# SGLang RadixAttention: Structured Output + JSON-Mode + Multi-Branch Caching

> Source: https://sukruyusufkaya.com/learn/fine-tuning-cookbook/ftc-sglang-radixattention-structured-output
> Updated: 2026-08-22T00:16:41.712Z
> Category: Fine-Tuning Cookbook (Model-by-Model)
> Module: Part XV — Serving Engineering
**TLDR:** SGLang (Zheng et al. 2024) — vLLM'in alternatif rakibi. RadixAttention: prefix cache'in Trie/Radix tree'de organize edilmiş hali → multi-branch sharing. Constrained decoding (regex, JSON schema), structured output native, agent workflows için optimize. RTX 4090'da Llama 3.1 8B SGLang serving + JSON-only response.