Reproduce “The Role of Quantization in Loss Spikes” 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
Zaiqian Chen, Aditya Somasundaram, Sophie Li, Sean McLeish, John Kirchenbauer, Tom Goldstein, Micah Goldblum
The authors test whether reduced precision contributes to delayed loss spikes in language-model pretraining. They propose that growing output logits make low-precision calculations lose accuracy in differences that determine next-token probabilities. In 28M-parameter, peak-learning-rate-0.1 experiments, 5/5 BF16-mixed and BF16-true runs time-bombed, versus 0/5 FP32 runs; this does not establish the cause of failures in real-world training.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.