The Autodesk Machine Learning Platform partnered with Outerbounds to make large-scale distributed training more accessible to the community by contributing to an extensible user-centric ML orchestration framework called Metaflow. Together, we developed the  ray_parallel decorator to seamlessly run Ray applications within Metaflow flows, leveraging AWS Batch and Ray for distributed computing. We successfully battle-tested this extension through applying many different Ray use cases, including fine-tuning a large language model using Ray Train. Additionally, we integrated functionality within Metaflow to easily provision an HPC cluster though AWS Batch, unlocking new possibilities for AI use cases that require training on massive amounts of data at high speeds with parallel performance. This collaboration will give users the tools needed to run next-generation ML pipelines for building their own state of the art models.

Discover how AI techniques are revolutionizing water management through pump optimization in distribution networks. By applying Reinforcement Learning (RL), researchers have developed intelligent systems that dynamically optimize pump operations, improving efficiency and energy consumption. This blog explores the advancements achieved with RL techniques, the transition from Genetic Algorithms to RL, and the benefits of using RLlib. Learn about the evaluation results, the hybrid RL approach for maintaining trust, and how RL combines with the query-based warm start approach for reliable and cost-effective solutions in water distribution networks.