Skip to content
Octopus Research Institute
Status: Active researchResearch software

octoryn-router (sovereign DP router)

A hand-written sovereign HTTP reverse-proxy router over a data-parallel inference fleet, with mid-stream byte-identical failover and an event-logged audit trail.

Description

Routes requests across a data-parallel fleet without a third-party proxy, and can fail over mid-stream between chips while keeping decode byte-identical. Every routing decision is written to an event-logged audit trail.

Highlights

  • Hand-written HTTP reverse-proxy in front of a data-parallel inference fleet — no third-party proxy in the request path
  • Mid-stream failover between chips or replicas with byte-identical decoded output
  • Every routing decision written to an event-logged audit trail
  • Routing and failover logic owned and inspectable, not delegated to an opaque load balancer
  • Apache-2.0 licence

octoryn-router is a hand-written HTTP reverse-proxy that sits directly in front of a data-parallel inference fleet. There is no third-party proxy in the request path: the routing and failover logic is code you own and can read, rather than behaviour delegated to an opaque load balancer. That is what "sovereign" means here — not a marketing adjective, but the concrete property that every decision about where a request goes is made by inspectable logic under your control.

How it works

The router's defining capability is mid-stream failover. When a chip or replica serving an in-flight generation becomes unavailable, the router can move that request to another replica while keeping the decoded output byte-identical, so a client that is already mid-generation does not observe a broken or divergent stream. This depends on cross-chip decode consistency: the same tokens must decode the same way on the replica that takes over, which is the subject of a related dataset on cross-chip decode consistency. On top of routing, every decision is written to an event-logged audit trail, so the path a request actually took can be reconstructed after the fact.

Why it matters

For teams running self-hosted, multi-chip inference, failover usually forces a trade-off between resilience and control — reaching for a third-party balancer means accepting behaviour you cannot fully see or verify. octoryn-router keeps both the failover and the record of it in your hands. It sits alongside the institute's wider work on local-sovereign-inference and governed-agent-systems, applying the same principle at the serving layer: decisions stay owned, inspectable, and evidenced. Licensed Apache-2.0.

Limitations

  • Byte-identical failover depends on cross-chip decode consistency (see the related dataset).
  • Repository is migrating off the retiring Octoryn Research domain.

Related research