Invisible InfrastructureCatalog record
03SystemFORTHCOMING2024-12-08

Query Planning for Large Language Model Inference

Jake Lawrence · AI Systems / ML Infrastructure · AI Systems theme

Read online The interactive edition of this pieceFull paper The complete paper, web editionDownload PDF The typeset paper, for citing and printing

Abstract

An LLM that produces the right answer at 100x the necessary cost has an infrastructure problem, not an intelligence problem.

Applies cost-based query planning from database systems to LLM inference. Uses contextual bandits to route between execution plans for constrained decoding.

Record details

Type
System
Method
Systems design with bandit optimization
Field
AI Systems / ML Infrastructure
Subjects
AI Systems, ML Infrastructure
Status
Working Paper (forthcoming)
Released
2024-12-08

Cite and listen

Thesis

Related in the catalog

  • Classification as InfrastructurePosition Paper

    LLM-QP classifies queries by execution cost. Every routing decision is a classification decision with consequences the user never sees.

  • SAGENSystem

    Two missing infrastructure layers. SAGEN builds awareness; LLM-QP builds cost-efficiency. Both are classification systems.

  • The Beautiful UnfinishedEssay

    LLM-QP tries to close the gap between valid and affordable output. The Beautiful Unfinished asks whether that gap can ever fully close.

Back to the release table · See it on the network · Machine-readable catalog

Need something like this built?

I design and ship AI tools, full-stack apps, and data pipelines — end to end, to production. Tell me the problem in a sentence; I'll give you an honest read on fit within a day.

Work with me →