Canonical, the distributor of Ubuntu, has launched version 1.7 of its Charmed Kubeflow MLOps platform. The new release allows companies to operate machine learning models on Kubernetes as event-driven serverless applications, thanks to the integration of Knative. In addition, the new version provides a Custom Resource Definition (CRD) for inference and model serving with KServe. The revision of the user interface for Katib should also make it easier for data scientists to perform hyperparameter tuning in Charmed Kubeflow.
The goal of Charmed Kubeflow is to automate workflows for training, tuning, and providing ML models. It expands its serverless capabilities with the integration of Knative, which aims to relieve developers and data scientists of routine infrastructure tasks. This automation allows them to work with a preferred ML framework and benefit from automatically scaled ML processes in serverless containers.
To optimize ML models more efficiently, Canonical has equipped the AutoML component Katib with a new user interface. This enables data scientists to have more direct access to logs, make hyperparameter tuning easier, and provide direct access to the test metrics in the database with a Tune API.
Charmed Kubeflow 1.7 is now connected to the open-source platform PaddlePaddle for the first time. It enables online training of large deep neural networks from distributed data sources with billions of features and trillions of parameters. There are also various dashboards that offer comprehensive observability, including the infrastructure. The new version has added Nvidia Triton as another framework for model serving and successfully completed the certification for Nvidia DGX.
For a complete overview of all the new features and improvements in Charmed Kubeflow 1.7, check out the Ubuntu blog post.
