RIDE-Lab.github.io/index.html at main · RIDE-Lab/RIDE-Lab.github.io · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
<!doctype html>
<html lang="en">
<head>
  <meta charset="utf-8">
  <meta name="viewport" content="width=device-width, initial-scale=1">
  <meta name="description" content="RIDE Lab builds an open, agent-native control plane above LLM inference engines: workflow, SLO, routing, placement, state, quality, and evidence.">
  <meta name="theme-color" content="#0a0f10">
  <link rel="canonical" href="https://ride-lab.github.io/">
  <meta property="og:type" content="website">
  <meta property="og:title" content="RIDE Lab | Agent-Native LLM Inference Control Plane">
  <meta property="og:description" content="An open control plane that turns agent workflow intent and live telemetry into bounded inference-fleet decisions.">
  <meta property="og:url" content="https://ride-lab.github.io/">
  <title>RIDE Lab | Agent-Native LLM Inference Control Plane</title>
  <link rel="icon" href="./assets/ride-mark.svg" type="image/svg+xml">
  <link rel="preconnect" href="https://fonts.googleapis.com">
  <link rel="preconnect" href="https://fonts.gstatic.com" crossorigin>
  <link href="https://fonts.googleapis.com/css2?family=Fira+Code:wght@400;500;600&family=Inter:wght@400;500;600;700;800;900&family=Sora:wght@600;700;800&display=swap" rel="stylesheet">
  <link rel="stylesheet" href="./assets/site.css?v=public-portfolio-20260825">
</head>
<body data-title-en="RIDE Lab | Agent-Native LLM Inference Control Plane" data-title-zh="RIDE Lab｜智能体原生大模型推理控制平面">
  <header class="site-header">
    <a class="brand" href="#top" aria-label="RIDE Lab home"><span class="brand-mark">R</span><span class="brand-copy">RIDE Lab<small data-en="Agent-native inference control" data-zh="智能体原生推理控制">Agent-native inference control</small></span></a>
    <nav aria-label="Primary navigation">
      <a href="#boundary" data-en="Boundary" data-zh="边界">Boundary</a><a href="#decisions" data-en="Decisions" data-zh="决策">Decisions</a><a href="#portfolio" data-en="Portfolio" data-zh="课题">Portfolio</a><a href="#evidence" data-en="Evidence" data-zh="证据">Evidence</a><a href="./members.html" data-en="People" data-zh="成员">People</a><a class="nav-action" href="https://github.com/RIDE-Lab">GitHub ↗</a>
    </nav>
    <button id="lang-toggle" class="lang-toggle" type="button" aria-label="切换到中文">中 / EN</button>
  </header>

  <main id="top">
    <section class="hero control-hero" aria-labelledby="hero-title">
      <div class="hero-grid" aria-hidden="true"></div><div class="hero-orbit orbit-one" aria-hidden="true"></div><div class="hero-orbit orbit-two" aria-hidden="true"></div>
      <div class="hero-core" aria-hidden="true"><span>DECIDE</span><i></i><i></i><i></i></div>
      <div class="hero-copy">
        <p class="signal"><span></span><span data-en="Observe → decide → actuate → verify" data-zh="观测 → 决策 → 执行 → 验证">Observe → decide → actuate → verify</span></p>
        <h1 id="hero-title">RIDE<span>Control</span></h1>
        <p class="hero-kicker" data-en="An open, agent-native control plane above LLM inference engines." data-zh="建设大模型推理引擎之上的开放、智能体原生控制平面。">An open, agent-native control plane above LLM inference engines.</p>
        <p class="hero-summary" data-en="Agent workflows expose DAG state, tool waits, deadlines, quality bounds, reuse, and recovery intent. RIDE turns those signals and live telemetry into bounded fleet actions." data-zh="智能体工作流暴露 DAG 状态、工具等待、截止时间、质量边界、复用与恢复意图；RIDE 将这些信号和实时遥测转化为有边界的集群动作。">Agent workflows expose DAG state, tool waits, deadlines, quality bounds, reuse, and recovery intent. RIDE turns those signals and live telemetry into bounded fleet actions.</p>
        <div class="actions"><a class="button primary" href="#boundary"><span data-en="See the system boundary" data-zh="查看系统边界">See the system boundary</span><span>↓</span></a><a class="button quiet" href="#portfolio"><span data-en="Explore the portfolio" data-zh="查看课题组合">Explore the portfolio</span><span>↘</span></a></div>
      </div>
      <dl class="hero-metrics"><div><dt>04</dt><dd data-en="decision domains" data-zh="决策域">decision domains</dd></div><div><dt>07</dt><dd data-en="incubating research lines" data-zh="孵化研究线">incubating research lines</dd></div><div><dt>01</dt><dd data-en="evidence contract" data-zh="统一证据合同">evidence contract</dd></div></dl>
    </section>

    <nav class="scope-rail" aria-label="Homepage sections">
      <a href="#boundary" data-code="01" data-en="System boundary" data-zh="系统边界">System boundary</a><a href="#decisions" data-code="02" data-en="Decision domains" data-zh="决策域">Decision domains</a><a href="#portfolio" data-code="03" data-en="Research portfolio" data-zh="研究课题">Research portfolio</a><a href="#evidence" data-code="04" data-en="Evidence contract" data-zh="证据合同">Evidence contract</a><a href="#ecosystem" data-code="05" data-en="Technical homes" data-zh="技术归属">Technical homes</a>
    </nav>

    <section id="boundary" class="section boundary-section">
      <div class="section-head"><div><p class="eyebrow" data-en="System boundary" data-zh="系统边界">System boundary</p><h2 data-en="Engines execute. The control plane decides." data-zh="引擎负责执行，控制平面负责决策。">Engines execute. The control plane decides.</h2></div><p data-en="RIDE is not another inference engine. It is the policy, coordination, and evidence layer that turns workload intent and live telemetry into bounded actions on an engine fleet." data-zh="RIDE 不是另一个推理引擎，而是把工作负载意图和实时遥测转化为有边界执行动作的策略、协调与证据层。">RIDE is not another inference engine. It is the policy, coordination, and evidence layer that turns workload intent and live telemetry into bounded actions on an engine fleet.</p></div>
      <div class="control-stack" role="img" aria-label="Applications and workflows feed the RIDE control plane, which actuates inference engines and infrastructure">
        <article class="stack-layer workload"><span>03 / INTENT</span><strong data-en="Agents, workflows, users" data-zh="智能体、工作流与用户">Agents, workflows, users</strong><p data-en="DAG state, deadlines, quality targets, budgets, reuse signals, and service policy." data-zh="DAG 状态、截止时间、质量目标、预算、复用信号与服务策略。">DAG state, deadlines, quality targets, budgets, reuse signals, and service policy.</p></article>
        <article class="stack-layer control"><span>02 / RIDE</span><strong data-en="Inference control plane" data-zh="推理控制平面">Inference control plane</strong><p data-en="Observe, predict, route, admit, place, budget, recover, and verify." data-zh="观测、预测、路由、准入、放置、预算、恢复与验证。">Observe, predict, route, admit, place, budget, recover, and verify.</p></article>
        <article class="stack-layer engine"><span>01 / EXECUTE</span><strong data-en="LLM inference engines" data-zh="大模型推理引擎">LLM inference engines</strong><p data-en="vLLM-HUST, vLLM, SGLang, and compatible serving backends execute prefill, decode, cache, and communication paths." data-zh="vLLM-HUST、vLLM、SGLang 及兼容后端执行 prefill、decode、缓存与通信路径。">vLLM-HUST, vLLM, SGLang, and compatible serving backends execute prefill, decode, cache, and communication paths.</p></article>
        <article class="stack-layer hardware"><span>00 / RESOURCE</span><strong data-en="Accelerators, memory, network" data-zh="加速器、内存与网络">Accelerators, memory, network</strong><p data-en="The data plane remains independently measurable and replaceable." data-zh="数据面保持可独立测量、可替换。">The data plane remains independently measurable and replaceable.</p></article>
      </div>
      <div class="boundary-note"><strong data-en="Boundary rule" data-zh="边界规则">Boundary rule</strong><p data-en="A RIDE mechanism must be expressible as a decision over observable state. Engine patches are allowed only as minimal reusable telemetry or actuation seams; kernels, KV transport, compilation internals, and hardware backends do not become control-plane work merely because they expose a flag." data-zh="RIDE 机制必须能表述为基于可观测状态的决策。引擎补丁只能是最小、可复用的遥测或执行接口；内核、KV 传输、编译内部和硬件后端不会因为暴露了一个开关就变成控制平面课题。">A RIDE mechanism must be expressible as a decision over observable state. Engine patches are allowed only as minimal reusable telemetry or actuation seams; kernels, KV transport, compilation internals, and hardware backends do not become control-plane work merely because they expose a flag.</p></div>
    </section>

    <section id="decisions" class="section decision-section">
      <div class="section-head"><div><p class="eyebrow" data-en="Decision domains" data-zh="决策域">Decision domains</p><h2 data-en="Control is a closed loop, not a dashboard." data-zh="控制平面是闭环，不是仪表盘。">Control is a closed loop, not a dashboard.</h2></div><p data-en="Each research line turns measurable state into a clear decision and a verifiable outcome." data-zh="每条研究线都把可测状态转化为明确决策与可验证结果。">Each research line turns measurable state into a clear decision and a verifiable outcome.</p></div>
      <div class="agenda-grid decision-grid">
        <article data-index="01"><b>WF</b><h3 data-en="Workflow & SLO" data-zh="工作流与 SLO">Workflow &amp; SLO</h3><p data-en="Schedule DAG stages, admit work, allocate capacity, and protect fairness along the critical path." data-zh="围绕关键路径调度 DAG 阶段、准入请求、分配容量并保护公平性。">Schedule DAG stages, admit work, allocate capacity, and protect fairness along the critical path.</p><small>ADMISSION · DEADLINE · FAIRNESS</small></article>
        <article data-index="02"><b>RP</b><h3 data-en="Routing & placement" data-zh="路由与放置">Routing &amp; placement</h3><p data-en="Choose a backend, cache holder, model replica, adapter, or continuation target from live fleet state." data-zh="依据实时集群状态选择后端、缓存持有者、模型副本、适配器或续算目标。">Choose a backend, cache holder, model replica, adapter, or continuation target from live fleet state.</p><small>SELECT · PLACE · RECOVER</small></article>
        <article data-index="03"><b>KV</b><h3 data-en="State & KV" data-zh="状态与 KV">State &amp; KV</h3><p data-en="Predict reuse, materialize or recompute state, select tiers, and account for transfer and survival horizons." data-zh="预测复用，选择状态物化或重算、分层位置，并计入传输与剩余生存时间。">Predict reuse, materialize or recompute state, select tiers, and account for transfer and survival horizons.</p><small>REUSE · MATERIALIZE · TIER</small></article>
        <article data-index="04"><b>QE</b><h3 data-en="Quality & economics" data-zh="质量与经济性">Quality &amp; economics</h3><p data-en="Trade quality, latency, cost, safety, and service risk without hiding failures behind aggregate throughput." data-zh="在质量、时延、成本、安全和服务风险之间决策，不用汇总吞吐掩盖失败。">Trade quality, latency, cost, safety, and service risk without hiding failures behind aggregate throughput.</p><small>QUALITY · COST · GUARDRAIL</small></article>
      </div>
    </section>

    <section id="portfolio" class="section systems-section portfolio-section">
      <div class="section-head"><div><p class="eyebrow" data-en="Control-plane portfolio" data-zh="控制平面课题">Control-plane portfolio</p><h2 data-en="Independent questions. One control-plane architecture." data-zh="独立课题，共用控制平面架构。">Independent questions. One control-plane architecture.</h2></div><p data-en="Each project retains its own owner, implementation, service, manifest, and results. RIDE provides the shared architectural family; repository links appear only after they are public." data-zh="每个课题保留独立负责人、实现、服务、manifest 与结果。RIDE 提供共同架构；仓库公开后才展示链接。">Each project retains its own owner, implementation, service, manifest, and results. RIDE provides the shared architectural family; repository links appear only after they are public.</p></div>
      <div class="portfolio-grid" data-project-portfolio data-source="./data/control-plane-projects.json?v=public-portfolio-20260825">
        <p class="project-status" data-project-status data-en="Loading the portfolio…" data-zh="正在加载课题组合……">Loading the portfolio…</p>
      </div>
    </section>

    <section id="evidence" class="section method-section">
      <div class="section-head"><div><p class="eyebrow" data-en="Decision lifecycle" data-zh="决策生命周期">Decision lifecycle</p><h2 data-en="From workload intent to accountable execution." data-zh="从工作负载意图到可追踪执行。">From workload intent to accountable execution.</h2></div><p data-en="RIDE connects observable workflow and system state to engine actions, then records outcomes for operations and evaluation." data-zh="RIDE 把可观测的工作流与系统状态连接到引擎动作，并记录结果以支持运维与评估。">RIDE connects observable workflow and system state to engine actions, then records outcomes for operations and evaluation.</p></div>
      <div class="method-path"><article><span>01 / OBSERVE</span><strong data-en="Build live context" data-zh="构建实时上下文">Build live context</strong><p data-en="Workload, topology, engine, model, queue, and resource signals." data-zh="汇聚工作负载、拓扑、引擎、模型、队列与资源信号。">Workload, topology, engine, model, queue, and resource signals.</p></article><article><span>02 / DECIDE</span><strong data-en="Choose an action" data-zh="选择控制动作">Choose an action</strong><p data-en="Admission, routing, placement, budgets, state, quality, and recovery." data-zh="覆盖准入、路由、放置、预算、状态、质量与恢复。">Admission, routing, placement, budgets, state, quality, and recovery.</p></article><article><span>03 / ACTUATE</span><strong data-en="Use engine connectors" data-zh="通过引擎连接器执行">Use engine connectors</strong><p data-en="Apply decisions through versioned telemetry and control interfaces." data-zh="通过版本化遥测与控制接口落实决策。">Apply decisions through versioned telemetry and control interfaces.</p></article><article><span>04 / REPORT</span><strong data-en="Report outcomes" data-zh="报告执行结果">Report outcomes</strong><p data-en="Correctness, latency, goodput, resource use, and fallback status." data-zh="报告正确性、时延、有效吞吐、资源使用与回退状态。">Correctness, latency, goodput, resource use, and fallback status.</p></article></div>
    </section>

    <section id="leadership" class="section leadership-section">
      <div class="leadership-card"><div><p class="eyebrow" data-en="Program leadership" data-zh="方向负责人">Program leadership</p><h2 data-en="Built as a coherent systems program." data-zh="以一致的系统研究计划推进。">Built as a coherent systems program.</h2></div><div><strong>Mao Yancan · <a href="https://github.com/yancanmao">yancanmao ↗</a></strong><p data-en="RIDE-Lab operations and Inference Control Plane program lead. Project-level academic ownership remains with each named owner." data-zh="RIDE-Lab 运营与推理控制平面方向负责人。各课题的学术责任仍由对应负责人独立承担。">RIDE-Lab operations and Inference Control Plane program lead. Project-level academic ownership remains with each named owner.</p><small data-en="Program coordination connects independent projects without replacing project-level accountability." data-zh="方向协调连接各独立课题，但不替代课题级责任。">Program coordination connects independent projects without replacing project-level accountability.</small></div></div>
    </section>

    <section id="ecosystem" class="section ecosystem-section">
      <div class="section-head"><div><p class="eyebrow" data-en="Technical homes" data-zh="技术归属">Technical homes</p><h2 data-en="One stack. Explicit ownership." data-zh="同一技术栈，明确技术归属。">One stack. Explicit ownership.</h2></div><p data-en="RIDE provides the control-plane architecture while each project retains the repository and organization that best match its technical ownership." data-zh="RIDE 提供控制面架构；各课题保留最符合其技术归属的仓库与组织。">RIDE provides the control-plane architecture while each project retains the repository and organization that best match its technical ownership.</p></div>
      <div class="architecture"><a href="https://lab.sage.org.ai/"><span>00</span><strong>IntelliStream</strong><small data-en="Cross-layer research incubator" data-zh="跨层研究孵化器">Cross-layer research incubator</small></a><a class="active" href="#top"><span>01</span><strong>RIDE Lab</strong><small data-en="Inference control plane" data-zh="推理控制平面">Inference control plane</small></a><a href="https://sage.org.ai/"><span>02</span><strong>SAGE</strong><small data-en="Programmable agent and workflow surface" data-zh="可编程智能体与工作流界面">Programmable agent and workflow surface</small></a><a href="https://vllm-hust.sage.org.ai/"><span>03</span><strong>vLLM-HUST</strong><small data-en="Inference engine and hardware execution" data-zh="推理引擎与硬件执行">Inference engine and hardware execution</small></a></div>
      <div class="shared-line"><strong>CONTROL LOOP</strong><i></i><p data-en="intent → telemetry → decision → engine action → evidence" data-zh="意图 → 遥测 → 决策 → 引擎动作 → 证据">intent → telemetry → decision → engine action → evidence</p></div>
    </section>
  </main>

  <footer><div><strong>RIDE Lab</strong><span data-en="Agent-native control above the engine." data-zh="在引擎之上构建智能体原生控制。">Agent-native control above the engine.</span></div><nav><a href="./members.html" data-en="People" data-zh="成员">People</a><a href="https://github.com/RIDE-Lab">GitHub</a><a href="https://github.com/RIDE-Lab/.github/blob/main/profile/README.md" data-en="About" data-zh="关于">About</a></nav></footer>
  <script src="./assets/site.js?v=public-portfolio-20260825"></script>
  <script src="./assets/portfolio.js?v=public-portfolio-20260825"></script>
</body>
</html>