GraSS: Scalable Data Attribution with Gradient Sparsification and Sparse Projection

Han Zhao (University of Illinois, Urbana Champaign) · Jiaqi Ma (University of Illinois Urbana-Champaign) · Pingbang Hu (University of Illinois Urbana-Champaign) · Joseph Melkonian (Washington University, Saint Louis) · Weijing Tang (Carnegie Mellon University)

billion-scale modelscomputational costdata influence fidelityfactgrassgradient compression algorithmgradient-based data attributiongrassinfluence functions linear layers memory cost model retrainingper-sample gradient computationsparsitysub-linear complexitythroughput

Gradient-based data attribution methods, such as influence functions, are critical for understanding the impact of individual training samples without requiring repeated model retraining. However, their scalability is often limited by the high computational and memory costs associated with per-sample gradient computation. In this work, we propose **GraSS**, a novel gradient compression algorithm and its variants **FactGraSS** for linear layers specifically, that explicitly leverage the inherent sparsity of per-sample gradients to achieve sub-linear space and time complexity. Extensive experiments demonstrate the effectiveness of our approach, achieving substantial speedups while preserving data influence fidelity. In particular, **FactGraSS** achieves up to 165% faster throughput on billion-scale models compared to the previous state-of-the-art baselines. Our code is publicly available at https://github.com/TRAIS-Lab/GraSS.