Skip to content

Latest commit

 

History

History
 
 

ngc

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
pytorch
pytorch
 
 
rapids
rapids
 
 
tensorflow
tensorflow
 
 
README.md
README.md
 
 
build.sh
build.sh
 
 

README.md

NGC-based Docker Containers

This directory contains several Dockerfiles for common Deep Learning and Machine Learning frameworks. These serve as an example for extending NGC Docker images.

By default, most NGC containers start a shell prompt. This is perfect for local development, but has limitations when deploying containers through an AI orchestrator (such as Kubernetes) or AI workflow platform (such as Kubeflow). The Dockerfiles in this folder are designed to start JupyterLab by default, allowing for easier integration into interactive notebook platforms such as Kubeflow and other DGX Ready Software Partners.

These examples are all based on NGC containers and provide a Dockerfile-minimal and Dockerfile version. The minimal version has the minimal changes required to deploy a NGC container in Kubeflow. The non-minimal version has a curated set of libraries and packages included that are useful for demos, monitoring, generating graphs, Kubernetes integration, and rendering interactive html5 elements.

Where possible, the package versions in these Dockerfiles are hard-coded to ensure repeatable builds and consistent execution of the included examples and tutorials.

Additional details about NGC can be found here.

Additional details about the Deep Learning Examples can be found here.

Building Docker Images

To build all example images run:

./build.sh minimal # For minimal images only
./build.sh # For all images

The full images can be built individually by running:

cd tensorflow
docker build -t tensorflow:deepops-kubeflow .

The minimal images can be built individually by running:

cd rapids
docker build -t rapids:deepops-kubeflow-minimal -f Dockerfile-minimal .

After building the docker images locally, they can be pushed to a private Docker registry.

Note: It is a violation of the NGC EULA to host a Docker image based off an NGC image in a public registry. All examples included in this directory should only be hosted in private registries, such as those provided by NGC.

Kubeflow Integration

As of v1.2.0, Kubeflow requires containers to start Jupyter, VSCode, or another web application in order to deploy through Kubeflow Notebooks.

A bare-minimum Dockerfile must define the FROM, ENTRYPOINT, and CMD, similar to below. Note that the base_url must be defined for Kubeflow to redirect properly.

FROM nvcr.io/nvidia/pytorch:20.12-py3

ENTRYPOINT ["/bin/sh"]

CMD ["-c", "jupyter lab  --notebook-dir=/workspace --ip=0.0.0.0 --no-browser --allow-root --port=8888 --NotebookApp.token='' --NotebookApp.password='' --NotebookApp.allow_origin='*' --NotebookApp.base_url=${NB_PREFIX}"]

As a quick-start, several pre-built Kubeflow containers with end-to-end AI workflow tutorial/demo containers can be found on NGC here.

Misc. Notes

  • JupyterLab is exposed at port 8888
  • Tensorboard is exposed at port 6006
  • The RAPIDS container has a different jupyter startup command due to it's use of Conda.
  • The WORKDIR contains the built-in NGC tutorials and example code. This is /workspace in most cases (/rapids for RAPIDS containers).

Included Frameworks