Skip to content

Router that virtually distributes inference across connected devices in the home.

Go
3.6 KB
Raw Blame
---
description: Backend-owned proxy routing and inference logging rules
alwaysApply: true
---
<!--
SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
SPDX-License-Identifier: Apache-2.0
-->

# Proxy and Inference Routing

Routing is owned by `nvpair-proxy` (one process hosting a facade per engine,
addressed by clients as `ollama-proxy:` / `lmstudio-proxy:` / `llamacpp-proxy:`),
`nvpair-job-scheduler`, and `nvpair-ui-broker`.

For model-bearing inference, each facade first filters a request-local discovery
snapshot to nodes whose per-engine inventory advertises the requested model.
Empty and non-matching inventories are ineligible; an empty owner set returns a
local `502` without contacting an engine. Ollama normalizes the implicit
`:latest` tag; LM Studio and llama.cpp IDs match exactly.

Within that eligible owner set, proxy target precedence is:

1. explicit manual selection;
2. the scheduler priority list;
3. deterministic proxy default ordering.

An ineligible manual selection cannot override the capability gate. Retryable
failover, including an upstream model `404` caused by stale positive inventory,
continues only through other advertised owners and never broadens to excluded
nodes.

The scanner and manual-node worker feed compact maximum-GPU telemetry into the
broker's source-aware cache; scanner observations take precedence. The broker
ages and relays this feed to `nvpair-job-scheduler`.

The scheduler smooths fresh utilization into pressure 0–3, treats invalid,
missing, or older-than-10-second telemetry as neutral pressure 1, and ranks by
`pending + gpuPressure`, then pressure, then stable UUID. It emits
`schedule:priority` snapshots containing order, pending counts, and pressure.
Each facade chooses with `pending + gpuPressure + localReservations`, so
concurrent bursts do not wait for workload feedback.

Those reservations are **process-wide, shared by every facade**. Two facades
bursting at once compete for the same node's GPU, so a dispatch through either
must be visible to the other; this is the reason the engines share one process.
A reservation is stamped with the snapshot generation it was counted against,
released when its request ends, and moved to the node a failover actually
lands on. A snapshot supersedes the reservations taken before it, so a release
arriving after one is ignored rather than double-counted.

## PAIR responsibilities

- Register and pass the scheduler path through the canonical modular binary
  inventory.
- Consume proxy readiness, node presence, workload, and error effects.
- Keep proxies in automatic routing mode.
- Treat model **Load** as an engine model action, never a route selection.

## Prohibited PAIR behavior

- Do not call `node/select`.
- Do not implement a TypeScript scheduler or load balancer.
- Do not derive routing decisions from renderer telemetry.
- Do not emit a separate proxy-node-switch log.
- Do not treat GPU pressure as GPU capability, VRAM capacity, or device affinity.
- Do not mirror scheduler smoothing or proxy load/reservation accounting in
  TypeScript.

The GPU chart's `inferenceHardwareIds` filter is display-only.

## Inference logging

Never log:

- prompts or chat messages;
- request bodies;
- stream chunks;
- full inference responses;
- credentials or pairing data.

Log only operational metadata such as engine, model, job ID, node ID, path,
stream flag, and normalized error text.

Use `getErrorString` or the relevant normalized backend error helper instead of
serializing raw errors. Keep high-frequency request and routing details below
the default info level.