Xiaohongshu dots infra open-sources BigMac: pushing past the Pareto frontier of multimodal large-model training
Original title:小红书 dots infra 开源BigMac : 突破多模态大模型训练的帕累托前沿
AI summary
Xiaohongshu open-sourced BigMac, a new pipeline-parallel training paradigm for native multimodal scenarios. Through dependency-safe nested pipelines, it efficiently enables multimodal pipeline training without adding LLM pipeline bubbles and while keeping activation memory bounded. Compared with baselines, BigMac improves training speed by 1.08-1.9x and maintains stable memory usage as global batch size grows. It is now used in production training of dots multimodal models.
Key Takeaways
- Xiaohongshu open-sourced BigMac,a new pipeline-parallel training paradigm for native multimodal…
- Through dependency-safe nested pipelines,it efficiently enables multimodal pipeline training…
- Compared with baselines,BigMac improves training speed by 1.
- 9x and maintains stable memory usage as global batch size grows.
- It is now used in production training of dots multimodal models.
Quick Facts
- Published
- Updated
- Source
- 公众号:小红书技术(dots.llm)
- Category
- Products & Apps
- AI score
- 70
通过在成熟的LLM流水线中嵌入依赖安全的编码器与生成器计算,BigMac让多模态大模型训练不再需要在速度和显存之间做非此即彼的选择,适合正在纠结训练策略的工程团队检查调度逻辑。
Source & copyright
This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.
How to Cite
When citing this article, keep the title, date, and permalink, and credit Orange News as the source.
Orange News: Xiaohongshu dots infra open-sources BigMac: pushing past the Pareto frontier of multimodal large-model training. July 22, 2026. https://new.nat6.net/en/news/cmsdu108a00ucrofzwlsyffxo/