LoQT: Low-Rank Adapters for Quantized Pretraining

Sebastian Bugge Loeschcke, Mads Toftrup, Michael Kastoryano, Serge Belongie, Vésteinn Snæbjarnarson

Keywords: Quantization, Low-Rank Adaptation, Memory Efficient Training, Large Language Models

TL;DR: LoQT enables efficient quantized pretraining of LLMs with results close to full-rank non-quantized models. It enables pretraining of a 13B LLM on a 24GB GPU without model parallel, checkpointing, or offloading strategies during training.

Abstract:

Despite advances using low-rank adapters and quantization, pretraining of large models on consumer hardware has not been possible without model sharding, offloading during training, or per-layer gradient updates. To address these limitations, we propose Low-Rank Adapters for Quantized Training (LoQT), a method for efficiently training quantized models. LoQT uses gradient-based tensor factorization to initialize low-rank trainable weight matrices that are periodically merged into quantized full-rank weight matrices. Our approach is suitable for both pretraining and fine-tuning models. We demonstrate this for language modeling and downstream task adaptation, finding that LoQT enables efficient training of models up to 7B parameters on a 24GB GPU. We also demonstrate the feasibility of training a 13B model using per-layer gradient updates on the same hardware.

Primary Area: Optimization for deep networks

Submission Number: 12149