具有乘积量化和统一索引的 LLM 增强型 GNN 加速器.pdf

《具有乘积量化和统一索引的 LLM 增强型 GNN 加速器.pdf》由会员分享，可在线阅读，更多相关《具有乘积量化和统一索引的 LLM 增强型 GNN 加速器.pdf（26页珍藏版）》请在三个皮匠报告上搜索。

1、Accelerator for LLM-Enhanced GNN with Product Quantization and Unified IndexingJiaming Xu*1,2,Jinhao Li*1,Jun Liu1,Hao Zhou1,Guohao Dai1,2Equal contribution*,Shanghai Jiao Tong University1,Infinigence-AI2,Correspondence to:Guohao Dai Jiaming Xu DAI Group,Shanghai Jiao Tong University2025/3/7OutlineO

2、utlineBackground and ChallengeTechniques Overview PQ-based MatMul Unified Architecture Extensible GFM-ISA DesignExperiment ResultsPage 2Jiaming Xu DAI Group,Shanghai Jiao Tong University2025/3/7Graph Neural Network for Graph DataGraph Neural Network for Graph DataPage 3Graph neural networks(GNNs)are

3、 designed for graph data problem.Social NetworkChemical MoleculeGraph data is everywhere.RecommendationCitation NetworkGraph DataGraph Neural NetworksGraph DataAggregation/Message PassingNode EmbeddingModel WeightCombinationJiaming Xu DAI Group,Shanghai Jiao Tong University2025/3/71 Zheng X,Zhang M,

4、Chen C,et al.Gnnevaluator:Evaluating gnn performance on unseen graphs without labelsJ.Advances in Neural Information Processing Systems,2024,36.Problem of Traditional GNNProblem of Traditional GNNPage 4Vulnerability of traditional GNN on unseen graphs.GNN Training and InferencePoor Generalization on

5、 Unseen Graphs1GNNGraph DatasetsTrainingNew Node in Training GraphInferenceGNNGraph Task(Eg.,Node Classification)New Node in New GraphInferenceGNNHigh Accuracy!Graph Task(Eg.,Node Classification)Poor Generalization!Acc.20%loss82.7%66.4%Training on CoraTraining on Wiki-CSInference on CoraJiaming Xu D

6、AI Group,Shanghai Jiao Tong University2025/3/7Era of Large Language Models(LLMs)Era of Large Language Models(LLMs)Page 5LLMs have attracted much attention.Rapid Development of LLMs1 Qin L,Chen Q,Zhou Y,et al.Multilingual large language model:A survey of resources,taxonomy and frontiersJ.arXiv prepri