PROMPTS: Performance Optimization via Multi-Agent Planning for LLM Training and Serving

简体中文

L1 profiler evidence ordering and constrained proposals; no Google TPU profiler or GSPMD.

Paper information

Field	Details
Paper	Paper or original source
Institution	University of Maryland / Google (first author).
First public release	2026-05-18
Author code	No public author repository found as of 2026-09-14.
Local adapter or method key	prompts
Local reproduction code	src/auto_research/agent_research/latest_20260914.py

Paper method and evidence

Coordinator, Analyzer and Proposal Agent jointly consult profiler and knowledge-base evidence, diagnose bottlenecks and propose explainable sharding changes.

flowchart LR
  A[Profiler and knowledge base] --> B[Diagnose bottleneck]
  B[Diagnose bottleneck] --> C[Propose constrained sharding]

Original paper figure

Figure 2 illustrates the multi-agent optimization flow. From the original paper; copyright remains with the authors.

Reported results

Paper reports up to 434% improvement on eight production workloads. Engineer selections always in top 3; top 1 hit 87.5%. Not a local run or inferred online A/B.

Local reproduction

MLSys 2026 Industry Track; repository identifier mlsys2026-prompts. Three-seed metrics use matched public fixtures and budget; only state transitions are validated.