NormFormer: Improved Transformer Pretraining with Extra Normalization

TL;DR

NormFormer introduces additional normalization steps in transformer layers to address gradient mismatch issues, leading to faster and more effective pretraining and improved downstream task performance with minimal computational overhead.

Contribution

The paper proposes NormFormer, a novel architecture that adds three normalization operations to each transformer layer, significantly enhancing training efficiency and model performance.

Findings

Faster convergence in pretraining (up to 24%)

Improved downstream task performance (up to 1.9% on GLUE)

Negligible increase in computational cost (+0.4% parameters)

Abstract

During pretraining, the Pre-LayerNorm transformer suffers from a gradient magnitude mismatch: gradients at early layers are much larger than at later layers. These issues can be alleviated by our proposed NormFormer architecture, which adds three normalization operations to each layer: a Layer Norm after self attention, head-wise scaling of self-attention outputs, and a Layer Norm after the first fully connected layer. The extra operations incur negligible compute cost (+0.4% parameter increase), but improve pretraining perplexity and downstream task performance for both causal and masked language models ranging from 125 Million to 2.7 Billion parameters. For example, adding NormFormer on top of our strongest 1.3B parameter baseline can reach equal perplexity 24% faster, or converge 0.27 perplexity better in the same compute budget. This model reaches GPT3-Large (1.3B) zero shot performance 60% faster. For masked language modeling, NormFormer improves fine-tuned GLUE performance by 1.9% on average. Code to train NormFormer models is available in fairseq https://github.com/pytorch/fairseq/tree/main/examples/normformer .