SGLang talk says RL post-training is really an inference problem

BanghuaZ · x · 2026-07-24

A talk at a dstack / Crusoe / SGLang event argues that reinforcement learning post-training is fundamentally an inference problem.

The speaker says the most expensive part of RL is rollouts, and rollouts are essentially serving.
That means inference-engine optimizations should directly carry over into RL training frameworks.
The example given is Miles, an SGLang-native RL framework, where serving and training share the same stack.
The broader point: one unified engine can reduce training/inference drift, avoid forks, and keep optimizations reusable across inference and RL.
The slides also claim the stack reached 500+ TPS in two weeks, with measurable gains such as 18–34% interactivity improvement and 10.17× at 1M context.

More from coding & agent