Day 3: SageMaker for GenAI
Learning Objectives
- - Understand SageMaker endpoint deployment patterns
- - Know when to use SageMaker JumpStart vs Bedrock
- - Understand LoRA, QLoRA, and full fine-tuning trade-offs
- - Use SageMaker Model Registry for version management
- - Apply SageMaker Clarify for bias detection and Model Monitor for drift
Tasks
Tasks
0/6 completed- Read30m
SageMaker JumpStart Foundation Models
How to deploy foundation models via JumpStart. Understand the deployment configuration options.
- Read20m
SageMaker Model Registry
Version management, approval workflows, deployment pipelines for models.
- Read20m
SageMaker Clarify for Bias Detection
Pre-training and post-training bias detection. Model explainability.
- Read20m
SageMaker Ground Truth and Ground Truth Plus
Generating evaluation datasets and human feedback loops for GenAI.
- Blog30m
Advanced Fine-Tuning Methods on SageMaker AI
Compares LoRA, QLoRA, and full fine-tuning. Key for exam questions about model customization.
- Blog20m
Tutorials Dojo AIP-C01 Exam Guide (SageMaker coverage)
Community exam guide covering SageMaker topics for the certification.
Exam Skills
Write your understanding, then reveal the reference answer.
Hands-On Lab
Build real muscle memory with these activities.
Deploy a JumpStart Foundation Model Endpoint
Use SageMaker JumpStart to deploy a foundation model endpoint and test inference.
- 1 Open SageMaker Studio and navigate to JumpStart
- 2 Search for a foundation model (e.g., Llama 3 8B or Mistral 7B)
- 3 Click Deploy and select an ml.g5.2xlarge instance
- 4 Wait for the endpoint to reach InService status (5-10 minutes)
- 5 Send a test inference request using the Studio notebook or AWS CLI and verify the response
Explore LoRA Fine-Tuning Configuration on Bedrock
Walk through the Bedrock fine-tuning console to understand LoRA configuration options.
- 1 Open the Bedrock console and navigate to Custom models → Fine-tuning
- 2 Select a supported model (e.g., Amazon Nova or Llama) and click Create fine-tuning job
- 3 Review the training data format requirements (JSONL with prompt/completion pairs)
- 4 Examine the hyperparameter options: epochs, batch size, learning rate, warmup steps
- 5 Note the S3 bucket requirements for training data and output artifacts (do not submit unless you have training data ready)
Review SageMaker Model Registry Workflow
Explore Model Registry for versioning and approval workflows using the SageMaker console.
- 1 Open SageMaker Studio and navigate to Model Registry
- 2 Create a new Model Package Group called 'genai-models'
- 3 Review the approval status workflow: PendingManualApproval → Approved → Deployed
- 4 Note how model versions are tracked with metadata, metrics, and lineage
Scenarios
Think through each scenario before revealing the answer.
Healthcare Model Fine-Tuning
- •What fine-tuning method minimizes compute cost while adapting to domain language?
- •Where do you store and version the fine-tuned model?
- •What compliance requirements does healthcare impose on deployment?
Practice Questions
11 questions across 3 difficulty levels.
Further Reading
Go deeper into today's topics.
Advanced Fine-Tuning Methods on SageMaker
Compares LoRA, QLoRA, and full fine-tuning. Key for exam questions about model customization.
Multi-LoRA Serving with vLLM on SageMaker
Dozens of LoRA adapters sharing one GPU with hot-swapping.
Reinforcement Fine-Tuning — 66% Accuracy Gain
66% average accuracy gain over base models with reinforcement fine-tuning on Bedrock.
LLM Fine-Tuning on AWS — SFT, Continued Pre-Training, RLHF
Comprehensive SageMaker fine-tuning: SFT, continued pre-training, RLHF — QLoRA, LoRA, full fine-tuning compared.
Bedrock Customization Workshop (Fine-Tuning Notebooks)
Hands-on model customization: continued pre-training, supervised fine-tuning, evaluation.