Ubuntu distributor Canonical has released version 1.7 of its Charmed Kubeflow MLOps platform. Thanks to the integration of Knative, the new release opens up the possibility for companies to run machine learning models on Kubernetes as event-driven serverless applications. In addition, KServe provides a Custom Resource Definition (CRD) for inference and model serving. A revised user interface for Katib should also simplify hyperparameter tuning in Charmed Kubeflow for data scientists.
More automation and flexible scaling for ML workflows
The to the initiated by Google Open-Source-Kubeflow Project The similar toolkit for automating the workflows for training, tuning and deploying ML models expands its serverless capabilities with the integration of Knative. According to the announcement in the Ubuntu blog, the declared goal is to relieve developers and data scientists of routine infrastructure tasks. On the one hand, they should benefit from the advantages of automatically scaled machine learning processes in serverless containers and, on the other hand, they should also be able to work with their preferred ML framework (TensorFlow, XGBoost, ScikitLearn, PyTorch, ONNX etc.) can concentrate.
With an eye on optimizing ML models more efficiently, Canonical has the AutoML-Komponente Katib equipped with a new user interface. This gives data scientists more direct access to logs – hyperparameter tuning should also be easier. A Tune API in Katib also provides direct access to the test metrics in the database, making it easier and faster to set up tuning experiments.
Industrial scale deep learning
To create deep learning models for industrial use, Charmed Kubeflow 1.7 offers a connection to the Open-Source-Plattform PaddlePaddlewhich enables, among other things, online training of large deep neural networks with billions of features and trillions of parameters from distributed data sources.
Other innovations in version 1.7 include various dashboards that are intended to contribute to more comprehensive observability – for example with regard to the infrastructure. Nvidia Triton has been added as another framework for model serving. In addition, Charmed Kubeflow has successfully completed the certification for Nvidia DGX.
Summarizes a complete overview of all improvements and new features in Charmed Kubeflow 1.7 the Ubuntu blog post announcing the new release.
(map)
