SGLang is a high-performance, open-source serving framework designed for low-latency and high-throughput inference of large language models (LLMs) and multimodal models. The engine utilizes advanced techniques such as RadixAttention for prefix caching, which can improve inference speeds by up to 5x, and speculative decoding via DFlash and Spec V2. It supports complex parallelization strategies, including tensor, pipeline, data, and large-scale expert parallelism (EP), as well as prefill-decode disaggregation. For example, deploying DeepSeek models on GB200 NVL72 clusters using PD disaggregation has demonstrated a 2.7x increase in decoding throughput and a 3.8x improvement in prefill throughput. The framework offers extensive hardware compatibility, running natively on NVIDIA GPUs (including GB300 and H100), AMD Instinct GPUs (MI300X), Google TPUs via the SGLang-Jax backend, and Intel CPUs. SGLang supports a diverse array of models, ranging from Llama and DeepSeek to diffusion models, and provides optimized serving for agentic workloads, such as reaching 500 TPS for GLM5.2 NVFP4. Currently hosted under LMSYS, SGLang powers trillions of tokens daily across more than 400,000 GPUs globally for major organizations like xAI, NVIDIA, and Google Cloud.