Decoupled Relative Learning Rate Schedules

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Ludziejewski, Jan, Małaśnicki, Jan, Pióro, Maciej, Krutul, Michał, Ciebiera, Kamil, Stefaniak, Maciej, Krajewski, Jakub, Sankowski, Piotr, Cygan, Marek, Adamczewski, Kamil, Jaszczur, Sebastian
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866918083413671936
author Ludziejewski, Jan
Małaśnicki, Jan
Pióro, Maciej
Krutul, Michał
Ciebiera, Kamil
Stefaniak, Maciej
Krajewski, Jakub
Sankowski, Piotr
Cygan, Marek
Adamczewski, Kamil
Jaszczur, Sebastian
author_facet Ludziejewski, Jan
Małaśnicki, Jan
Pióro, Maciej
Krutul, Michał
Ciebiera, Kamil
Stefaniak, Maciej
Krajewski, Jakub
Sankowski, Piotr
Cygan, Marek
Adamczewski, Kamil
Jaszczur, Sebastian
contents In this work, we introduce a novel approach for optimizing LLM training by adjusting learning rates across weights of different components in Transformer models. Traditional methods often apply a uniform learning rate across all network layers, potentially overlooking the unique dynamics of each part. Remarkably, our introduced relative learning rates, RLRS, method accelerates the training process by up to $23\%$, particularly in complex models such as Mixture of Experts (MoE). Hyperparameters of RLRS can be efficiently tuned on smaller models and then effectively reused on models up to $27\times$ larger. This simple and effective method results in a substantial reduction in training time and computational resources, offering a practical and scalable solution for optimizing large-scale neural networks.
format Preprint
id arxiv_https___arxiv_org_abs_2507_03526
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Decoupled Relative Learning Rate Schedules
Ludziejewski, Jan
Małaśnicki, Jan
Pióro, Maciej
Krutul, Michał
Ciebiera, Kamil
Stefaniak, Maciej
Krajewski, Jakub
Sankowski, Piotr
Cygan, Marek
Adamczewski, Kamil
Jaszczur, Sebastian
Machine Learning
In this work, we introduce a novel approach for optimizing LLM training by adjusting learning rates across weights of different components in Transformer models. Traditional methods often apply a uniform learning rate across all network layers, potentially overlooking the unique dynamics of each part. Remarkably, our introduced relative learning rates, RLRS, method accelerates the training process by up to $23\%$, particularly in complex models such as Mixture of Experts (MoE). Hyperparameters of RLRS can be efficiently tuned on smaller models and then effectively reused on models up to $27\times$ larger. This simple and effective method results in a substantial reduction in training time and computational resources, offering a practical and scalable solution for optimizing large-scale neural networks.
title Decoupled Relative Learning Rate Schedules
topic Machine Learning
url https://arxiv.org/abs/2507.03526