如何优化 AWS 存储以适应生成式 AI 和 ML 工作负载.pdf

1、 2024,Amazon Web Services,Inc.or its affiliates.All rights reserved.2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.S T G 3 4 8How to optimize AWS storage for generative AI and ML workloadsJohn Mallory(he/him)WWSO Storage SpecialistAmazon Web ServicesLaura Shepard(she/her)WWSO Sto

2、rage SpecialistAmazon Web Services 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.What 100,000 ML customers on AWS are teaching us?01Everyone needs an AI data strategy02Your unique data is the key to AI business value03A modern data foundation accelerates PoC to Production04Optim

3、al storage decisions reduce costs and deployment times 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.Foundations of a GenAI/ML data strategy Work backward from desired outcomesDetermine your data needsAssess what you haveAcquire what you needIngest from all sourcesPut your datat

4、o workCatalogSecure&GovernPrepare dataIntegrate ToolsPrepareto evolveRapidly experiment&iterateImprove model accuracy withyour dataAdopt new tools and technologiesCost effectively retain more dataStart with the right data foundation to avoid one-way doors 2025,Amazon Web Services,Inc.or its affiliat

5、es.All rights reserved.Using your data for GenAIRAGSpecialized knowledge through prompt augmentation Enterpriseknowledge corpusNo change to the foundation models AgenticLong term memory for knowledge and customization Short term memory for context and stateUse multiple models Fine-tuningSpecialized

6、knowledge for specific tasksSmall number of labeled examplesChange a copy of the foundation model Training/Continued pre-trainingGeneralized and specialized knowledge Trained on yourdata onlyCreate a new model from scratch,or change a model copyEASIERMORE COMPLEXA D O P T I O N 2025,Amazon Web Servi