Llumnix Documentation

Llumnix is a full-stack solution for distributed LLM inference serving, featuring fully dynamic request scheduling for modern LLM deployments.

---

🚀 Getting Started

New to Llumnix? Start here.

• Quick Start — Deploy your first Llumnix cluster in minutes

• Deployment Guide - Full deployment guide covering all modes: neutral, PD, and PD-KVS

• Benchmark - Performance benchmarks for Llumnix on Kubernetes

• Llumlet Configuration — Configure Llumlet

---

🛠️ Development Guide

• Development Setup — Environment setup, build commands, and e2e unit tests guide

• Build Images — Manually build and push component images

---

🏗️ Design

Understand how Llumnix works internally.

• Architecture Overview — Full-stack component design

• Gateway

  • Gateway Architecture - Gateway architecture and basic functionalities

  • PDD Forwarding Protocol — Prefill-Decode disaggregation forwarding protocol

  • Batch Inference — Batch inference support

  • Traffic Splitting — Split traffic across internal and external endpoints with fallback

• Scheduler

  • Scheduling Policy Framework — Scheduling policy design

  • Instant and Accurate Load — Instance load obervation and modeling

  • Cache-aware Scheduling — Scheduling with KV cache state awareness

  • Predictor-Enhanced Scheduling — Online latency prediction for accurate prefill load estimation

  • SLO-aware Scheduling — Scheduling with SLO awareness

  • Adaptive PD Scheduling — Adaptive P/D role assignment to maximize SLO attainment

  • Rescheduler — Continuous rescheduling via request migration

• Llumlet

  • Llumlet and Llumlet Proxy — Engine-side agent bridging local engine and global scheduler

  • Real-time Instance Status Tracking — How Llumnix tracks engine state with minimal delay and overhead

  • Migration — Request migration implementation

---