Tool · 2026-07-25

vLLM v0.26.0

vLLM v0.26.0 ships 411 commits from 212 contributors, headlined by a full support stack for the new Inkling model family.

vLLM Releases2026-07-25About 3 min readWorth tracking

Marked important, but published more than 30 days ago, so it moved down to this band.

Inference & deployment
On this page13 sections

The full write-up and the editors' notes are in Chinese; the summary above is from the original announcement.

Release line

This signal has a follow-up

#

You are reading an earlier release on this line. 3 more releases follow; the latest is “v0.32.0”.

  1. 2026-06-30vLLM v0.24.0
  2. 2026-07-11vLLM v0.25.0
  3. 2026-07-25vLLM v0.26.0Current
  4. 2026-08-10vLLM v0.27.0
  5. 2026-09-22v0.30.0
  6. 2026-10-18v0.32.0Latest

承接 v0.25.0 的 Model Runner V2 默认化,继续往架构适配走。

Record

The signal

#

这次发布做了什么

vLLM v0.26.0 汇集了 212 位贡献者(其中 61 位是新面孔)的 411 个提交,是一次体量很大的版本。发布说明把「新增 Inkling 模型家族的完整支持栈」放在了第一条。

为什么是 Inkling

Inkling 是 Thinking Machines 在 7 月中旬开源的万亿参数级多模态 MoE 模型——本站已经把它作为当日头条信号收录。开源权重和「能在主流推理引擎上跑起来」之间,通常还隔着几周到几个月的适配工作:注意力核、图捕获策略、量化路径、并行切分,每一项都要针对具体架构重写。v0.26.0 把这段距离压缩到了十天左右。

支持栈包含四个层次:

  • 基础建模:模型结构本身在 vLLM 内的实现;
  • 分段 CUDA Graph(piecewise CUDA graph):把前向过程切成可以分别捕获的片段,让 MoE 这类含动态路由的结构也能吃到 CUDA Graph 的调度收益,而不必整图捕获;
  • Hopper FA4 相对注意力:针对 H100 一代硬件的注意力核实现;
  • MTP=1 推测解码:多 token 预测形态的推测解码,用小步长的草稿换取解码吞吐。

值得注意的信号

把这四条放在一起看,重点不是「又一个模型被支持了」,而是开源推理栈对新架构的适配周期正在明显缩短,并且适配是成套的——不是先能跑通、几个版本后才补上性能路径。对于要在本地或自有集群上评估新开源模型的团队,这直接改变了「等多久才能真正测起来」的预期。

Why now

Why it matters

#

开源万亿参数模型从「放出权重」到「主流推理引擎上可用」的间隔被压到了十天量级,而且是基础建模、图捕获、注意力核、推测解码一次配齐。这改变了自建推理团队评估新模型的排期假设。

Background

Technical context

#

分段 CUDA Graph 解决的是 MoE 动态路由与整图捕获冲突的问题;FA4 相对注意力是面向 Hopper 架构的注意力核;MTP=1 推测解码属于多 token 预测形态,用草稿模型的小步长预测换取解码阶段吞吐。三者叠加才构成一条可用的生产推理路径。

At your level

Have AI explain this signal at your level

Pick your level of experience and AI writes an explanation for it on the spot.

Audience

Who should care

#
自建或自运维推理服务的平台团队需要在本地评估新开源模型的算法与评测团队关注 MoE 与推测解码落地成本的架构师
Areas it may affect
推理服务容量规划开源模型选型周期MoE 架构的部署成本解码吞吐优化

Next

Learning path

#
  1. 先理解 MoE 架构为什么让图捕获和量化变复杂
  2. 再看推测解码如何用草稿模型换解码吞吐
  3. 最后结合自有硬件(是否 Hopper)判断这套支持栈能吃到多少收益

Getting started

Have AI draft a learning path

From the related knowledge and skills on this page, AI drafts a path from the basics to applying it.

Network

Where it sits in the network

#

How this signal connects to neighbouring technologies, skills and background knowledge. Follow any link to keep exploring.

Comparison

Compare with another technology

Pick another published technology and AI writes a comparison of what they share, how they differ and when to use each.

Further thought

Follow-up questions

#
  • 在非 Hopper 硬件上,这套支持栈能保留多少性能收益?
  • MTP=1 推测解码对长上下文任务的加速比是多少?
  • 分段 CUDA Graph 是否会成为后续 MoE 模型的默认适配路径?

Publication record

Featured in

#

The briefings that carried this signal, newest first; the last row is the weekly review of its week. The tag is the section it ran in.

  1. 技术简报 - 2026-07-27

    Immediate attention

Sources

vLLM Releases

vLLM ProjectOpen-source community2026-07-25
Open original sourcehttps://github.com/vllm-project/vllm/releases/tag/v0.26.0