BoA introduces a novel approach to post-training quantization that significantly improves efficiency of large language models without costly retraining.

• Proposes attention-aware optimization to preserve critical model components
• Achieves state-of-the-art performance for 4-bit LLM quantization
• Eliminates the need for backpropagation, making it practical for billion-parameter models
• Enables deployment on resource-constrained devices with minimal accuracy loss

This engineering breakthrough makes powerful AI models more accessible while maintaining their capabilities, addressing critical deployment challenges in real-world applications.

BoA: Attention-aware Post-training Quantization without Backpropagation