Skip to content

README.md

Latest commit

 

History

History
113 lines (76 loc) · 9.02 KB

README.md

File metadata and controls

113 lines (76 loc) · 9.02 KB
Raw

InferMatrix

Open to Infer. Built to Evolve.

InferMatrix 开源推理生态展示网站。当前版本提供中文/英文切换、七个项目的分类筛选、架构图缩放查看、生态方向探索、视频资源专栏，以及原始海报下载。

A bilingual, buildless website showcasing seven open-source inference projects: vLLM-Omni, afd-plugin, vLLM Router, agent-infer, vllm-rlt, vllm-omni-mlx, and InferRSI.

本地运行 / Run locally

不需要 npm 安装、构建步骤或后端。使用 Python 3 在仓库根目录运行：

python3 -m http.server 8000 --bind 127.0.0.1

Open http://localhost:8000. The deployable files are already at the repository root: index.html, styles.css, assistant.css, app.js, showcase.js, inferrsi.js, assistant.js, video-resources.js, robots.txt, sitemap.xml, and assets/.

The stylesheet requests DM Sans and Manrope from Google Fonts. System-font fallbacks remain available when the external font service cannot be reached.

检查与维护 / Check and maintain

可选维护工具只依赖 Node.js 18+ 的内置模块，无需安装第三方包。

Optional maintenance checks use Node.js 18+ built-in modules only:

node --check app.js
node --check video-resources.js
node --check prepare-initial-content.cjs
node --check qa-interactions.cjs
node qa-interactions.cjs

The interaction suite reads the current index.html and runs app.js in a simulated DOM using Node's VM. It covers filtering, bilingual content, keyboard navigation, project/poster dialogs, focus return, motion preferences, blocked storage, and all six architecture viewers. This is a logic regression check; it does not replace real-browser visual or accessibility testing.

项目资料和中英文文案集中在 app.js。修改项目卡片内容后，重新生成 HTML 中供无 JavaScript 环境阅读的卡片：

After editing project content in app.js, refresh the pre-rendered no-JavaScript cards:

node prepare-initial-content.cjs
node qa-interactions.cjs

ASSET_PROVENANCE.json records the included image sources, dimensions, and SHA-256 hashes. SHA256SUMS covers every file in this source snapshot except the checksum file itself. Verify on Linux with sha256sum -c SHA256SUMS, or on macOS with shasum -a 256 -c SHA256SUMS. After intentional edits, refresh the relevant provenance hashes and checksum manifest.

网站地址 / Site address

Live websites: https://infermatrix.si for the brand website and https://recipes.infermatrix.si for the Recipes catalog. All three Recipes entries on the brand site link to the Recipes subdomain.

Both sites use Cloudflare Pages Free with GitHub automatic deployments. The Pages addresses remain available at https://infermatrix.pages.dev and https://infermatrix-recipes.pages.dev.

For the brand site, select framework None, leave the build command empty, and use the repository root (.) as the build output directory. For InferMatrixLab/recipe, use pnpm build, output directory out, Node.js 24, and pnpm 11.25.0. Leave NEXT_PUBLIC_BASE_PATH unset for the Recipes subdomain.

Bind each custom domain in its Pages project. The registrar must delegate infermatrix.si to the Cloudflare nameservers assigned to the zone before the custom domains can become active. Both sites are public and require no sign-in.

Latest additions

  • A bilingual Video resources column for external learning tutorials and project talks, with project/type filters and search. Eleven links cover vLLM-Omni, afd-plugin and vLLM Router, with available publisher information and verification sources retained.
  • A repo-grouped video showcase with five vLLM-Omni videos, local poster previews, click-to-play and fullscreen controls.
  • InferRSI as a seventh project and an evidence-driven optimization direction, with a brand-styled, scalable architecture SVG.
  • Qwen3-TTS and Drama 3 dedicated-engine examples with pinned, workload-specific C1/C64 audio-throughput comparisons. See INFERRSI.md and BENCHMARK_EVIDENCE.json; these are author-reported results, not independently rerun benchmarks.
  • MiniCPM-o 4.5 single-GPU streaming speech: H200 target configuration with explicitly labelled L20X source C1/C32 measurements and reproduction links. The all-metric 2× target is unmet; the measured r11 profile has a HiFT numerical failure and pending speech-quality gates.
  • Infer RSI in the focus strip and a fifth bilingual autonomous-inference-optimization explorer direction.
  • Public Recipes catalog links in the navigation and InferRSI section.

MIRA assistant guide

The header avatar and a small full-body floating sticker open the same bilingual project guide. The floating sticker appears after the header leaves the viewport; its dismiss control appears on mouse hover or keyboard focus and remains available on touch devices. Dismissal is remembered, and opening the guide from the header or footer restores the sticker.

This is an explicitly labeled local-content navigation demo, not a connected AI model. Replies use preset project material and source links. No model endpoint, API key, or chat logging service is configured. See LOCAL_PREVIEW.md for the optional dependency-free Node preview and interaction checklist.

视频资源 / Video resources

首页导航的「视频资源」指向 #video-resources，与「演示展示」分别维护。专栏收录学习教程、技术分享和宣讲回放的外部链接，点击后在新窗口打开原发布平台，不嵌入或下载视频。

资源清单集中在 assets/video-resources.json。当前收录 11 条：vLLM-Omni 8 条、afd-plugin 2 条、vLLM Router 1 条；其他四个项目暂未找到可核实的教程或宣讲回放。用户指定合集分集时保留 p 参数，并移除分享链接中的跟踪参数；可用 B 站官方分页元数据核实分集标题、讲者（标题明确标注时）与时长。每添加一条资源，填写以下格式；示例仅说明字段，不是已收录的视频：

{
  "id": "unique-video-id",
  "project": "omni",
  "type": "tutorial",
  "title": {"zh": "视频中文标题", "en": "Video title"},
  "description": {"zh": "介绍视频内容与适合的观众。", "en": "Describe the content and intended audience."},
  "url": "https://www.youtube.com/watch?v=REPLACE_WITH_VIDEO_ID",
  "platform": "YouTube",
  "publisher": "发布者或活动主办方",
  "speaker": "讲者姓名",
  "publishedAt": "2026-10-28",
  "duration": "35:20"
}

项目 ID：omni（vLLM-Omni）、afd（afd-plugin）、router（vLLM Router）、agent（agent-infer）、rlt（vllm-rlt）、mlx（vllm-omni-mlx）、inferrsi（InferRSI）。类型为 tutorial（学习教程）或 talk（技术分享 / 宣讲）。publisher、speaker、publishedAt、duration 可省略；链接必须使用 HTTPS。标题与简介提供中英文，列表按清单顺序显示。发布前核实链接、项目归属和视频内容，清单中不要保留示例占位链接。

每条已收录资源还保留 sourceUrl、verifiedAt 和 sourceNote，供维护者核对来源、核实日期及证据范围；有依据时记录 eventDate。这些维护字段不显示在视频卡片上。publishedAt 只填写已确认的视频上传日期，不能用会议日期或配套文章日期代替。中英文标题是便于浏览的编辑文案；原始标题与信息请以来源页面为准。首批资料来源包括 vLLM 官方 Office Hours、Conf42、Red Hat Developer、B 站原视频页面以及 vLLM 小课堂的原视频与课件索引。B 站的部分页面/API 会对自动抓取返回 412；其链接由原始页面索引或配套资料交叉核对，未逐条验证实际播放。

The library is rendered by video-resources.js. Project counts reflect the full catalog; results combine project, content-type, and text filters. Results are paginated at 6 videos per page, with bilingual previous/next controls and the current visible range. Changing a filter or search resets to page one; changing language preserves the current page and filters. Pagination is hidden for a single page, empty results, loading and errors. Search covers both languages, project names, platforms, publishers, and speakers. Empty catalogs, no matches, and failed requests have distinct states; failed requests can be retried.

The catalog is fetched with cache: 'no-store' and a fresh request parameter so edits to the JSON appear on the next page reload, even when the document or previous JSON response was cached. Changing only the page URL's query string does not invalidate a separately fetched resource. The script entry also carries a version parameter to refresh older copies of the loader.

素材与许可 / Assets and licensing

See ATTRIBUTION.md for asset-specific sources and licenses/ for the retained upstream licenses. The two official upstream diagrams retain their Apache-2.0 terms. The logo, poster, and poster-based vector redraws are not assigned an open-source license by this repository. No blanket license is asserted for all website code and branding.