CONCEPT Cited by 1 source

FMOps pipeline

Definition

An automated, versioned model customization workflow for foundation models — analogous to MLOps but specialized for the foundation-model lifecycle (fine-tuning, evaluation, packaging, deployment). Enables repeatable model customization triggered by new data without manual intervention.

In the edge AI context, an FMOps pipeline produces quantized model artifacts deployable to hardware-constrained devices, creating a factory-like model production process:

Raw documentation → LLM-generated training data → Fine-tuning → Evaluation → Packaging → Deployment

Key properties

• Repeatable and versioned: Triggered whenever new documentation/data is added
• Automated training data generation: Uses a large foundation model (e.g., Amazon Nova Pro) to produce structured Q&A pairs from raw documentation — avoiding manual curation
• Quantization as a pipeline stage: Model must be converted to edge-deployable format (GGUF) before artifact storage
• Feedback loop: Edge usage data feeds back into the next pipeline run

Seen in

• sources/2026-07-22-aws-architecting-offline-first-generative-ai-applications-for-edge-deployments — Amazon Bedrock (Nova Pro) generates Q&A training data → SageMaker AI Pipelines fine-tunes gpt-oss-20b → S3 artifact → IoT Greengrass deploys to edge.