Navigating EU AI Act requirements for LLM fine-tuning on Amazon SageMaker AI

This article explains how to track computational resources (FLOPs) during LLM fine-tuning on Amazon SageMaker AI to comply with EU AI Act requirements.

• EU AI Act requires tracking FLOPs to determine if fine-tuning reclassifies you as a GPAI model provider
• Default compliance threshold is 3.3×10²² FLOPs; 30% of actual pretraining compute if known
• Fine-Tuning FLOPs Meter toolkit automates FLOPs calculation and compliance tracking
• Supports both architecture-based (analytical) and hardware-based (upper bound) FLOPs calculation methods
• Integrates with Hugging Face training via single configuration flag: compute_flops: true
• Generates audit-ready JSON documentation with compliance status and threshold determination
• Handles parameter-efficient methods like LoRA with enhanced formulas for accurate estimates
• Toolkit available in Amazon SageMaker Generative AI recipes repository on GitHub

The Fine-Tuning FLOPs Meter simplifies EU AI Act compliance by automating computational resource tracking and generating regulatory documentation for LLM fine-tuning workloads on SageMaker AI.