Adam-mini: Use Fewer Learning Rates To Gain More

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Zhang, Yushun, Chen, Congliang, Li, Ziniu, Ding, Tian, Wu, Chenwei, Kingma, Diederik P., Ye, Yinyu, Luo, Zhi-Quan, Sun, Ruoyu
Format: Preprint
Published: 2024
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866912243094913024
author Zhang, Yushun
Chen, Congliang
Li, Ziniu
Ding, Tian
Wu, Chenwei
Kingma, Diederik P.
Ye, Yinyu
Luo, Zhi-Quan
Sun, Ruoyu
author_facet Zhang, Yushun
Chen, Congliang
Li, Ziniu
Ding, Tian
Wu, Chenwei
Kingma, Diederik P.
Ye, Yinyu
Luo, Zhi-Quan
Sun, Ruoyu
contents We propose Adam-mini, an optimizer that achieves on par or better performance than AdamW with 50% less memory footprint. Adam-mini reduces memory by cutting down the learning rate resources in Adam (i.e., $1/\sqrt{v}$). By investigating the Hessian structure of neural nets, we find Adam's $v$ might not function at its full potential as effectively as we expected. We find that $\geq$ 99.9% of these learning rates in $v$ could be harmlessly removed if we (1) carefully partition the parameters into blocks following our new principle on Hessian structure; (2) assign a single but good learning rate to each parameter block. We then provide one simple way to find good learning rates and propose Adam-mini. Empirically, we verify that Adam-mini performs on par or better than AdamW on various language models sized from 39M to 13B for pre-training, supervised fine-tuning, and RLHF. The reduced memory footprint of Adam-mini also alleviates communication overheads among GPUs, thereby increasing throughput. For instance, Adam-mini achieves 49.6% higher throughput than AdamW when pre-training Llama 2-7B on $2\times$ A800-80GB GPUs, which saves 33% wall-clock time for pre-training.
format Preprint
id arxiv_https___arxiv_org_abs_2406_16793
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Adam-mini: Use Fewer Learning Rates To Gain More
Zhang, Yushun
Chen, Congliang
Li, Ziniu
Ding, Tian
Wu, Chenwei
Kingma, Diederik P.
Ye, Yinyu
Luo, Zhi-Quan
Sun, Ruoyu
Machine Learning
Artificial Intelligence
We propose Adam-mini, an optimizer that achieves on par or better performance than AdamW with 50% less memory footprint. Adam-mini reduces memory by cutting down the learning rate resources in Adam (i.e., $1/\sqrt{v}$). By investigating the Hessian structure of neural nets, we find Adam's $v$ might not function at its full potential as effectively as we expected. We find that $\geq$ 99.9% of these learning rates in $v$ could be harmlessly removed if we (1) carefully partition the parameters into blocks following our new principle on Hessian structure; (2) assign a single but good learning rate to each parameter block. We then provide one simple way to find good learning rates and propose Adam-mini. Empirically, we verify that Adam-mini performs on par or better than AdamW on various language models sized from 39M to 13B for pre-training, supervised fine-tuning, and RLHF. The reduced memory footprint of Adam-mini also alleviates communication overheads among GPUs, thereby increasing throughput. For instance, Adam-mini achieves 49.6% higher throughput than AdamW when pre-training Llama 2-7B on $2\times$ A800-80GB GPUs, which saves 33% wall-clock time for pre-training.
title Adam-mini: Use Fewer Learning Rates To Gain More
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2406.16793