HALO: Hadamard-Assisted Lower-Precision Optimization for LLMs

TL;DR

HALO introduces a novel quantization-aware training method for LLMs that uses Hadamard rotations to enable accurate, low-precision fine-tuning with significant speedups, maintaining near-full-precision accuracy.

Contribution

The paper presents HALO, a new approach combining Hadamard rotations, high-performance kernels, and FSDP for efficient low-precision LLM training and fine-tuning.

Findings

Achieves near-full-precision accuracy during fine-tuning.

Provides up to 1.41x speedup on RTX 4090 GPUs.

Supports both standard and parameter-efficient fine-tuning.

Abstract

Quantized training of Large Language Models (LLMs) remains an open challenge, as maintaining accuracy while performing all matrix multiplications in low precision has proven difficult. This is particularly the case when fine-tuning pre-trained models, which can have large weight and activation outlier values that make lower-precision optimization difficult. To address this, we present HALO, a novel quantization-aware training approach for Transformers that enables accurate and efficient low-precision training by combining 1) strategic placement of Hadamard rotations in both forward and backward passes, which mitigate outliers, 2) high-performance kernel support, and 3) FSDP integration for low-precision communication. Our approach ensures that all large matrix multiplications during the forward and backward passes are executed in lower precision. Applied to LLAMA-family models, HALO achieves near-full-precision-equivalent results during fine-tuning on various tasks, while delivering up to 1.41x end-to-end speedup for full fine-tuning on RTX 4090 GPUs. HALO efficiently supports both standard and parameterefficient fine-tuning (PEFT). Our results demonstrate the first practical approach to fully quantized LLM fine-tuning that maintains accuracy in 8-bit precision, while delivering performance benefits. Code is available at https://github.com/IST-DASLab/HALO.