Back to system overview

server

High-scale model serving infrastructure. The distributed RPC server executes inference efficiently across local hardware and networked devices.

Serving infrastructure, not routing logic

hirn_server focuses on fast and reliable model execution at scale. It does not predict intent; it serves the requests selected by router policy.

The server exposes a unified RPC interface and can target multiple backends:

  • llama.cpp: For highly optimized CPU/GPU offloading on consumer hardware.
  • vLLM: For high-throughput serving on larger VRAM setups.
  • exo / cactus: For experimental and specialized local routing architectures.

Distributed across devices

Running large models often requires more VRAM than a single machine can provide. The server pools capacity across multiple GPUs and multiple devices, turning your local network into one serving fabric.

Operationally simple

The goal is high-scale serving without heavy ops burden: consistent RPC contracts, backend-agnostic execution, and predictable resource behavior under load.

In short: router decides what should run, server decides how to serve it efficiently across available hardware.

High Throughput

Optimized for sustained multi-request serving with efficient caching, batching and backend-specific execution paths.

Cross-Platform

Natively optimized for Windows, Linux, and macOS while still scaling across heterogeneous hardware.

Auto-Scaling

Dynamically assigns and reclaims VRAM/resources so large models remain practical in real local deployments.