7 and gives an example of creating a conda configuration using python=3.6
When following these directions using the NVIDIA Deep Learning AMI 23.03.0 on an EC2 instance, the Tao Toolkit Instance configures with the 4.0 docker images (e.g., 4.0.0-tf1.15.5).
If I configure conda with a Python 3.8 environment, the quickstart_launcher.sh install scripts pull the 5.0 version of Tao (as expected).
Is the documentation in the Quickstart Guide out of date, or should the quickstart_launcher install the Tao Configuration for 5.0 with a Python 3.6 environment?
When python version >= 3.6.9, it is recommended to set up a python environment using miniconda. The conda environment with python 3.6 version is verified. If conda environment with python 3.8 version may also work which is not mentioned in the guide.
More info can also be found in Mismatch in python environment on AWS EC2 image - #9 by tom_s.
Thanks @Morganh – it looks like nvidia-docker2 was missing from the NVIDIA Deep Learning AMI 23.03.0.
Once I installed that, running bash setup/quickstart_launcher.sh --install from the 5.0.0 Getting Started package on a miniconda env with Python=3.6 resulted in the correct version of the Tao Toolkit being installed.