A*STAR CDF | Algorithm-System Co-Design for Efficient and IP-Protected LLMs
PI. A*STAR Career Development Fund (CDF). S$250,000. 31 recipients selected from 194 applicants.
English
Title: Algorithm-System Co-Design for Efficient and IP-Protected LLMs: From Model Optimization to Cluster Deployment
Role: Principal Investigator (PI)
Funding: S$250,000
Funder: A*STAR Career Development Fund (CDF) Official announcement: A*STAR CDF 2024 Recipients
Selection: 31 recipients selected from 194 applicants
Objectives
This project addresses two core challenges in large-scale LLM deployment:
1. Inference Efficiency
Algorithm-system co-design to reduce inference latency and memory footprint, focusing on:
- Sparse Mixture-of-Experts (MoE) routing and expert caching
- Token scheduling and batch management at cluster scale
2. Intellectual Property Protection
Designing fingerprinting and attribution mechanisms so that model ownership can be verified even after fine-tuning or distillation:
- Routing-based model fingerprinting
- IP attribution in model merging and deployment pipelines
Affiliation
Agency for Science, Technology and Research (A*STAR), Singapore Centre for Frontier AI Research (CFAR)
中文版本
项目名称： 面向高效与知识产权保护的大语言模型算法-系统协同设计：从模型优化到集群部署
职责： 项目负责人（PI）
经费： 25 万新加坡元（约 135 万人民币）
资助方： 新加坡科学技术研究局职业发展基金（A*STAR Career Development Fund, CDF） 官方公告：A*STAR CDF 2024 获奖名单
遴选： 194 位申请者中择优录取 31 位
研究目标
本项目聚焦大规模 LLM 集群部署中两个核心难题：
1. 推理效率
通过算法与系统协同设计降低推理延迟和显存开销，重点方向：
- 稀疏混合专家（MoE）路由优化与专家缓存
- 集群级别的 Token 调度与批处理管理
2. 知识产权保护
设计模型指纹与归属认证机制，使模型所有权在微调或蒸馏后仍可被验证：
- 基于路由路径的模型指纹嵌入
- 模型合并与部署流水线中的 IP 归属认证
所属机构
新加坡科学技术研究局（A*STAR） 未来人工智能研究中心（CFAR）