Ray Summit 2026, held August 24-26 at the San Francisco Marriott Marquis, drew more than 2,000 attendees to focus on reinforcement learning as a production engineering problem, with Robert Nishihara, …
Anyscale and NVIDIA cuDF have integrated GPU-native operators into Ray Data, enabling up to 3x better total cost of ownership (TCO) on select data curation workloads compared to CPU-based solutions. T…
Anyscale's Ray team reports that after a round of Ray Core improvements, batch inference on 500 nodes is 23% faster end to end, Ray Data shuffle is 24% faster, and placement groups are 62× faster to r…
Ray Serve LLM, developed by Anyscale, introduces token-load-aware routing to improve LLM serving efficiency, moving beyond naive KV cache reuse strategies. The new approach addresses the challenges of…
Anyscale co-founder Christian Stano outlines a three-step maturity curve for companies to build 'learning loops'—compounding flywheels of data curation, custom model training, and inference powered by…
SkyRL now supports FP8-accelerated training and rollout with on-policy weight sync, reducing end-to-end step time by up to 23% while closely tracking BF16 convergence in long-run RL experiments. The s…
Ray 2.58 introduces native sandboxing built on gVisor, allowing users to create isolated environments from OCI container images and manage them with Ray primitives. The feature, developed in partnersh…
Ray Serve's async inference feature was used to build a video-indexing service that returns a task ID immediately and processes videos in the background, with the IndexingIngress deployment enqueuing …
Ray Direct Transport (RDT), a feature in Ray Core, enables RDMA-backed weight transfer in reinforcement learning for LLMs, achieving up to 7.5x speedup over naive RDMA implementations on GB200 nodes u…
Anyscale, the company behind the open-source Ray framework, has signed a definitive agreement to join Nscale, a vertically integrated neocloud provider. The deal will expand investment in Ray and the …
Anyscale Inc. today introduced the Anyscale Physical AI Skill, a new workload skill for building robotics and autonomous-driving systems with Ray and Anyscale that generates code, dependency pins, com…
Ray's new Resource Isolation feature, built using Linux kernel control groups (cgroup-v2), delivers up to 1.5x speedup and eliminates node deaths and job failures on previously unstable workloads such…
Ray Data 2.56 introduces memory-aware execution and improved prefetching to reduce out-of-memory failures and unnecessary object spilling in AI data pipelines. The update includes automatic batch size…
Researchers from Anyscale and NVIDIA demonstrated a method to scale robot policy evaluation by disaggregating GPU-heavy simulation and policy inference workloads using Ray and Isaac Lab. The approach …
Ray Serve LLM, in partnership with Google Kubernetes Engine, announced major performance improvements achieving up to 4.4x higher throughput on prefill-heavy workloads and 24x higher on decode-heavy w…
Ray Data LLM, a library for large-scale batch inference, achieves 2x throughput over vLLM's synchronous LLM engine in production-scale workloads by optimizing hardware utilization and providing fault …
Engineers achieved up to 67% cost savings and 2.7x better goodput by using Prefill-Decode disaggregation with Ray and vLLM on AMD MI325X GPUs, separating prefill and decode phases onto dedicated hardw…
Ray Serve LLM and vLLM on AMD MI325X achieve up to 67% cost savings by disaggregating prefill and decode phases in LLM serving, separating them onto dedicated GPUs to eliminate interference and improv…
Alibaba's 1.7B parameter Qwen3-TTS voice cloning model was fine-tuned using Fully Sharded Data Parallel (FSDP) with PyTorch and Ray, demonstrating memory-efficient distributed training across 4 GPUs. …
Anyscale released agent skills for debugging Ray workloads, including /anyscale-platform-fix and /anyscale-platform-inspect, which automate troubleshooting of failing pipelines. A user demonstrated fi…