Skip to content

Commit

Permalink
Merge pull request #1 from yuudiiii/patch-1
Browse files Browse the repository at this point in the history
  • Loading branch information
sparanoid authored Nov 4, 2024
2 parents 13e2bff + 02d737d commit bb65b8d
Showing 1 changed file with 7 additions and 6 deletions.
13 changes: 7 additions & 6 deletions src/components/HomepageFeatures/index.tsx
Original file line number Diff line number Diff line change
Expand Up @@ -10,27 +10,28 @@ type FeatureItem = {

const FeatureList: FeatureItem[] = [
{
title: "高效编程",
title: "高效的服务吞吐量",
Svg: require("@site/static/img/undraw_docusaurus_mountain.svg").default,
description: (
<>
vLLM 采用创新的内存管理与执行架构,大幅提升大模型推理的速度与效率
vLLM 采用独特的 PagedAttention 技术和动态批处理机制,且支持并行采样、波束搜索等多种解码算法,极大提升了服务吞吐量和响应速度
</>
),
},
{
title: "实时编译",
title: "内存管理大师",
Svg: require("@site/static/img/undraw_docusaurus_tree.svg").default,
description: (
<>支持高度并发的请求处理,vLLM 可同时服务数千用户,提升吞吐量和响应速度</>
<>创新的内存管理与执行架构,通过将 kv 缓存分割为若干块进行精细管理,把内存浪费控制在序列的最后一块,能实现接近最优的内存使用且支持内存共享,浪费比例低至不到 4%
</>
),
},
{
title: "灵活的迭代空间结构",
title: "灵活易用",
Svg: require("@site/static/img/undraw_docusaurus_react.svg").default,
description: (
<>
兼容多种深度学习框架,vLLM 易于集成到现有机器学习管道,部署更便捷
vLLM 可无缝集成各类模型,兼容 NVIDIA、AMD、Intel 等多种硬件平台 GPU 和 CPU,并提供简洁的接口和文档,易于上手
</>
),
},
Expand Down

0 comments on commit bb65b8d

Please sign in to comment.