GLM-5 RL Training with vLLM | Prime Intellect | Ray Summit 2026
Training trillion-parameter MoE models on long-horizon agentic tasks demands that inference and optimization scale independently. At Ray Summit 2026, Sami Jaghouar and Radhika Gaonkar from Prime Intellect present the asynchronous prime-rl stack, using GLM-5.1 as a case study: runs at up to 131K sequence length with 256 rollouts and sub-five-minute steps on 28 H200 nodes. They explain how vLLM powers low-precision inference, prefill/decode disaggregation, prefix-aware routing, and distributed KV cache storage, while the trainer combines FSDP, context parallelism, and expert parallelism. You'll leave seeing how these pieces reduce idle time and make frontier-scale agentic post-training more efficient. This session was part of the first vLLM Conference at Ray Summit, hosted by Inferact. Liked this video? Check out other Ray Summit vLLM session recordings: https://www.youtube.com/playlist?list=PLXMguE8Nc9o4 Subscribe to our YouTube channel to stay up-to-date on the future of AI: https://www.youtube.com/c/anyscale š Connect with us: LinkedIn: https://www.linkedin.com/company/joinanyscale/ X: https://x.com/anyscalecompute Website: https://www.anyscale.com/