FM customization deployment and lifecycle management

Deploy and manage customized models.

Customization can include:

• fine-tuning
• parameter-efficient adaptation
• adapters
• LoRA

Customization may be useful when the base model cannot reliably produce the required behavior.

A lifecycle can look like:

Training Data
    ↓
Customization / Fine-tuning
    ↓
Model Evaluation
    ↓
Model Registry
    ↓
Deployment
    ↓
Monitoring
    ↓
Rollback / Replacement / Retirement

• Amazon SageMaker AI
• SageMaker Model Registry
• Amazon S3
• CI/CD tooling
• Amazon CloudWatch

Low-Rank Adaptation (LoRA) updates a smaller set of trainable parameters rather than modifying the entire base model.

Benefits can include:

• lower training cost
• smaller adaptation footprint
• faster experimentation

Development
 → Evaluation
 → Versioning
 → Deployment
 → Monitoring
 → Rollback
 → Retirement

Customization is not automatically the answer.

Consider:

1. Prompt engineering
2. RAG
3. Fine-tuning/customization

Choose the simplest approach that meets the requirement.

---