Union, the company behind the open source AI Orchestrator Flyte, with a mission to unify work streams across ML, Platform, Data and Ops teams to efficiently create AI products, announced the early availability of its Serverless offering. The new service offers AI developers a frictionless experience to run a wide variety of AI workloads including fine-tuning off-the-shelf models, LLM batch inference and model training workflows.
Finding a suitable model among the various adaptations and fine-tuned variations of popular LLM models on Hugging Face, or deciding to fine-tune your own, can be challenging. The lack of rigor, infrastructure management, productionization, and tracking for how models are trained and released makes further tuning or continuous training difficult. Union addresses these challenges by unifying processes such as validating datasets, storage and tracking, versioning training code, building container images, and allocating right-size machines. This provides robust, reproducible workflows that can run seamlessly both locally and in the cloud.
“Union Serverless is a critical step towards democratizing AI infrastructure, making it accessible to organizations regardless of their existing capabilities,” said Ketan Umare, CEO of Union. “Our goal with this release is to empower teams to become self-sufficient AI powerhouses. This innovation is part of our long-term vision of creating a unified fabric that leverages both customer cloud resources and pooled resources available to Union, all through an intuitive user experience.”
Also Read: Databricks Agrees to Acquire Tabular, the Company Founded by the Original Creators of Apache Iceberg
The Ultimate Productivity Tool for Machine Learning Developers
Union Serverless expands the range of possible deployment options, including BYOC (“bring your own cloud”) model in which Union manages resources in a customer’s VPC. Many enterprises and startups leverage Union to develop and ship AI-powered products. Examples include autonomous driving use cases at Woven by Toyota, data preparation and training for computer vision models at Flawless, running cancer predictions at ArteraAI, and more. Union Serverless offers flexible access to GPUs and other resources by the minute, with the option of migrating to Union BYOC for full control over data locality, cloud resources, and spending.
Union Serverless is the perfect tool for AI developers looking to develop workflows that address core business opportunities quickly. Developers focus on applying their domain knowledge to authoring and iterating on their models, without having to solve the added challenge of deploying and running code on the cloud at scale. Union Serverless provides an out-of-the-box solution for managing underlying compute infrastructure and optimizing the scheduling of developer workloads. Application developers can use Union Serverless’s lightweight Python library to write code to run on specialized GPUs, integrate with their proprietary datasets, and incorporate third-party APIs like ChatGPT to unlock productivity and create new AI applications.
“In less than five minutes, I connected to my Serverless namespace and designed, submitted, and executed an ML pipeline I had developed locally on to large-scale distributed compute. I didn’t have to worry about authentication tokens, Dockerfiles, or managing a Kubernetes cluster,” said Grantham Taylor, ex-Principal Data Scientist at Capital One.
“Union is my backend. It allowed me to develop, iterate, and test locally before seamlessly deploying workflows that handle the heavy lifting for our end-user application.” Vikram Advani, founder AI stealth company, cofounder Viro Media & UpNext (Amazon) and ex Meta.
SOURCE: PRNewsWire