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.

Original post →

More from coding & agent

• Anthropic engineer says the right way to use Claude is to build self-prompting agents — mathemagie · 2026-07-24
• Wayfinder says its AI trading agents now rewrite strategies every cycle — templecrash · 2026-07-24
• Beginner asks what really matters in AI automation before selling it to clients — Responsible-Box-4905 · 2026-07-24
• Pi 0.82.0 lets bash commands detect the loaded model and provider — mitsuhiko · 2026-07-24
• Claude Code is being used to generate ComfyUI workflows instead of downloading them — My-NameWasTaken · 2026-07-24
• Claude’s code-comment narration looks like the same failure mode as Bash-comment reasoning — mertdumenci · 2026-07-24

coding & agent channel →