使用 Amazon SageMaker AI 和 SGLang 通过自定义模型扩展 AI 代理.pdf

1、 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.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 8 7Amit ModiDan FergusonScale AI Agents with Custom Models using Amazon SageMaker AI and S

2、GLangHe/himSenior Manager,Inference and ModelOpsAWSHe/himSr WW Specialist SA,GenAIAWSYing ShengShe/herCo-creator of SGLangLMSYS Corp 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.Enterprises are doubling down on agentsGartner,“Top strategic Technology Trends for 2025,33%2025,Ama

3、zon Web Services,Inc.or its affiliates.All rights reserved.Bloomberg,“Chesky Says OpenAI Tools Not Ready for ChatGPT Tie-Up With Airbnb App”“Were relying a lot on Alibabas Qwen.Its very good.Its also fast and cheap.“Brian Chesky,CEO Airbnb 2025,Amazon Web Services,Inc.or its affiliates.All rights re

4、served.Gartner,“AI Maturity Matters:Proportion of AI and GenAI Prototypes Making It Into Production”41%of generative AI prototypes reached production 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.Challenges with Model Customization at ScaleCustomization workflows are manual and

5、time-consumingOptimizing inference for cost-efficiency is complexModel and agent observability is fragmentedExperimentation and production workflows are disconnectedDifficult to track,version,and audit models 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.Train and tune models wi

6、th SageMaker AIBroadest choice of modelsBroadest choice of fine-tuning recipesFully managed training jobs and HyperPod for training 2025,Amazon Web Services,Inc.or its affiliates.All rights reserved.Resilient Infrastructure for TrainingNodefailureInstancerestoreTrainingCheckpointsAutomatic node repl