BranchNorm: Robustly Scaling Extremely Deep Transformers

Recently, DeepNorm scales Transformers into extremely deep (i.e.,in0 layers) and reveals the promising potential of deep scaling. To stabilize the training of deep models, Deep-Norm (Wang et al., 2022a) attempts to constrain the model update to a constant value. Although applying such a constraint can benefit the early stage of model training, it may lead to undertrained models during the whole training procedure. In this paper, we propose Branch-Norm, which dynamically rescales the nonresidual branch of Transformer in accordance with the training period. BranchNorm not only theoretically stabilizes the training with smooth gradient norms at the early stage, but also encourages better convergence in the subsequent training stage. Experimental results on multiple translation tasks demonstrate that BranchNorm achieves a better trade-off between training stability and converge performance.