MLCommons seeks to simplify sharing ML models with MLCube

No standardized way of packaging ML models

Model creators use unique and disparate methods for packaging their models with custom scripts and configuration settings. Due to lack of standardization of model packaging many hours are spent simply trying to get models run on a new system. 

MLCube provides a way to create and specify the information needed to run the model. An MLCube can be constructed by any model creator who follows a simple checklist for packaging the model into a shareable container with the standardized interface common to all MLCubes. MLCommons provides tools to help validate and run these packages.

Difficulties in sharing models at both the training and inference steps of the model lifecycle

MLCube is not a siloed solution, it is designed as an ecosystem enabler that helps remove friction and share models broadly. MLCubes can be used for training or inference across multiple platforms: Linux, Kubernetes, and public clouds. 

The Best Practices Working Group provides reference runners - simple software to help with executing MLCubes, but it also encourages third parties to adopt MLCube for use with more powerful infrastructure. 

Red Hat is working with a number of MLCommons members to further develop the MLCube concept. Specifically, we rely on Podman and Red Hat Universal Base Image (UBI) as  foundational OCI-compliant components that are used to create an MLCube. 

Podman is designed to run, build, share and deploy applications using OCI container images. A great advantage of Podman vs. other container runtimes, such as Docker, is it does not require root privileges to run containers, making it a safer and less error-prone tool. 

UBI is based on Red Hat Enterprise Linux (RHEL) and allows you to build, share and collaborate on your containerized application where you want. Model creators can choose to use freely distributable UBI as the base layer of their OCI containers and use Podman to run them, making these enterprise-grade software technologies available to the users of their MLCubes.

Red Hat has a history of contributing and shaping open source solutions that benefit the entire industry. Through our involvement in the best practices working group in general, and in the development of the MLCube concept in particular, Red Hat is aiming to provide a complete and effective solution that could greatly improve the adoption of many ML applications. 

Stay tuned for future updates on the progress MLCommons is making in democratizing access to AI and machine learning for everyone.