Stable-Transformer: Towards a Stable Transformer Training
Keywords: Transformer, Stable Transformer
TL;DR: We introduce Stable-Transformer that achieves a more stable training process.
Abstract: The scale of parameters in Transformers has expanded dramatically—from hundreds of millions to several trillion. A key challenge when scaling the model to trillions is the training instability. Although many practical tricks, such as learning rate warmup, query-key normalization and better weight initialization, have been introduced to mitigate the training instability, a rigorous mathematical understanding of why such instabilities happen and why the above-mentioned tricks work well is still unclear. In this paper, we give a theoretical analysis of the initialization, normalization and attention mechanism in Transformers, and present a set of stabilized designs of the initialization, normalization and attention mechanism, which are thus termed as StableInit, StableNorm and StableAtten, individually. In experiments, we demonstrate that each of our stabilized designs, i.e., StableInit, StableNorm and StableAtten, exhibits better stability. Furthermore, by putting the stabilized designs together, we propose a stabilized Transformer, termed Stable-Transformer, and show in experiments on large model (1B parameters) and deep model (200 layers) that our proposed Stable-Transformer
achieves a more stable training process.
Primary Area: optimization
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 2907
Loading