Research
SGLang v0.5.20 Returns Per-Step Sampling Masks So RL Trainers Can Replay a Rollout Exactly
SGLang shipped v0.5.20 on 18 September 2026 with 713 PRs from 237 contributors. The notable feature for RL practitioners is `return_sampling_mask`, which makes each decode step return the exact token support the sampler drew from along with the log-probability of the sampled token under it, so a trainer can replay a rollout without reconstructing top-k or top-p itself. The release also adds GLM-5.3-Flash, Hy4-Preview, Qwen3.8-Flash-Next, K2 Horizon and Nanbeige4.2 on the autoregressive side, plus SenseNova-U1.5-8B-MoT, FastH3 and VDN-H3 on the diffusion side. Given the training-inference mismatch work published the same week, exact rollout replay is directly relevant to anyone debugging RL instability.
↳ Follow the thread