Join top executives in San Francisco on July 11-12, to hear how leaders are integrating and optimizing AI investments for success. Learn More
Transformers are revolutionizing the capabilities of machine learning (ML), leading to a new era of generative AI. But how can data scientists build out models that fully take advantage of the power of transformers? That’s a question the open source Kubeflow effort is looking to help answer.
Kubeflow 1.7 became generally available today, providing the first update to the widely used open-source MLops platform since the debut of Kubeflow 1.6 in Sept. 2022. At its core, Kubeflow is an open-source ML toolkit that helps organizations to deploy and run ML workflows on cloud-native Kubernetes infrastructure. Among the themes of the Kubeflow 1.7 update is a focus on helping to better support transformer based models.
As model developers switch to using transformer-based models, they must also learn to utilize resources effectively. Kubeflow 1.7 can assist in workload placement and autoscaling, which can reduce resource usage and simplify operations. In particular, the Kubeflow Pipelines component in the 1.7 update benefits from the introduction of ‘Parallelfor’ statements which enables developers to more efficiently use parallel processes across AI accelerator hardware.
“Kubeflow 1.7 is a large release with hundreds of commits so the benefits and themes could be written many ways,” Josh Bottum, Kubeflow Community Product Manager, told VentureBeat. “We choose to highlight how model developers, that are moving to transformer model architectures, will benefit from 1.7’s python and Kubernetes native workflows, which speed model iteration and provide for efficient infrastructure utilization.”
Join us in San Francisco on July 11-12, where top executives will share how they have integrated and optimized AI investments for success and avoided common pitfalls.
MLops security gets a boost in Kuberflow 1.7
There is a lot to process about the Kubeflow update overall.
“The Kubeflow 1.7 release is the largest Kubeflow release to date,” Amber Graner, VP Community and Marketing at Arrikto Inc, told VentureBeat.
Graner noted over 250 people contributed code to the release with significant contributions and changes to Pipelines, Katib, and the Notebooks components, among other changes. Beyond the core code changes Graner said that one of the items that she’s most excited about for this release is the formation of the Kubeflow Security Team.
“During this release, the team was formed, identified a set of core images to scan, has identified vulnerabilities, and will begin to address these upstream rather than waiting for a downstream distribution to find and fix these vulnerabilities,” Graner said.
As an open source project, there is the core upstream technology and then individual vendors like Arrikto, Canonical or Red Hat for example can choose to create a packaged distribution for their own users.
“What users can expect to see with Kubeflow, as a project, product and community, is continued growth in both contributions and contributors, which ensures a healthy and more stable release and Kubeflow ecosystem,” she said.
KNative, KServe and Kubeflow
Kubeflow 1.7 also benefits from integration with a growing array of cloud native technologies that can help to aid in the deployment of MLops workflows.
Two such technologies are Knative for serverless deployment and KServe, for serveless ML inference. Andreea Munteanu, Product Manager at Canonical, which develops the ‘Charmed Kubeflow’ distribution, told VentureBeat that there are multiple benefits of adding KServe and KNative to Kubeflow.
Munteanu said that first and most important, organizations will be able to run serverless workloads, which unburdens developers to focus on scheduling the infrastructure underneath. She explained that Knative is designed to plug easily into existing DevOps toolchains, offering the flexibility and control customers need to adapt the system to their own unique requirements. “At the same time, KServe allows the deployment of single or multiple trained models onto model servers such as TFServing, TorchServe, ONNXRuntime or Triton Inference Server,” she said. “It expands extensively the number of applications that Kubeflow can support, allowing users to stay flexible with their choices and reducing operational costs.”