SGLang 发布: v0.5.20

来源摘要

# Highlights *713 PRs from 237 contributors.* **New models in this release** (see the [cookbook](https://docs.sglang.io/cookbook) for all supported models): | Model | Type | PRs | Cookbook | |---|---|---|---| | GLM-5.3-Flash | Autoregressive | [#36507](https://github.com/sgl-project/sglang/pull/36507), [#38621](https://github.com/sgl-project/sglang/pull/38621) | [link](https://docs.sglang.io/cookbook/autoregressive/GLM/GLM-5.3-Flash) | | Hy4-Preview | Autoregressive | [#36805](https://github.com/sgl-project/sglang/pull/36805) | [link](https://docs.sglang.io/cookbook/autoregressive/Tencent/Hy4-Preview) | | Qwen3.8-Flash-Next | Autoregressive | [#37500](https://github.com/sgl-project/sglang/pull/37500) | [link](https://docs.sglang.io/cookbook/autoregressive/Qwen/Qwen3.8-Flash-Next) | | K2 Horizon | Autoregressive | [#37654](https://github.com/sgl-project/sglang/pull/37654), [#38033](https://github.com/sgl-project/sglang/pull/38033) | [link](https://docs.sglang.io/cookbook/autoregressive/IFM/K2-Horizon) | | Nanbeige4.2 | Autoregressive | [#32151](https://github.com/sgl-project/sglang/pull/32151) | | | SenseNova-U1.5-8B-MoT | Diffusion | [#36606](https://github.com/sgl-project/sglang/pull/36606) | [link](https://docs.sglang.io/cookbook/diffusion/SenseNova/SenseNova-U1.5-8B-MoT) | | FastH3 (4-step MiniMax-H3 distill) | Diffusion | [#37480](https://github.com/sgl-project/sglang/pull/37480) | [link](https://docs.sglang.io/cookbook/diffusion/MiniMax/MiniMax-H3) | | VDN-H3 (hybrid-attention MiniMax-H3 distill) | Diffusion | [#37903](https://github.com/sgl-project/sglang/pull/37903) | [link](https://docs.sglang.io/cookbook/diffusion/MiniMax/MiniMax-H3) | **Sampling masks for RL rollouts.** With `return_sampling_mask`, each decode step returns the exact token support the sampler d