English

Optimal low-rank stochastic gradient estimation for LLM training

Machine Learning 2026-03-24 v1

Abstract

Large language model (LLM) training is often bottlenecked by memory constraints and stochastic gradient noise in extremely high-dimensional parameter spaces. Motivated by empirical evidence that many LLM gradient matrices are effectively low-rank during training, we present an unbiased, memory-efficient, low-rank matrix estimator with the lowest variance that is applicable across common stochastic gradient estimation paradigms. The core idea is to project a high-dimensional stochastic gradient estimator onto a random low-dimensional subspace and lift it back, reducing memory while keeping the estimator unbiased and controlling mean-squared error via an optimally designed projection distribution, including Haar--Stiefel projections. The projection distribution is derived by solving a constrained functional optimization problem, yielding an optimal random projector that guides algorithm design. Empirically, the resulting low-rank gradient estimators deliver both practical memory savings and improved training behavior. In RoBERTa-large fine-tuning, our method attains the lowest peak GPU memory among compared methods (e.g., 3.83GB versus 16.7GB for full BP) while remaining competitive in accuracy; in autoregressive LLM pretraining (LLaMA-20M/60M/100M), our method outperforms the traditional methods, supporting the benefit of the proposed optimal projection strategy.

Keywords

Cite

@article{arxiv.2603.20632,
  title  = {Optimal low-rank stochastic gradient estimation for LLM training},
  author = {Zehao Li and Tao Ren and Zishi Zhang and Xi Chen and Yijie Peng},
  journal= {arXiv preprint arXiv:2603.20632},
  year   = {2026}
}
R2 v1 2026-07-01T11:30:59.660Z