Algebraic Integer Quantization for Machine Learning Applications
The authors propose Algebraic Integer Number Formats (AINF) to create highly configurable and accurate compressed computation for large language models (LLMs). They introduce Algebraic Integer Quantization (AIQ) for LLM inference, achieving accuracy comparable to the industry-standard MX format at equivalent bit budgets. This method uses algebraic integers to enable error-free arithmetic and optimized matrix multiplication.