2026-09-12
每日一文 · 长文精读
NVIDIA Personal AI Router Distributes AI Tasks across Local Compute
NVIDIA 个人 AI 路由器：在本地计算设备间分发 AI 任务
作者：Sergio De Simone · InfoQ 原文
摘要：NVIDIA 推出 Personal AI Router (PAIR) 测试版，可将本地网络中多台计算机的推理能力整合，自动分发 AI 请求。它主要针对本地多智能体工作负载，解决单个 GPU 的瓶颈问题。PAIR 无缝集成 Ollama 和 LM Studio，通过代理接收请求并选择合适节点执行。演示显示结合多设备可缩短完成时间约 2 倍。PAIR 支持 Windows、Linux、macOS 及不同架构，但不合并 GPU 或 VRAM。
NVIDIA Personal AI Router (PAIR), now available in beta, lets you combine the inference capacity of multiple computers on your local network and automatically distribute AI requests among them.
NVIDIA 个人 AI 路由器 (PAIR) 现已推出测试版，它允许你整合本地网络中多台计算机的推理能力，并自动在它们之间分发 AI 请求。
It is primarily designed for local multi-agent AI workloads, where multiple independent model calls can otherwise overwhelm one GPU.
它主要针对本地多智能体 AI 工作负载设计，在这种场景下，多个独立的模型调用可能会使单个 GPU 不堪重负。
NVIDIA says a breadth-first approach to distributing agentic tasks is becoming increasingly common, with a lead agent dispatching subtasks to sub-agents or multiple agents working together to complete more complex tasks.
NVIDIA 表示，广度优先分发智能体任务的方式正变得越来越普遍，即由主智能体将子任务分派给子智能体，或多个智能体协同完成更复杂的任务。
However, this approach can create a bottleneck on the local GPU when it receives too many requests.
然而，当本地 GPU 收到过多请求时，这种方法可能会造成瓶颈。
To address this challenge, NVIDIA PAIR maximizes the AI compute available locally by distributing individual inference requests across available systems.
为了应对这一挑战，NVIDIA PAIR 通过将单个推理请求分发到可用系统上，最大化本地可用的 AI 计算能力。
It integrates seamlessly with popular local inference services such as Ollama and LM Studio without requiring changes to the underlying architecture or agent harness.
它与 Ollama 和 LM Studio 等流行的本地推理服务无缝集成，无需更改底层架构或智能体框架。
Agents can send a request through the familiar local interface it expects.
智能体可以通过它期望的熟悉本地接口发送请求。
PAIR receives the request through its proxy, identifies its engine and model requirements, and selects one eligible node.
PAIR 通过其代理接收请求，识别请求的引擎和模型要求，并选择一个符合条件的节点。
That node executes the request from start to finish and sends the response back through PAIR.
该节点从头到尾执行请求，并通过 PAIR 发回响应。
The agent continues to see one connection while PAIR handles placement behind it.
智能体始终只看到一个连接，而 PAIR 在后台处理分发。
To demonstrate PAIR’s capabilities, NVIDIA released a demo combining Hermes Desktop, Ollama, and PAIR showing roughly a 2x reduction in completion time when combining an RTX Spark, a DGX Spark, and an RTX 5090 via PAIR compared with running the workload on a single RTX Spark laptop.
为了展示 PAIR 的能力，NVIDIA 发布了一个结合 Hermes Desktop、Ollama 和 PAIR 的演示，显示通过 PAIR 组合 RTX Spark、DGX Spark 和 RTX 5090 时，完成时间相比在单个 RTX Spark 笔记本电脑上运行工作负载大约减少了 2 倍。
In the demo, Hermes breaks the task into five independent specialist analyses, delegates them, reconciles their findings, and synthesizes the final plan—covering what needs to happen tonight, this week, later, or not at all.
在演示中，Hermes 将任务分解为五个独立的专业分析，分派它们，协调它们的发现，并综合出最终计划——涵盖今晚、本周、以后或根本不需要做的事情。
PAIR handles the distribution of these inference requests across available nodes, while Ollama runs the model on whichever node PAIR selects.
PAIR 负责将这些推理请求分发到可用节点，而 Ollama 在 PAIR 选择的任何节点上运行模型。
NVIDIA notes, however, that the demo should not be taken as a performance guarantee, as results depend on multiple factors including workload parallelism, model, engine settings, hardware, network, and node availability.
不过，NVIDIA 指出，该演示不应被视为性能保证，因为结果取决于多种因素，包括工作负载并行性、模型、引擎设置、硬件、网络和节点可用性。
NVIDIA PAIR can be used on Windows 11, Linux, and macOS, with support for both x64 and arm64 systems.
NVIDIA PAIR 可用于 Windows 11、Linux 和 macOS，支持 x64 和 arm64 系统。
It can also pair nodes running different operating systems, dispatching a task to a given node only if the required model or engine is known to be compatible with that node.
它还可以配对运行不同操作系统的节点，仅当所需模型或引擎已知与该节点兼容时，才将任务分派给该节点。
NVIDIA explicitly notes that PAIR does not "merge GPUs or pool VRAM into one larger accelerator".
NVIDIA 明确表示，PAIR 不会“合并 GPU 或将 VRAM 池化为一个更大的加速器”。
Instead, it distributes individual inference requests across available systems.
相反，它将单个推理请求分发到可用系统上。
Despite these disclaimers, NVIDIA's announcement sparked some confusion on social media, with users interpreting PAIR as a solution for sharing their available compute with third-parties or for running complex models by combining less capable compute.
尽管有这些免责声明，NVIDIA 的公告仍在社交媒体上引发了一些困惑，用户将 PAIR 解读为一种与第三方共享可用计算资源或通过组合性能较低的计算设备来运行复杂模型的解决方案。
Reddit user Vegetable-Warthog81 described their positive experience with PAIR using it to distribute inference across three RTX 5090s running Qwen 3.8 27B using Ollama:
Reddit 用户 Vegetable-Warthog81 描述了他们对 PAIR 的积极体验，他们使用 PAIR 将推理分发到三台运行 Qwen 3.8 27B（使用 Ollama）的 RTX 5090 上：
PAIR makes distributing jobs across the three machines pretty painless.
PAIR 使得在三个机器之间分发任务变得非常轻松。
For long, repetitive “grunt work” where I care more about stability and just keeping all the GPUs busy than squeezing out maximum tokens/sec, it’s been surprisingly nice.
对于长期、重复的“苦力活”，我更关心稳定性和让所有 GPU 保持忙碌，而不是榨取最大 tokens/秒，它出奇地好用。
PAIR can be downloaded from GitHub.
PAIR 可以从 GitHub 下载。
For a step-by-step introduction to how to use it, see the Getting Started documentation.
有关如何使用它的分步介绍，请参阅入门文档。
If you are looking for a platform that enables GPU compute to be shared across a network of parties, check Petals or Mesh LLM.
如果你正在寻找一个能够在多方网络间共享 GPU 计算的平台，可以查看 Petals 或 Mesh LLM。
Mesh LLM also supports splitting models that are too large to fit on a single machine using Skippy.
Mesh LLM 还支持使用 Skippy 拆分那些太大而无法放入单台机器的模型。
阅读理解
1. What is the primary purpose of NVIDIA PAIR?
2. How does PAIR handle a request from an agent?
3. What was the result of the demo combining Hermes Desktop, Ollama, and PAIR?