AWS Trainium 上的端到端基础模型生命周期.pdf

1、 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.A I M 3 5 1End-to-End Foundation Model Lifecycle on AWS TrainiumKamran KhanBusiness Development Annapurna Labs,AWS Matt McCleanCustomer EngineeringAnnapurna Labs,AWS

2、Randeep Bhatia CTOSplash Music 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.The Circle of Life:AI Model 3Use case discovery,prioritization,selectionData collectionModel SelectionAdapt your ModelOffline Model EvaluationOptimize for DeploymentEval for Production MetricsDeploy and

3、 scaleBusiness valueAI Model LifecycleDecision pointLearn,repeat&accelerate 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.The Circle of Life:AI Model 4Use case discovery,prioritization,selectionData collectionModel SelectionAdapt your ModelOffline Model EvaluationOptimize for De

4、ploymentEval for Production MetricsDeploy and scaleBusiness valueAI Model LifecycleDecision pointLearn,repeat&accelerate 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.5Iterations:Slow fine-tuning and deployment cycles limit innovation and increase time to market.Compute Scarcity

5、:Access to the right compute for your workloads impacts cost-efficiency of your AIRising Compute Costs:Selecting the right model can help optimize,can make building and serving models too expensiveSlow Inference:High inference latency will lead to underwhelming experiences and high cost AI Bottlenec

6、ksChallenges we need to over come to build and deploy models efficiently and cost effectively 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.6Iterations:Faster Iterations with open source models Democratizing AI:Easier access to the right compute AWS Inferentia or TrainiumAWS Tra