Some of our interesting observations from working on multi-turn text2SQL: - Data-efficient RL works pretty well: We did very typical GRPO settings; Just make sure to use "hard-enough" samples and no KL. KL can stabilize learning early on but will always bring down rewards

Wide-EP and prefill/decode disaggregation APIs for vLLM are now available in Ray 2.52 🚀🚀 Validated at 2.4k tokens/H200 on Anyscale Runtime, these patterns maximize sparse MoE model inference efficiency, but often require non-trivial orchestration logic. Here’s how they

1/n 🚀 Introducing SkyRL-Agent, a framework for efficient RL agent training. ⚡ 1.55× faster async rollout dispatch 🛠 Lightweight tool + task integration 🔄 Backend-agnostic (SkyRL-train / VeRL / Tinker) 🏆 Used to train SA-SWE-32B, improving Qwen3-32B from 24.4% → 39.4%

🧑‍🏫On-Policy Distillation is available as an example on SkyRL! The implementation required no library code changes, and we were able to reproduce AIME math reasoning experiments from the @thinkymachines blogpost. Check out our detailed guide to see how! https://t.co/TqDS649oFQ

🔥 New Blog: “Disaggregated Inference: 18 Months Later” 18 months in LLM inference feels like a new Moore’s Law cycle – but this time not just 2x per year: 💸 Serving cost ↓10–100x 🚀 Throughput ↑10x ⚡ Latency ↓5x A big reason? Disaggregated Inference. From DistServe, our

☁️SkyRL now runs seamlessly with SkyPilot! Let @skypilot_org handle GPU provisioning and cluster setup, so you can focus on RL training with SkyRL. 🎯 Launch distributed RL jobs effortlessly ⚙️ Auto-provision GPUs across clouds 🤖 Train your LLM agents at scale Get started

🚀 SkyRL has day-zero support for OpenEnv!! This initial integration with OpenEnv highlights how easily new environments plug into SkyRL. Train your own LLM agents across containerized environments with simple, Gym-style APIs 🔥 👉 Check it out:

⚙️ Engineering for scale and speed. Next-gen RL libraries are redefining how agents learn and interact. 🎤 Open Agent Summit Speakers: • @sumanthrh (Anyscale) – “SkyRL: A Modular RL Library for LLM Agents” • @tyler_griggs_ (UC Berkeley) – “Lessons in Agentic RL Modeling and

https://t.co/BNJRwUtP5D

🚀 Excited to release our new paper: “Barbarians at the Gate: How AI is Upending Systems Research” We show how AI-Driven Research for Systems (ADRS) can rediscover or outperform human-designed algorithms across cloud scheduling, MoE expert load balancing, LLM-SQL optimization,

SkyRL now supports Megatron! Training massive MoE models demands more than just ZeRO-3/FSDP sharding. The Megatron backend for SkyRL unlocks high throughput training with: ✅ 5D parallelism (tensor + pipeline + context + expert + data) ✅ Efficient training for 30B+ MoEs

Training our advisors was too hard, so we tried to train black-box models like GPT-5 instead. Check out our work: Advisor Models, a training framework that adapts frontier models behind an API to your specific environment, users, or tasks using a smaller, advisor model (1/n)!

when bro suddenly starts submitting PRs with detailed descriptions, correct grammar and casing, and testing steps

RT @NovaSkyAI: Scaling agentic simulations is hard, so in collaboration with @anyscalecompute we wrote up our experience using Ray for agen…

Do you find it challenging to run RL / agent simulations at a large scale (e.g. dealing with docker and remote execution)? Check out our blog post https://t.co/iNPivIzbc2 where we show how to do it with Ray and mini-swe-agent (kudos to @KLieret)

🚀🚀🚀🚀

An amazing collaboration with the Biomni team! We introduce Biomni-R0, a reasoning-enabled multi-task multi-tool biomedical research agent, trained through end-to-end RL. The results are impressive -- 2x stronger than the base model and >10% better than GPT 5 and Claude 4 Sonnet!

2. Online quantization: Policy model runs in BF16, while rollouts run in FP8/ Int8. FlashRL patches vLLM’s weight loading APIs to ensure compatibility. We’ve ported over these patches for v0.9.2 and further optimized weight syncing! Try it out:

We’ve implemented both ingredients in FlashRL: 1. Truncated Importance Sampling: Ensures that the difference in rollout and policy logprobs doesn’t hurt performance This is a simple token-level correction factor for your policy loss. Can help stabilize training even without

SkyRL now has native support for FlashRL! We now support Int8 and FP8 rollouts, enabling blazing fast inference - upto 1.7x for the DAPO recipe - without compromising performance!

I’m confused by ppl saying GPT-5 is bad. At least with thinking mode, I find it way more focused, more reasonable, and less rambling. Sometimes I even have to ask it to go into more detail! IMO best model from OAI so far.