This April 2024 paper introduces Atom, a novel low-bit quantization method designed to enhance the efficiency and accuracy of Large Language Model (LLM) serving. The core challenge addressed is the high computational and memory costs associated with LLMs, especially when accommodating numerous user requests. Atom tackles this by quantizing both weights and activations to low-bit representations, like 4-bit, which significantly reduces memory consumption and boosts throughput by leveraging modern GPU capabilities. It maintains accuracy through mixed-precision quantization, fine-grained group quantization, and dynamic quantization, demonstrating substantial improvements in tokens per second with negligible accuracy loss compared to existing methods. The paper provides a detailed analysis of Atom's design, implementation, and comprehensive evaluation across various LLM models and tasks.Source: https://arxiv.org/pdf/2310.19102

Ver todos los episodios