Saved in:
| Main Authors: | , , , , , , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2026
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2605.25451 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866911715329835008 |
|---|---|
| author | Zhang, Zili Yang, Chengxu Zhang, Shenglong Wang, Chenyu Zhang, Yufan Dai, Tuo Li, Zhouyang Ge, Yuhong Jin, Chao Jin, Xin Liu, Yuliang |
| author_facet | Zhang, Zili Yang, Chengxu Zhang, Shenglong Wang, Chenyu Zhang, Yufan Dai, Tuo Li, Zhouyang Ge, Yuhong Jin, Chao Jin, Xin Liu, Yuliang |
| contents | Training multimodal large language models (MLLMs) is challenged by both model and data heterogeneity. Existing systems redesign the training pipeline to address these challenges, but remain bound by a Pareto frontier between compute and memory efficiency, improving one only at the expense of the other. We present BigMac, a new training pipeline for multimodal LLMs. The core idea of BigMac is to elegantly nest the encoder and generator computation into the original LLM pipeline, forming a dependency-safe nested pipeline structure. With this design, BigMac reduces the activation memory complexity of the encoder and generator to O(1) while keeping the activation memory complexity of the LLM unchanged. At the same time, it achieves the same computational efficiency as the idealized setting with unlimited memory. As a result, BigMac breaks the Pareto frontier between computational efficiency and memory usage, enabling simultaneous optimization of both computation and memory in MLLM training. We evaluate BigMac on multiple MLLMs and training workloads. Experimental results show that BigMac achieves a 1.08$\times$-1.9$\times$ training speedup over baseline systems while maintaining stable memory usage as batch size increases. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2605_25451 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | BigMac: Breaking the Pareto Frontier of Compute and Memory in Multimodal LLM Training Zhang, Zili Yang, Chengxu Zhang, Shenglong Wang, Chenyu Zhang, Yufan Dai, Tuo Li, Zhouyang Ge, Yuhong Jin, Chao Jin, Xin Liu, Yuliang Machine Learning Training multimodal large language models (MLLMs) is challenged by both model and data heterogeneity. Existing systems redesign the training pipeline to address these challenges, but remain bound by a Pareto frontier between compute and memory efficiency, improving one only at the expense of the other. We present BigMac, a new training pipeline for multimodal LLMs. The core idea of BigMac is to elegantly nest the encoder and generator computation into the original LLM pipeline, forming a dependency-safe nested pipeline structure. With this design, BigMac reduces the activation memory complexity of the encoder and generator to O(1) while keeping the activation memory complexity of the LLM unchanged. At the same time, it achieves the same computational efficiency as the idealized setting with unlimited memory. As a result, BigMac breaks the Pareto frontier between computational efficiency and memory usage, enabling simultaneous optimization of both computation and memory in MLLM training. We evaluate BigMac on multiple MLLMs and training workloads. Experimental results show that BigMac achieves a 1.08$\times$-1.9$\times$ training speedup over baseline systems while maintaining stable memory usage as batch size increases. |
| title | BigMac: Breaking the Pareto Frontier of Compute and Memory in Multimodal LLM Training |
| topic | Machine Learning |
| url | https://arxiv.org/abs/2605.25451 |