SGLang v0.5.19 Released with New Model Support and Performance Optimizations
SGLang released version v0.5.19, introducing support for new autoregressive models including Qwen3.8, Ling-3.0, and Granite 4.2, as well as diffusion models like LongCat-Image-Edit. The update also brings beam search capability, DeepEP v2 for MoE models, LayerNorm sequence parallelism, and FP8 activation quantization for MoE experts on Hopper GPUs. As open-weight LLMs evolve rapidly, inference engines must quickly adopt new model architectures and hardware-level optimizations to maintain high throughput. These updates allow teams deploying SGLang to run newer models with lower latency and higher output throughput across various hardware platforms. The new beam search feature allows users to request the top best sequences via `beam_width`, though it currently does not mix with speculative decoding or disaggregation. Additionally, LayerNorm sequence parallelism reduces prefill time by up to 5.6% on NVIDIA B200 GPUs, while DeepEP v2 standardizes fixed-size buffers so decode can run under CUDA graphs across multiple nodes.
## BACKGROUND
SGLang is an open-source serving framework hosted under LMSYS designed for high-performance inference of large language and multimodal models. It provides specialized execution backends, memory management routines, and quantization support to serve AI models efficiently at enterprise scale.