Ray Summit 2026: RL Post-Training Forces Open-Source AI Infrastructure to Converge

Today, Anyscale’s Ray Summit 2026 opens its main conference days at the San Francisco Marriott Marquis — and for the first time, a second community is in the building. Inferact’s inaugural vLLM Conference runs concurrently under the same roof, with a single pass granting access to both tracks. That co-location is not a scheduling convenience. It reflects a structural shift in how production AI is built: reinforcement learning post-training now requires inference and training infrastructure to coordinate simultaneously in real time, collapsing what were once separate engineering disciplines into one systems problem.

The clearest articulation of why is scheduled to come from NVIDIA’s Bryan Catanzaro, VP of Applied Deep Learning Research, whose keynote on building and post-training open models with Nemotron frames RL post-training explicitly as a systems engineering challenge: agentic RL requires data generation, inference, and model weight updates to run in balance across CPUs and GPUs simultaneously. Pause the inference side to let training catch up and GPU utilization plummets; let inference race ahead and the policy the model trains against becomes stale. Ray — Anyscale’s open-source Python-native distributed framework — is the orchestration layer that keeps those three workloads coordinated. vLLM is the inference engine that feeds rollouts fast enough to keep training from idling. The NVIDIA Ray Summit program frames this directly: agentic reinforcement learning requires balanced data generation, inference, and training across CPUs and GPUs.

The two communities showing up in the same building this week is the natural consequence of that interdependency.

What Ray Summit 2026 Is

Ray Summit is Anyscale’s annual conference for engineers building production AI infrastructure on the Ray open-source framework. The 2026 edition runs August 24–26 at the San Francisco Marriott Marquis, with more than 3,000 confirmed attendees. Tuesday and Wednesday (August 25–26) are the main conference days — morning keynote blocks from 9:30 to 11:30 AM PT, afternoon breakouts from 12:30 to 5:30 PM PT. Monday (August 24) was a dedicated hands-on training day; four of the six workshop tracks sold out before the event began.

Ray itself is a Python-native, open-source unified framework for scaling AI and machine learning applications across distributed compute. It was created at UC Berkeley’s RISELab in 2016 under the direction of Ion Stoica — who also co-created Apache Spark and SkyPilot, and is listed as a creator of vLLM — making him one of the few researchers whose projects span both communities now sharing this venue. Details on Anyscale’s founding and Stoica’s role are available at Anyscale’s about page. Ray’s production deployments include OpenAI’s model training infrastructure, Spotify’s internal Hendrix LLM training framework, Apple Maps’ batch inference pipeline, and Netflix’s ML platform, among others.

Passes cost $400 for the two main conference days (which include the vLLM Conference track) and $450 for the full three-day access including Monday’s training day. The pass also covers post-event session recordings for all registered attendees.

Why RL Post-Training Makes Inference and Training Infrastructure Inseparable

The central technical argument behind the Ray+vLLM convergence is this: RL post-training is not a training-only workload. The standard RL pipeline requires at minimum four concurrent processes — rollout generation (inference), reward scoring, policy gradient updates (training), and reference model inference for KL divergence bounding. In synchronous setups, GPU utilization drops by as much as 50 to 80 percent because training nodes sit idle while they wait for the slowest rollout in each batch to complete. A survey of asynchronous RL post-training frameworks documents this synchronization penalty across Laminar, LlamaRL, AsyncFlow, and other leading approaches.

The solution adopted by leading RL infrastructure frameworks — Laminar, LlamaRL, AsyncFlow, and Anyscale’s own tooling — is full decoupling of rollout generation and training into separate distributed node pools that communicate asynchronously. That decoupling only works if the inference side (rollout generation) can keep pace with the training side’s appetite for new trajectories. This is what makes vLLM’s role critical: it is the serving infrastructure that generates rollouts fast enough that training workers are never the ones waiting. Anyscale’s RL for LLMs documentation describes this coordination architecture in detail.

Ray’s actor-model abstraction — which lets arbitrary Python processes communicate, share state, and coordinate resource allocation across heterogeneous clusters — is what makes the orchestration layer possible. The production RL pipelines running in the Spotify, Discord, and Mercor sessions at this conference are, in architectural terms, Ray-orchestrated clusters where vLLM is one class of actor and training workers are another.

GRPO (Group Relative Policy Optimization) is the algorithm most commonly running on this infrastructure in 2026 — it eliminates the critic network required by classic PPO, halving the number of models that need to stay in GPU memory during training, which directly improves the GPU utilization math that makes large-scale RL post-training economically viable.

What Today’s Keynote Block Will Cover

The main stage opens at 9:30 AM PT this morning with an address from Anyscale co-founders Robert Nishihara and Ion Stoica. CEO Keerti Melkote and Nscale CEO Josh Payne are scheduled to speak together, with Nscale’s full-stack AI cloud positioned as the compute layer for Ray-based training and inference workflows.

Catanzaro’s NVIDIA keynote is the session most attendees have circled. He is expected to detail how Nemotron’s data curation and RL workflows are built on Ray — covering the trade-offs of using open model weights and training recipes, and identifying where current systems still fall short in coordinating CPU/GPU workloads for agentic RL. Details on NVIDIA’s full Ray Summit program include additional sessions on Blackwell-hardware serving and distributed inference.

On the vLLM Conference track, the full schedule is available at the vLLM Conference 2026 page. The "State of vLLM 2026" keynote — covering the Flat Model and Model Runner V2 migrations, disaggregated serving with multi-tier key-value cache offloading, speculative decoding progress toward 1,000-plus tokens per second, and production-grade quantized KV cache compression — will be delivered by Woosuk Kwon, vLLM creator and Inferact co-founder/CTO, alongside Zachary Xi.

Also on the main stage: Andrew Beam, CTO of Lila Sciences, speaking on AI-first drug discovery infrastructure; Liam Fedus, co-founder of Periodic Labs and co-author of the 2021 Switch Transformer paper that established sparse mixture-of-experts scaling, building in the foundation model space since leaving OpenAI; and Elvis Nava, co-founder of robotics startup Mimic, on training generalist robotic policies with Ray.

What the vLLM Roadmap Says About Where LLM Serving Is Headed

The technical state of vLLM in 2026 illuminates how far LLM inference engineering has moved in three years. When Woosuk Kwon and collaborators published the PagedAttention paper in 2023, the core insight was borrowed from operating systems: instead of pre-allocating contiguous GPU memory blocks for each sequence’s key-value cache — a practice that caused enormous fragmentation — treat KV cache like virtual memory, divided into fixed-size pages that can be allocated non-contiguously. Research on multi-tier KV cache memory management details how this approach evolved into the disaggregated serving architectures on vLLM’s 2026 roadmap.

PagedAttention enabled continuous batching (serving multiple requests simultaneously by sharing GPU time across partial sequences) and prefix caching (reusing cached attention states when multiple requests share a common prefix, such as a long system prompt). Those two capabilities are now table stakes.

The 2026 roadmap targets harder problems. For context on why: serving a LLaMA-2 70B model at a batch size of 32 across 2,048-token contexts requires roughly 640 gigabytes for the KV cache alone — far beyond the 80-gigabyte capacity of a single H100 GPU. Analysis in the LMCache KV memory paper documents this memory wall in detail. The multi-tier KV offloading approach on vLLM’s roadmap manages this by spilling KV cache from GPU memory to CPU RAM to disk to remote storage in a hierarchy, maintaining high throughput by overlapping compute with cache movement. Quantized KV cache compression reduces the memory footprint of each cache entry by representing it at lower precision — trading a small quality degradation for a large reduction in VRAM consumption.

Speculative decoding — where a small draft model proposes multiple tokens at once and the larger target model verifies them in a single forward pass — is the mechanism vLLM is targeting for 1,000-plus token-per-second throughput. The draft model is fast; the verifier is accurate; the combination produces both speed and quality if the acceptance rate (how often the large model accepts the draft model’s proposals) stays high. At the vLLM Conference track today, Siyuan Fu of NVIDIA will describe how full-stack optimization from FlashInfer kernels through NVIDIA Dynamo’s distributed serving layer produced the record-setting DeepSeek and MiniMax inference results in the SemiAnalysis InferenceX Benchmark.

Production AI at Scale: Who Is Building on Ray and What They Have Learned

The afternoon breakout sessions represent some of the most operationally specific conference content in the AI infrastructure calendar. Several sessions are worth flagging for engineers evaluating production Ray deployments:

Spotify’s Aamir Anwar will present engineering lessons from Hendrix — Spotify’s internal LLM training framework built on Ray Train — including failure modes and scaling decisions made in production. Discord’s Serrana Aguirregaray will describe the platform’s ML infrastructure evolution, covering the transition from open-source Ray to Anyscale’s managed platform to serve a consumer application at hundreds of millions of users. Apple’s Jack Li will detail how Apple Maps uses Ray for large-scale batch inference and LLM evaluation at Apple’s operational scale.

Recursion Pharmaceuticals’ Aleksandar Djuric will present one of the more technically unusual sessions: how one of the most compute-intensive biotech companies in the world orchestrates multimodal AI across a hybrid cloud environment that combines vast biological imaging data with genomic and chemical datasets — a use case that requires exactly the heterogeneous hardware coordination Ray’s actor model is designed for.

Adyen’s Hanna van der Vlis will present work on pretraining a tabular foundation model for payments on trillion-token-scale transaction data — an application of foundation model techniques to domain-specific structured data that illustrates how far the paradigm has traveled from natural language. The full Ray Summit 2026 agenda lists all session details.

A session from Shaowei Su of Microsoft AI will cover how the team used Ray to build MAI-Thinking-1, Microsoft’s flagship reasoning model. Microsoft’s own MAI-Thinking-1 launch announcement confirms the model scored 97.0% on AIME 2025, trained without distillation from third-party models.

Beyond Autoregression: Diffusion LLMs Challenge the Token-by-Token Paradigm

One session that stands apart from the infrastructure-and-scale focus is Volodymyr Kuleshov’s breakout, in which the Inception AI Lab co-founder and chief scientist will argue that diffusion-based large language models represent a genuine alternative efficiency profile for long-context and structured generation tasks.

Standard autoregressive LLMs predict the next token sequentially, left to right — one forward pass per token. Diffusion LLMs instead start from a fully corrupted or masked sequence and iteratively denoise it, generating all tokens in parallel across multiple forward passes. Research on diffusion versus autoregressive language models details the architectural trade-offs. The trade-off: multiple forward passes versus one per token, offset by the parallelism of each pass. For long outputs, the arithmetic favors diffusion; for short commonsense tasks, autoregressive models remain competitive. Inception AI Lab’s Mercury Coder has demonstrated generation rates exceeding 1,000 tokens per second in code generation tasks — a performance profile that changes the inference economics calculation for specific workload types.

Andrew Dai of Elorian AI will also present a critical perspective on visual AI limitations — the session title, "Why Current AI Cannot Truly See the World," signals a challenge to the field’s confident benchmark progress on perception tasks.

Physical AI: Where Robotics Meets Infrastructure

The conference’s Physical AI track reflects another convergence — between the robotics community and the AI infrastructure community — that is only beginning to be institutionalized. Two keynote speakers represent this: Felix Heide of Torc (autonomous trucking) on multi-modal AI workloads for physical systems, and Elvis Nava of Mimic on training generalist robotic policies.

An Industry Roundtable on Physical AI during Tuesday’s lunch block offers structured discussion for robotics practitioners and autonomous systems engineers on the shared infrastructure questions that both communities need to solve: how to train policies that generalize across physical environments, how to manage the data pipelines that feed sensor-rich robotic simulation, and where the compute infrastructure for physical AI differs from language model training.

Wednesday’s keynote block will include NVIDIA sessions covering adapting Ray to scale RL training for Nemotron trillion-parameter models, and serving frontier mixture-of-experts models at the lowest token cost using NVIDIA Blackwell hardware.

Sponsor Ecosystem Reflects Ray’s Infrastructure-Critical Status

Platinum sponsors this year are AWS, CoreWeave, Google, and Nscale — a combination that spans hyperscalers and next-generation AI compute providers. Vincent Gonguet, Head of Technical Foundations at Amazon Bedrock, is scheduled to keynote on how Bedrock’s technical stack leverages Ray-based infrastructure patterns at Amazon scale, providing rare detail on how a major cloud provider’s managed AI service is architected under the hood.

Gold sponsors include NVIDIA, Nebius, Lila Sciences, Datadog, Simplismart, Intel, Encord, Lambda Labs, VAST Data, Verda, and Microsoft Azure. Session contributions come from BMW Group, Alibaba, JPMorgan Chase, Hyundai Motor Company, Pinterest, Autodesk, Grab, Google DeepMind, Red Hat, Netflix, and Zoox, among others. The breadth of that list — spanning automotive, financial services, consumer internet, biotech, and defense-adjacent autonomous systems — reflects Ray’s expansion beyond AI-native companies into sectors where distributed compute is a new but now-essential capability.

A note on Ray’s security posture: organizations evaluating Ray for production deployment should be aware that a critical authentication gap (CVE-2023-48022) in the framework has been actively exploited since 2023 and remains unpatched as of 2026. The ShadowRay security disclosure documents the initial campaign; ShadowRay 2.0 was still actively exploiting the vulnerability as of late 2025, converting exposed clusters into cryptomining and DDoS infrastructure. Anyscale has released tooling to verify proper port exposure and updated security documentation; the risk is mitigated when Ray clusters are deployed in isolated networks per the framework’s design intent — but the vulnerability itself has not been patched at the framework level.

Wednesday, August 26: What to Watch on Day Two

Day 2 follows the same structure — keynotes from 9:30 to 11:30 AM PT, a Life Sciences Industry Roundtable at lunch, and breakouts through 5:30 PM PT. The vLLM Conference continues with sessions on agentic inference workloads, AMD and Intel hardware backends for vLLM, speculative decoding at scale, Google DeepMind’s Gemma quantization work, and MiniMax’s production serving experience. Red Hat’s Greg Pereira and Michael Goin will present llm-d, Red Hat’s production-grade distributed inference platform built on vLLM, covering enterprise scheduling and reliability patterns. The conference ends Wednesday at 5:30 PM PT.

Session recordings will be available to all registered attendees following the event through the Linux Foundation registration portal.

---

Frequently Asked Questions

What is Ray Summit, and who is it for?

Ray Summit is Anyscale’s annual conference for engineers building production AI infrastructure using the Ray open-source framework. It focuses specifically on distributed systems running real workloads — training foundation models, building multimodal data pipelines, orchestrating RL post-training — rather than research previews or demos. The 2026 edition includes the inaugural co-located vLLM Conference, so a single pass covers both communities. It is best suited for ML infrastructure engineers, ML platform leads, and distributed systems engineers who need architectural depth rather than product announcements.

Why are the Ray and vLLM communities now at the same conference — and what does that mean for AI infrastructure?

The co-location reflects a technical reality: reinforcement learning post-training requires inference and training infrastructure to run simultaneously and coordinate in real time. Generating rollouts (inference) and updating model weights (training) are no longer sequential steps — they must happen in parallel on the same cluster, with inference results feeding training continuously. Ray is the orchestration layer that coordinates those workloads across heterogeneous hardware; vLLM is the inference engine that keeps rollout generation fast enough for training workers to stay utilized. The two communities now share the same production systems problem.

How does PagedAttention work, and why is it the foundation of modern LLM serving?

PagedAttention, the core architectural innovation in vLLM, applies virtual memory concepts from operating systems to LLM inference. Traditional LLM servers pre-allocated contiguous blocks of GPU memory for each request’s key-value cache, causing massive memory fragmentation and low throughput. PagedAttention divides KV cache into small fixed-size pages that can be allocated non-contiguously — the same insight that lets an OS serve many applications simultaneously without each one owning a contiguous physical memory region. This enables continuous batching (mixing partial requests from many users), prefix caching (reusing shared system-prompt attention states), and disaggregated serving (transferring KV cache pages across GPU nodes). It became the de facto standard for LLM serving engines after 2023 and now underpins vLLM, TensorRT-LLM, and SGLang.

What companies are using Ray in production, and what can their sessions at Ray Summit teach infrastructure teams?

Spotify built Hendrix — its internal LLM training framework — on Ray and will discuss production scaling lessons at this year’s summit. Apple uses Ray for batch inference and LLM evaluation inside Apple Maps. Discord migrated from open-source Ray to Anyscale’s managed platform as its ML workloads scaled to hundreds of millions of users. Recursion Pharmaceuticals uses Ray to orchestrate multimodal AI across hybrid cloud environments combining biological imaging, genomic, and chemical data. Adyen built a payments-domain foundation model on trillion-token tabular transaction data using Ray Data. Amazon Bedrock’s keynote will describe how AWS’s managed AI service is built on Ray-based infrastructure patterns at hyperscale.