Mixture-of-Depths Attention

TL;DR

Mixture-of-depths attention (MoDA) enhances deep language models by allowing attention across different layer depths, improving performance with minimal computational overhead, and is efficiently implementable on hardware.

Contribution

Introduces MoDA, a novel attention mechanism enabling cross-depth attention in deep models, with an efficient algorithm and demonstrated performance gains.

Findings

MoDA improves perplexity by 0.2 on validation benchmarks.

Increases downstream task performance by 2.11%.

Achieves 97.3% of FlashAttention-2 efficiency at sequence length 64K.

Abstract

Scaling depth is a key driver for large language models (LLMs). Yet, as LLMs become deeper, they often suffer from signal degradation: informative features formed in shallow layers are gradually diluted by repeated residual updates, making them harder to recover in deeper layers. We introduce mixture-of-depths attention (MoDA), a mechanism that allows each attention head to attend to sequence KV pairs at the current layer and depth KV pairs from preceding layers. We further describe a hardware-efficient algorithm for MoDA that resolves non-contiguous memory-access patterns, achieving 97.3% of FlashAttention-2's efficiency at a sequence length of 64K. Experiments on 1.5B-parameter models demonstrate that MoDA consistently outperforms strong baselines. Notably, it improves average perplexity by 0.2 across 10 validation benchmarks and increases average performance by 2.11% on 10 downstream tasks, with a negligible 3.7% FLOPs computational overhead. We also find that combining MoDA with post-norm yields better performance than using it with pre-norm. These results suggest that MoDA is a promising primitive for depth scaling. Code is released at https://github.com/hustvl/MoDA .