Reproduce “Balanced and Elastic End-to-end Training of Dynamic LLMs” 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
Mohamed Wahib, Muhammed Abdullah Soyturk, Didem Unat
DynMo balances changing workloads in pipeline-parallel large-language-model training by profiling layer execution times and GPU memory, then redistributing layers across stages; when memory permits, it can also consolidate work onto fewer GPUs. In a 40-layer setting, layer-freezing schemes show up to 40% GPU idleness. For LLaMA-MoE-3.5B continual training, DynMo reaches 1.23× static-baseline throughput, in the reported setup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.