UniMoD: Efficient Unified Multimodal Transformers with Mixture-of-Depths
Fuente:
arXiv
Salvato in:
| Autori principali: | , , |
|---|---|
| Natura: | Preprint |
| Pubblicazione: |
2025
|
| Soggetti: | |
| Accesso online: | |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866929707008655360 |
|---|---|
| author | Mao, Weijia Yang, Zhenheng Shou, Mike Zheng |
| author_facet | Mao, Weijia Yang, Zhenheng Shou, Mike Zheng |
| contents | Unified multimodal transformers, which handle both generation and understanding tasks within a shared parameter space, have received increasing attention in recent research. Although various unified transformers have been proposed, training these models is costly due to redundant tokens and heavy attention computation. In the past, studies on large language models have demonstrated that token pruning methods, such as Mixture of Depths (MoD), can significantly improve computational efficiency. MoD employs a router to select the most important ones for processing within a transformer layer. However, directly applying MoD-based token pruning to unified transformers will result in suboptimal performance because different tasks exhibit varying levels of token redundancy. In our work, we analyze the unified transformers by (1) examining attention weight patterns, (2) evaluating the layer importance and token redundancy, and (3) analyzing task interactions. Our findings reveal that token redundancy is primarily influenced by different tasks and layers. Building on these findings, we introduce UniMoD, a task-aware token pruning method that employs a separate router for each task to determine which tokens should be pruned. We apply our method to Show-o and Emu3, reducing training FLOPs by approximately 15% in Show-o and 40% in Emu3, while maintaining or improving performance on several benchmarks. Code will be released at https://github.com/showlab/UniMoD. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2502_06474 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | UniMoD: Efficient Unified Multimodal Transformers with Mixture-of-Depths Mao, Weijia Yang, Zhenheng Shou, Mike Zheng Computer Vision and Pattern Recognition Unified multimodal transformers, which handle both generation and understanding tasks within a shared parameter space, have received increasing attention in recent research. Although various unified transformers have been proposed, training these models is costly due to redundant tokens and heavy attention computation. In the past, studies on large language models have demonstrated that token pruning methods, such as Mixture of Depths (MoD), can significantly improve computational efficiency. MoD employs a router to select the most important ones for processing within a transformer layer. However, directly applying MoD-based token pruning to unified transformers will result in suboptimal performance because different tasks exhibit varying levels of token redundancy. In our work, we analyze the unified transformers by (1) examining attention weight patterns, (2) evaluating the layer importance and token redundancy, and (3) analyzing task interactions. Our findings reveal that token redundancy is primarily influenced by different tasks and layers. Building on these findings, we introduce UniMoD, a task-aware token pruning method that employs a separate router for each task to determine which tokens should be pruned. We apply our method to Show-o and Emu3, reducing training FLOPs by approximately 15% in Show-o and 40% in Emu3, while maintaining or improving performance on several benchmarks. Code will be released at https://github.com/showlab/UniMoD. |
| title | UniMoD: Efficient Unified Multimodal Transformers with Mixture-of-Depths |
| topic | Computer Vision and Pattern Recognition |
| url | https://arxiv.org/abs/2502.06474 |