[Hands-on Lab] RL Training Lab: Launch a Real Training Job with the Miles RL Framework

Reinforcement learning post-training has become a critical stage in building capable foundation models — yet most open-source practitioners still struggle to set up a stable, high-throughput RL pipeline. This hands-on lab brings SGLang's battle-tested RL infrastructure directly to the GOSIM Paris community. Attendees will walk through building an end-to-end RL training loop using SGLang as the rollout backend, learn how to integrate it with the Miles framework, and tackle real challenges like training-inference mismatch and rollout efficiency at scale. SGLang today powers RL post-training for frontier models across the industry, running on over 400,000 GPUs worldwide. This session distills that production experience into practical, reproducible techniques — giving open-source developers and researchers a concrete path to running robust RL training on their own infrastructure.