Reproduce “MegaScale-Omni: A Hyper-Scale, Workload-Resilient System for MultiModal LLM Training in Production” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Chunyu Xue, Yangrui Chen, Jianyu Jiang, Ningxin Zheng, Junda Feng, Jingji Chen, Shixiong Zhao, Shen Yan and 9 more
MegaScale-Omni, developed by ByteDance Seed and Shanghai Jiao Tong University, introduces an MLLM training system with an encoder-LLM multiplexing scheme and workload-resilient orchestration, achieving up to 7.57x throughput improvement and a 2.21x memory reduction compared to baselines in hyper-scale production environments with dynamic multimodal workloads. It effectively scales to sequence lengths up to 512K and improves Model FLOPs Utilization by up to 23.0%.
Key insights
- Decoupling parallelization and resource allocation between encoders and the LLM backbone is critical for efficiently handling dynamic multimodal workloads.
- Flexible colocation and temporal orchestration of encoders and the LLM enable reactive adaptation to workload shifts, mitigating bottlenecks and device idling.
- Specialized parallelization (e.g., LSSP for dynamic sequence lengths) and workload-aware data management (e.g., grouped reordering) are essential for achieving hyper-scale performance stability.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.