LLM.int8(): 8-bit Matrix Multiplication for Transformers at Scale proposes a novel quantization method to significantly reduce the memory footprint of Large Language Models (LLMs) during inference without sacrificing predictive performance.

Here is a short summary of its key points:

• The Problem: Large pre-trained language models require massive amounts of GPU memory for inference. While quantizing parameters to 8-bit can cut memory use in half, traditional quantization methods fail and degrade performance for models larger than 6.7 billion parameters.
• The Cause of Degradation: The authors discovered that as transformers scale beyond 6.7B parameters, they develop highly systematic "outlier" features with extreme magnitudes. These sparse outliers dominate the transformer's predictive performance, and forcing them into standard 8-bit bins ruins the model's precision and accuracy.
• The Solution (LLM.int8()): The paper introduces a two-part quantization procedure to handle these outliers:
• The Impact: LLM.int8() allows zero-degradation 8-bit quantization for models up to 175 billion parameters. This drastically reduces memory requirements, making it possible to run massive models like OPT-175B or BLOOM on a single server with consumer-grade GPUs, vastly improving the accessibility of large-scale AI research.