Saved in:
| Main Authors: | , , , , , , , , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2026
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2603.15619 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866915867550285824 |
|---|---|
| author | Zhu, Lianghui Fang, Yuxin Liao, Bencheng Wang, Shijie Cheng, Tianheng Huang, Zilong Chen, Chen Wei, Lai Zeng, Yutao Wang, Ya Lin, Yi Li, Yu Wang, Xinggang |
| author_facet | Zhu, Lianghui Fang, Yuxin Liao, Bencheng Wang, Shijie Cheng, Tianheng Huang, Zilong Chen, Chen Wei, Lai Zeng, Yutao Wang, Ya Lin, Yi Li, Yu Wang, Xinggang |
| contents | 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 . |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2603_15619 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Mixture-of-Depths Attention Zhu, Lianghui Fang, Yuxin Liao, Bencheng Wang, Shijie Cheng, Tianheng Huang, Zilong Chen, Chen Wei, Lai Zeng, Yutao Wang, Ya Lin, Yi Li, Yu Wang, Xinggang Computation and Language Artificial Intelligence 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 . |
| title | Mixture-of-Depths Attention |
| topic | Computation and Language Artificial Intelligence |
| url | https://arxiv.org/abs/2603.15619 |