Summary

Multi-model inference is the practice of running multiple specialized AI models (embedding, reranking, OCR, vision, extraction, generation) in a single pipeline, ideally sharing GPU resources to minimize cost and latency.

The Problem

Standard serving tools (vLLM, TEI) each serve one model type and assume exclusive GPU ownership:

  • vLLM serves LLMs, occupies ~90% of GPU at startup
  • TEI serves embedding/reranking models
  • Neither covers OCR, vision, or document parsing
  • Result: 4 servers + 4 APIs + 4 deployments, each with dedicated GPU

Why GPU Sharing Matters

  • Small specialized models are accurate enough for their tasks
  • But switching to small models doesn’t make inference cheaper by itself
  • Savings only appear when models share GPUs
  • Requires a single engine that can run all model types

Solutions

  • superlinked-inference-engine — Open-source engine running all model types on shared GPUs
  • Smart batching by compute cost (not item count)
  • Dynamic model loading/eviction (LRU cache pattern)
  • Gateway-based shared queue across workers
  • harness — The program that runs LLMs
  • inference — General inference optimization
  • rag — Retrieval-augmented generation pipelines

Evidence Timeline

  • 2026-07-28: Avi Chawla’s article detailed multi-model inference challenges and SIE solution (sources/articles/avichawla-superlinked-sie-2026.md)