SpanNorm: Reconciling Training Stability and Performance in Deep Transformers
This addresses a critical problem for researchers and practitioners training deep Transformer models, offering a novel solution to improve both stability and performance, though it is incremental in refining normalization techniques.
The paper tackles the trade-off between training stability and performance in deep Transformers by proposing SpanNorm, which integrates PreNorm's stability and PostNorm's performance. Empirically, SpanNorm consistently outperforms standard normalization schemes in dense and Mixture-of-Experts scenarios.
The success of Large Language Models (LLMs) hinges on the stable training of deep Transformer architectures. A critical design choice is the placement of normalization layers, leading to a fundamental trade-off: the ``PreNorm'' architecture ensures training stability at the cost of potential performance degradation in deep models, while the ``PostNorm'' architecture offers strong performance but suffers from severe training instability. In this work, we propose SpanNorm, a novel technique designed to resolve this dilemma by integrating the strengths of both paradigms. Structurally, SpanNorm establishes a clean residual connection that spans the entire transformer block to stabilize signal propagation, while employing a PostNorm-style computation that normalizes the aggregated output to enhance model performance. We provide a theoretical analysis demonstrating that SpanNorm, combined with a principled scaling strategy, maintains bounded signal variance throughout the network, preventing the gradient issues that plague PostNorm models, and also alleviating the representation collapse of PreNorm. Empirically, SpanNorm consistently outperforms standard normalization schemes in both dense and Mixture-of-Experts (MoE) scenarios, paving the way for more powerful and stable Transformer architectures.