SGLang is a high-performance serving framework for large language models and multimodal models, now the de facto industry standard deployed on over 400,000 GPUs worldwide. It introduces RadixAttention for KV cache reuse, a zero-overhead CPU scheduler, and compressed finite state machines for faster structured output decoding. SGLang supports a wide range of models including Llama, Qwen, DeepSeek, Kimi, GLM, and diffusion models, and runs across NVIDIA, AMD, Intel, Google TPU, and Ascend NPU hardware.