Reproduce “Online Dynamic Batching with Formal Guarantees for LLM Training” 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
Dian Li, Zekun Wang, Yaoru Wang, Jiahong Yan
ODB forms variable-size batches at the DataLoader boundary, after preprocessing, tokenization, and visual-token expansion reveal each sample’s length, then aligns batch steps across distributed ranks without changing the model or attention kernel. On Qwen3-VL-8B Full FT with ShareGPT4o, it reaches 5.83 ± 0.04 emitted samples/s versus 2.37 ± 0.00 for Standard, a 2.46× speedup; quality is Standard-comparable broadly, but metrics differ.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.