Ray ecosystem expands with NVIDIA integration and Red Hat support

Anyscale has updated its Ray Data library to include GPU-native operators through integration with NVIDIA’s cuDF and RapidsMPF. This update aims to improve the total cost of ownership for GPU-accelerated AI data processing by up to 3x. The integration allows for cuDF batch format support, which eliminates manual format conversions, and introduces GPU-native shuffling via RapidsMPF to bypass CPU memory bottlenecks.

Separately, Red Hat OpenShift AI has integrated Ray to support enterprise AI workloads. Starting with version 3.5, the platform includes preinstalled Training Hub within a Ray CUDA runtime image. This enables users to perform various fine-tuning algorithms, such as Low-Rank Adaptation (LoRA), on Ray clusters without manual dependency management, even in air-gapped environments. Ray offers elastic scaling and a unified runtime for data preprocessing, training, and serving.