Reproduce “GradientStabilizer:Fix the Norm, Not the Gradient” 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
Tianjin Huang, Zhangyang Wang, Haotian Hu, Zhenyu Zhang, Gaojie Jin, Xiang Li, Li Shen, Jiaxing Shang and 5 more
GradientStabilizer stabilizes gradient magnitudes during neural-network training while preserving each gradient’s direction. It replaces the norm using exponential moving averages of past norms and squared norms, rather than applying a clipping threshold. With zero-initialized norm averages and qualifying decay rates, the transformed gradient has a bound independent of its raw norm. On FP4 LLaMA-350M, validation perplexity was 18.84 versus ZClip’s 21.39 after 6.6B C4 tokens.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.