SGLang announced version 0.5.22 on GitHub on October 19, 2026, with release notes listing standard PyTorch support for Apple Silicon and broader model support. According to the maintainers, the PyTorch backend joins the existing MLX option for Apple hardware. Sources: SGLang
The release notes list new support for Clef/Clef Flash, EmbeddingGemma 2 and LLaDA2.2. They also list HunyuanImage-3.0-Instruct among image model additions and ComfyUI integration for MiniMax-H3 and Qwen-Image 2.1. Sources: SGLang
PyPI records package publication on October 16, 2026, three days before GitHub’s release date, and lists Python 3.11 as the minimum requirement. The upgrade notes deprecate --enable-dp-attention in favor of --attn-dp-size N and change --dp-size to count model replicas, with a DeepSeek context-parallelism exception. Affected deployments should consult the upgrade notes before changing these flags. Sources: SGLang, PyPI
The maintainers also report a fix for incorrect Mamba state restoration in hybrid linear models. According to the release notes, the error could occur when speculative decoding was combined with structured output. Sources: SGLang
What to do
- Before upgrading to SGLang 0.5.22, confirm that the deployment uses Python 3.11 or later.
- Review the parallelism upgrade notes before changing flags, especially for deployments affected by the DeepSeek context-parallelism exception. Outside that exception, replace --enable-dp-attention with --attn-dp-size N and check that --dp-size reflects the intended replica count. Sources: SGLang, PyPI