PV-Tuning: Beyond Straight-Through Estimation for Extreme LLM Compression

TL;DR

This paper introduces PV-Tuning, a novel fine-tuning framework for extreme quantization of large language models, surpassing prior methods in accuracy and efficiency, especially at 1-2 bits per parameter.

Contribution

PV-Tuning is a representation-agnostic, improved fine-tuning approach that outperforms existing methods and guarantees convergence, enabling Pareto-optimal quantization of Llama 2 models at 2 bits.

Findings

PV-Tuning outperforms prior quantization techniques on Llama and Mistral models.

Achieves Pareto-optimal 2-bit quantization for Llama 2 models.

Demonstrates the limitations of straight-through estimators in extreme LLM compression.

Abstract

There has been significant interest in "extreme" compression of large language models (LLMs), i.e., to 1-2 bits per parameter, which allows such models to be executed efficiently on resource-constrained devices. Existing work focused on improved one-shot quantization techniques and weight representations; yet, purely post-training approaches are reaching diminishing returns in terms of the accuracy-vs-bit-width trade-off. State-of-the-art quantization methods such as QuIP# and AQLM include fine-tuning (part of) the compressed parameters over a limited amount of calibration data; however, such fine-tuning techniques over compressed weights often make exclusive use of straight-through estimators (STE), whose performance is not well-understood in this setting. In this work, we question the use of STE for extreme LLM compression, showing that it can be sub-optimal, and perform a systematic study of quantization-aware fine-tuning strategies for LLMs. We propose PV-Tuning - a representation-agnostic framework that generalizes and improves upon existing fine-tuning strategies, and provides convergence guarantees in restricted cases. On the practical side, when used for 1-2 bit vector quantization, PV-Tuning outperforms prior techniques for highly-performant models such as Llama and Mistral. Using PV-Tuning, we achieve the first Pareto-optimal quantization for Llama 2 family models at 2 bits per parameter.