December 2025 Summaries

KubeRay v1.5 introduces significant enhancements to the Ray on Kubernetes ecosystem, including increased security, performance, and scalability. This release incorporates over 180 commits from around 50 contributors, offering features such as full support for Ray authentication token mode to secure cluster access, incremental upgrades in RayService to reduce resource overhead, and new submission modes in RayJob to streamline job management. Advanced deletion policies allow more flexible resource cleanup while a lightweight job submitter improves startup times. Integration with batch schedulers like Yunikorn and Volcano enables efficient gang scheduling, while the new Ray Label Selector API enhances autoscaling decisions. Additionally, KubeRay v1.5 integrates with NVIDIA's Kai-Scheduler for gang scheduling and workload prioritization, and it enhances the API Server and Dashboard for improved user experience. The update also includes a tuning guide for large-scale deployments, enabling more effective management of extensive Kubernetes clusters.

Multimodal AI models, which integrate various types of data like text, images, and audio, are advancing AI technology but require complex and resource-intensive training processes. This blog post explores how disaggregated hybrid parallelism using Ray can enhance training efficiency by applying specific parallelization strategies to different modules within a model, such as using sequence parallelism for vision encoders and tensor parallelism for language models. By implementing this approach on Ray and testing it with the Qwen-VL 32B model, the authors achieved a 1.26–1.37x improvement in throughput over traditional tensor parallelism and enabled training sequences up to 7x longer than with DeepSpeed ZeRO3. The strategy allows for better memory efficiency and avoids out-of-memory errors commonly encountered with monolithic parallelization methods, demonstrating Ray's capability to handle the demands of state-of-the-art multimodal AI models effectively. The post encourages further exploration of this method across different hardware and model architectures, inviting feedback and contributions through their GitHub repository.