Scheduler

Llumnix’s scheduler handles dynamic request scheduling for distributed LLM inference.

Contents

• Policy Framework
  • Scheduling modes: full-mode and lite-mode
  • Metrics: quantifying instance load and latency
  • Filters: filtering candidate instances
  • Selectors: choosing the final instance
  • Built-in policies and modes
  • Scheduling features
  • Future work
• Instant and Accurate Load
  • Goal and definition
  • Existing paradigms
  • Limitations of existing paradigms
  • Llumnix Ial solutions
• Cache-aware Scheduling
  • Introduction
  • Design and implementation
• Predictor-Enhanced Scheduling
  • Introduction
  • Design and Implementation
  • Current Limitations and Future Direction
• SLO-aware Scheduling
  • Overview
  • Generating Profiling Data
  • Latency Prediction
  • Scheduling Pipeline
• Adaptive PD Scheduling
  • Introduction
  • Prerequisites
  • Overview
  • Reserve State Management
  • Scheduling Strategy
  • Rescheduling Strategy
  • Performance
  • Future Work
• Rescheduler
  • Introduction
  • Architecture
  • Rescheduling Policies
  • Migration Implementation