SGLang 0.5.20 replays sampling steps for RL training

SGLang released version 0.5.20 on September 18, and the release note that matters most is a new flag, return_sampling_mask. Set it, and each decode step returns the exact set of tokens the sampler drew from. It also returns the log probability of whichever token got picked, according to SGLang's release notes.

That matters for anyone training reinforcement learning models on generated rollouts. Without the mask, a trainer replaying how a rollout was produced has to reconstruct the sampling behavior after the fact, guessing at top-k or top-p cutoffs from partial logs. With the mask returned directly, replay is exact instead of approximate.

The release is large: 237 contributors merged 713 pull requests this cycle. Alongside the sampling change, SGLang added support for five new language models, GLM-5.3-Flash, Hy4-Preview, Qwen3.8-Flash-Next, K2 Horizon, and Nanbeige4.2, plus three diffusion models, SenseNova-U1.5-8B-MoT, FastH3, and VDN-H3.

Model support additions are routine for a serving framework. The sampling mask isn't. It's the kind of feature that only gets built once RL-on-generated-data pipelines are common enough that unreproducible rollouts cost real debugging time.