Saved in:
Bibliographic Details
Main Authors: Torroba-Hennigen, Lucas, Lang, Hunter, Guo, Han, Kim, Yoon
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2502.13811
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866915435871469568
author Torroba-Hennigen, Lucas
Lang, Hunter
Guo, Han
Kim, Yoon
author_facet Torroba-Hennigen, Lucas
Lang, Hunter
Guo, Han
Kim, Yoon
contents We study memory-efficient optimization of neural networks (in particular language models) with linear gradient transformations, where the gradients are linearly mapped to a lower dimensional space than the full parameter space, thus saving memory required for gradient accumulation and optimizer state persistence. The model parameters are updated by first performing an optimization step in the lower dimensional space and then going back into the original parameter space via the linear map's transpose. We show that optimizing the model in this transformed space is equivalent to reparameterizing the original model through a linear adapter that additively modifies the model parameters, and then only optimizing the adapter's parameters. When the transformation is Kronecker-factored, this establishes an equivalence between GaLore and one-sided LoRA. We show that this duality between gradient transformations and adapter-based reparameterizations unifies existing approaches to memory-efficient training and suggests new techniques for improving training efficiency and memory use.
format Preprint
id arxiv_https___arxiv_org_abs_2502_13811
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle On the Duality between Gradient Transformations and Adapters
Torroba-Hennigen, Lucas
Lang, Hunter
Guo, Han
Kim, Yoon
Machine Learning
Computation and Language
We study memory-efficient optimization of neural networks (in particular language models) with linear gradient transformations, where the gradients are linearly mapped to a lower dimensional space than the full parameter space, thus saving memory required for gradient accumulation and optimizer state persistence. The model parameters are updated by first performing an optimization step in the lower dimensional space and then going back into the original parameter space via the linear map's transpose. We show that optimizing the model in this transformed space is equivalent to reparameterizing the original model through a linear adapter that additively modifies the model parameters, and then only optimizing the adapter's parameters. When the transformation is Kronecker-factored, this establishes an equivalence between GaLore and one-sided LoRA. We show that this duality between gradient transformations and adapter-based reparameterizations unifies existing approaches to memory-efficient training and suggests new techniques for improving training efficiency and memory use.
title On the Duality between Gradient Transformations and Adapters
topic Machine Learning
Computation and Language
url https://arxiv.org/abs/2502.13811