[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-120355-en":3,"doc-seo-120355-105":30,"detail-sidebar-cat-0-en-105":91},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":4,"is_deleted":4,"is_public":20,"is_downloadable":20,"audit_status":20,"page_count":21,"language":22,"language_code":23,"site_id":24,"html_lang":23,"table_of_contents":25,"faqs":26,"seo_title":27,"seo_description":14,"update_tm":28,"read_time":29},120355,1099514067438,"River Wang","https://ap-avatar.wpscdn.com/avatar/100002539ee87300030?x-image-process=image/resize,m_fixed,w_180,h_180&k=1780474512215547542",8,"Research & Report","Efficient Data - Algorithm Optimization for Modern Machine Learning System - dissertation","Modern machine learning systems face two escalating pressures: rapidly increasing data volumes and growing model architecture complexity, which strain GPU-based cloud deployment environments. For model efficiency, large vision-language architectures require substantial KeyValue (KV) cache memory during auto-regressive decoding, and text-oriented compression methods underperform due to modality-specific sparsity. For data reliability, Data Quality (DQ) problems accumulate across thousands of recurring pipelines, demanding costly manual monitoring. This dissertation introduces VL-Cache for KV compression and an optimization-based approach for automated DQ anomaly detection with provable guarantees.","UCLA  \nUCLA Electronic Theses and Dissertations  \nTitle  \nEfficient Data, Algorithm Optimization for Modern Machine Learning System  \nPermalink  \n[https://escholarship.org/uc/item/2p8246f5](https://escholarship.org/uc/item/2p8246f5)  \nAuthor  \nTU, Dezhan  \nPublication Date  \n2025  \nPeer reviewed|Thesis/dissertation  \n[eScholarship.org](eScholarship.org) Powered by the California Digital Library  \nUniversity of California  \nUNIVERSITY OF CALIFORNIA  \nLos Angeles  \nEfficient Data, Algorithm Optimization for Modern Machine Learning System  \nA dissertation submitted in partial satisfaction of the requirements for the degree  \nDoctor of Philosophy in Electrical and Computer Engineering  \nby  \nDezhan Tu  \n© Copyright by Dezhan Tu  \n2025  \nABSTRACT OF THE DISSERTATION  \nEfficient Data, Algorithm Optimization for Modern Machine Learning System  \nby  \nDezhan Tu  \nDoctor of Philosophy in Electrical and Computer Engineering University of California, Los Angeles, 2025  \nProfessor Lei He, Chair  \nModern machine learning systems (MLSys) are evolving rapidly, facing two critical challenges: the exponential growth of data volumes and the increasing complexity of model architectures. These challenges place significant pressure on ML deployment environment, particularly in GPU-based cloud infrastructures. On the model side, emerging large-scale architecture such as Vision-Language Models (VLMs) require substantial memory for KeyValue (KV) caches during auto-regressive decoding. Existing compression methods, designed for text-based Large Language Models (LLMs), fail to account for the modality-specific sparsity patterns in VLMs, leading to inefficiencies in both memory usage and inference speed. On the data side, as data volumes scale, Data Quality (DQ) issues often creep into machine learning pipelines. Engineers have to spend considerable effort manually monitoring and resolving issues in modern ML systems that serve thousands of recurring pipelines.  \nAutomating DQ monitoring and maintenance is essential to ensure reliability and reduce operational costs. This dissertation addresses these challenges with two key contributions:  \n1) Model Compression: we propose VL-Cache, a novel KV cache compression framework tailored for VLMs, incorporating layer-adaptive budget allocation and modality-aware token scoring to optimize memory usage without compromising accuracy; 2) Data Optimization: we formalize DQ detection as an optimization problem over historical pipeline statistics and develop approximation algorithms with provable guarantees for efficient anomaly detection.  \nExtensive experiments show that our methods exhibit effectiveness and efficiency in both domains, achieving significant memory savings and inference speedups, while also eliminating the need for human intervention in the loop of data operations.  \nThe dissertation of Dezhan Tu is approved.  \nMau-Chung Frank Chang Chang-Jin Kim  \nLin Yang Lei He, Committee Chair  \nUniversity of California, Los Angeles 2025  \nTo my wife Xiangxiang, for her love and accompanying. To my parents Changrong and Qiurong, for their support and trust.  \nv  \nContents  \n1 Introduction 1  \n1.1 Motivation ..................................... 3  \n1.2 Research Objective ................................ 3  \n1.3 Dissertation Outline ............................... 3  \n2 Background 5  \n2.1 Transformer-based Models ............................ 5  \n2.1.1 LLM Inference .............................. 5  \n2.1.2 VLM Inference .............................. 6  \n2.1.3 KV Cache Compression ......................... 7  \n2.2 GPU Architecture ................................. 7  \n3 Efficient Model Inference 9  \n3.1 Introduction .................................... 9  \n3.2 Preliminary: Sparse Attention Pattern ..................... 11  \n3.2.1 Measuring Attention Sparsity ...................... 12  \n3.2.2 Measuring Cache Hit Rate ........................ 13  \n3.3 VL-Cache Method ................................ 15  \n3.3.1 Sparsit","cbCaib1csA7kN38T","https://ap.wps.com/l/cbCaib1csA7kN38T","pdf",21843824,1,107,"English","en",105,"# 1 Introduction\n## 1.1 Motivation\n## 1.2 Research Objective\n## 1.3 Dissertation Outline\n# 2 Background\n## 2.1 Transformer-based Models\n## 2.2 GPU Architecture\n# 3 Efficient Model Inference\n## 3.1 Introduction\n## 3.2 Preliminary: Sparse Attention Pattern\n## 3.3 VL-Cache Method\n## 3.4 Experiments\n# 4 Low-Rank Adaption Fine-Tuning\n## 4.1 Introduction\n## 4.2 KVLoRA Method\n## 4.3 Experiments\n# 5 Automatic Data Validation\n## 5.1 Introduction\n## 5.2 Preliminary: DQ in MLSys\n## 5.3 Auto-Validate-by-History Method\n## 5.4 Experiments\n# 6 Conclusion and Future Work\n## 6.1 Research Summary\n## 6.2 Conclusion and Future Work","[{\"question\":\"What core challenges does the dissertation target in modern machine learning systems?\",\"answer\":\"It targets exponential data growth and increasing model complexity that pressure GPU-based deployment, alongside two practical issues: large KV-cache memory demands in vision-language decoding and data quality problems that require manual intervention.\"},{\"question\":\"How does VL-Cache improve inference efficiency for vision-language models?\",\"answer\":\"VL-Cache compresses KV caches using layer-adaptive budget allocation and modality-aware token scoring, reducing memory usage while maintaining accuracy and improving inference speed.\"},{\"question\":\"How is automatic data validation formulated in the dissertation?\",\"answer\":\"Data quality detection is formalized as an optimization problem over historical pipeline statistics, followed by approximation algorithms with provable guarantees for efficient anomaly detection.\"}]","Efficient Data - Algorithm Optimization for Modern Machine Learning System - dissertation | PDF",1785729641,270,{"code":4,"msg":31,"data":32},"ok",{"site_id":24,"language":23,"slug":33,"title":13,"keywords":34,"description":14,"schema_data":35,"social_meta":86,"head_meta":88,"extra_data":90,"updated_unix":28},"efficient-data-algorithm-optimization-for-modern-machine-learning-system-dissertation","",{"@graph":36,"@context":85},[37,54,68],{"@type":38,"itemListElement":39},"BreadcrumbList",[40,44,48,51],{"item":41,"name":42,"@type":43,"position":20},"https://docshare.wps.com","Home","ListItem",{"item":45,"name":46,"@type":43,"position":47},"https://docshare.wps.com/document/","Document",2,{"item":49,"name":12,"@type":43,"position":50},"https://docshare.wps.com/document/research-report/",3,{"item":52,"name":13,"@type":43,"position":53},"https://docshare.wps.com/document/efficient-data-algorithm-optimization-for-modern-machine-learning-system-dissertation/120355/",4,{"url":52,"name":13,"@type":55,"author":56,"headline":13,"publisher":58,"fileFormat":61,"inLanguage":23,"description":14,"dateModified":62,"datePublished":62,"encodingFormat":61,"isAccessibleForFree":63,"interactionStatistic":64},"DigitalDocument",{"name":9,"@type":57},"Person",{"url":41,"name":59,"@type":60},"DocShare","Organization","application/pdf","2026-08-03",true,{"@type":65,"interactionType":66,"userInteractionCount":4},"InteractionCounter",{"@type":67},"ViewAction",{"@type":69,"mainEntity":70},"FAQPage",[71,77,81],{"name":72,"@type":73,"acceptedAnswer":74},"What core challenges does the dissertation target in modern machine learning systems?","Question",{"text":75,"@type":76},"It targets exponential data growth and increasing model complexity that pressure GPU-based deployment, alongside two practical issues: large KV-cache memory demands in vision-language decoding and data quality problems that require manual intervention.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"How does VL-Cache improve inference efficiency for vision-language models?",{"text":80,"@type":76},"VL-Cache compresses KV caches using layer-adaptive budget allocation and modality-aware token scoring, reducing memory usage while maintaining accuracy and improving inference speed.",{"name":82,"@type":73,"acceptedAnswer":83},"How is automatic data validation formulated in the dissertation?",{"text":84,"@type":76},"Data quality detection is formalized as an optimization problem over historical pipeline statistics, followed by approximation algorithms with provable guarantees for efficient anomaly detection.","https://schema.org",{"og:url":52,"og:type":87,"og:title":13,"og:site_name":59,"og:description":14},"article",{"robots":89,"canonical":52},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":92},[93,97,101,105,110,115,120,123,128,131,135],{"id":20,"doc_module":4,"doc_module_name":46,"category_name":94,"show_sort_weight":95,"slug":96},"Story & Novel",90,"story-novel",{"id":47,"doc_module":4,"doc_module_name":46,"category_name":98,"show_sort_weight":99,"slug":100},"Literature",80,"literature",{"id":53,"doc_module":4,"doc_module_name":46,"category_name":102,"show_sort_weight":103,"slug":104},"Exam",70,"exam",{"id":106,"doc_module":4,"doc_module_name":46,"category_name":107,"show_sort_weight":108,"slug":109},5,"Comic",60,"comic",{"id":111,"doc_module":4,"doc_module_name":46,"category_name":112,"show_sort_weight":113,"slug":114},6,"Technology",50,"technology",{"id":116,"doc_module":4,"doc_module_name":46,"category_name":117,"show_sort_weight":118,"slug":119},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":46,"category_name":12,"show_sort_weight":121,"slug":122},30,"research-report",{"id":124,"doc_module":4,"doc_module_name":46,"category_name":125,"show_sort_weight":126,"slug":127},9,"Religion & Spirituality",20,"religion-spirituality",{"id":126,"doc_module":4,"doc_module_name":46,"category_name":129,"show_sort_weight":126,"slug":130},"World Cup","world-cup",{"id":132,"doc_module":4,"doc_module_name":46,"category_name":133,"show_sort_weight":132,"slug":134},10,"Lifestyle","lifestyle",{"id":136,"doc_module":4,"doc_module_name":46,"category_name":137,"show_sort_weight":106,"slug":138},19,"General","general"]