← Runners

SGLangSGLang

LMSYS / sgl-project

开源Open source Apache2.0Apache2.0 服务化推理Inference serving RadixAttentionRadixAttention 多模态Multimodal partial lifecycle: active 核验Checked 2026-10-30

SGLang 是服务化推理的「前缀复用」一代：RadixAttention 把多轮/共享前缀的 KV 缓存复用做成主轴， agent 工作流场景收益直接。Apache-2.0，LMSYS 系学术开源。自报数字走 models 站 guides 快照口径。

SGLang is the prefix-reuse generation of serving engines: RadixAttention makes KV-cache reuse across turns and shared prefixes the main axis, so agent workloads benefit directly. Apache-2.0, LMSYS-line academic open source. Vendor figures follow the models-site guides snapshot.

适合与不适合When to use it, when not to

适合：agent 多轮调用、大量共享 system prompt/few-shot 的服务化部署、要官方结构化输出特性。不适合：单机开箱（Ollama）；要最宽生态兼容面（vLLM 更成熟）。

Fits: agent multi-turn calls, serving workloads with shared system prompts/few-shots, official structured-output features. Not a fit: casual single-machine use; needing the widest ecosystem surface (vLLM is broader).

中文生态China availability 国产 / 国内可直接用 / 备案状态 · 快照domestic / direct access / CAC filing · snapshot 2026-10-21

出品属地Origin: 国外出品foreign-made　sgl-project（LMSYS 系学术开源项目）。sgl-project (LMSYS-line academic open-source project).
国内使用Direct access: 大陆可直连direct OK　docs.sglang.ai 直连实测 200（2026-10-21）。docs.sglang.ai measured 200 direct (2026-10-21).
备案状态CAC filing: 未见于备案名单not in filing list：未见于网信办公示名单（2026-10-21 检索）；开源项目未面向大陆公众提供服务，备案不适用。not in the CAC public registry (searched 2026-10-21); the open-source project does not serve the mainland public, so filing does not apply.

备案信息以网信办《生成式人工智能服务已备案信息》公示名单为准（按批次滚动更新，本站只抄编号不解读）；「大陆可直连」是本站从大陆网络直连实测的结果，不是厂商承诺。GitHub 类站点当日实测有波动，口径写在依据里。Filing status is quoted from the CAC public registry (updated batch by batch; this site copies the number and does not interpret it). "Direct access" is this site's own reachability measurement from a mainland network, not a vendor promise. GitHub-hosted sites fluctuated on the measurement day — the note says so.

固定坐标系Fixed coordinate system

8 个维度dimensions, 与同分区其他对象逐项可比each comparable item by item against others in the same section.

模型与开放条件Model & openness: safetensors/HF 系权重为主；支持模型清单以官方 docs 为准，未逐条清点。官方定位含「large language models and multimodal models」（仓库描述直抓所见）。
safetensors/HF-line weights are the mainstay; the supported-model list follows the official docs, not inventoried item by item. The official positioning covers "large language models and multimodal models" (repo description, direct-fetched).
运行位置Where it runs: 服务化推理框架：RadixAttention 前缀缓存复用是官方招牌（自报数字在 models 站 guides serving-deep-dive.md 已核验快照，出处同源）；GitHub 38,370★（API 核验 2026-10-21）。
A serving inference framework: RadixAttention prefix-cache reuse is the official signature (the self-reported figures were verified as a snapshot in the models-site guides serving-deep-dive.md, same source); GitHub 38,370★ (API verified 2026-10-21).
本地文件Local files: 自托管部署，数据在自有环境；未见官方托管云线——第三方云托管不算官方口径。
Self-hosted deployment, data in your own environment; no official managed cloud line seen — third-party cloud hosting does not count as official.
关机后的任务Tasks after shutdown: 产品形态就是常驻服务（OpenAI 兼容 server）；agent 多轮场景的前缀复用是叙事重点。
The product form is a resident service (an OpenAI-compatible server); prefix reuse in agent multi-turn scenarios is the narrative focus.
工具与扩展Tools & extensions: OpenAI 兼容 API 含 tool calling；结构化输出（官方语言特性线）在文档内；MCP 支持未逐条核验。
The OpenAI-compatible API includes tool calling; structured output (the official language-features line) is in the docs; MCP not verified item by item.
上下文与记忆Context & memory: RadixAttention 对多轮/共享前缀的上下文复用是主轴；细节未逐条核验。
RadixAttention's reuse of context across multi-turn/shared prefixes is the main axis; details not verified item by item.
权限与限制Permissions & limits: 自托管无账号体系；API 鉴权由部署方自配；多租户/隔离面未逐条核验。
Self-hosted with no account system; API authentication configured by the deploying party; multi-tenancy/isolation surfaces not verified item by item.
适合什么任务Best for: 适合：agent 多轮调用/共享 system prompt 的服务化场景（前缀复用收益直接）、要结构化输出语言特性。不适合：单机开箱跑量化档；要最大生态兼容面（vLLM 生态更宽）。
Fits: serving scenarios with agent multi-turn calls/shared system prompts (prefix-reuse gains are direct), wanting structured-output language features. Does not fit: single-machine out-of-the-box quantized tiers; wanting the widest ecosystem-compatibility surface (vLLM's ecosystem is broader).

头号误解Top misconception

把 RadixAttention 自报数字当本站实测——官方口径快照，本站不实测Taking RadixAttention vendor numbers as our own benchmarks — official-claim snapshot, no local benchmarking
与 vLLM 二选一时只看吞吐——agent 多轮场景前缀复用才是 SGLang 的差异点Choosing between it and vLLM on raw throughput alone — prefix reuse is the differentiator for agent multi-turn workloads
把 LMSYS 项目当 LMSYS 产品——SGLang 是独立项目（sgl-project），Chatbot Arena 是另一条线Confusing the LMSYS project with an LMSYS product — SGLang is its own project (sgl-project); Chatbot Arena is a separate line

未知项清单Known unknowns

支持模型全清单（未逐条清点）
MCP 支持状态（未逐条核验）
结构化输出语言（SGLang 官方 DSL）的当前稳定度（未逐条核验）

The full supported-model list (not inventoried item by item)
MCP support status (not verified item by item)
The current stability of the structured-output language (SGLang's official DSL) (not verified item by item)

证据来源Evidence sources

判断可回到以下一手源复核。本站核验日Every judgement can be rechecked against the primary sources below. Verified on 2026-10-30，内容更新日content updated on 2026-10-21。

类型Type	名称Name	链接Link
repo	sgl-project/sglang · 仓库（Apache-2.0；38,370★，最近推送 2026-10-30，GitHub API 核验）sgl-project/sglang · repo (Apache-2.0; 38,370★, last push 2026-10-30, GitHub API verified)	https://github.com/sgl-project/sglang
docs	SGLang · 官方文档SGLang · official docs	https://docs.sglang.ai
changelog	sgl-project/sglang · Releases（changelog）sgl-project/sglang · Releases (changelog)	https://github.com/sgl-project/sglang/releases

实测记录Measured runs

本档未跑吞吐与显存实测；结论只覆盖运行时能力与许可。本档案复用同一快照出处。运行器专用实测协议本站未立。

This site has not completed measured runs. The models-site guides serving-deep-dive.md has already verified SGLang's official framing (RadixAttention self-reported figures); this dossier reuses the same snapshot source. No runner-specific test protocol exists on this site yet.