TrainSDC: Characterizing and Mitigating Silent Data Corruption in Large Language Model Training
2026-08-31 • Machine Learning
Machine Learning
AI summaryⓘ
The authors studied how silent data corruption (SDC) affects the training of large language models (LLMs), focusing on which parts of the Transformer computations are most vulnerable during training. They found that errors in the forward pass, especially in the Q/K paths, cause persistent problems, while in the backward pass, the vulnerability depends more on the distribution of gradient exponents than on location. Based on this, they developed TrainSDC, a method that selectively protects these sensitive parts and adjusts gradients to reduce error impact. Tests on two LLMs showed TrainSDC keeps training close to normal while adding only a small extra runtime cost.
silent data corruptionlarge language modelsTransformerforward passbackward passQ/K pathgradient exponentrecomputationfault injectiontraining robustness
Authors
Zhipeng Xia, Haotian Xu, Siyu Yun, Liqi Lin, Hu Liu, Yu Li, Cheng Zhuo
Abstract
LLM training is increasingly vulnerable to silent data corruption (SDC), yet existing protection methods largely treat Transformer computations uniformly because their vulnerability remains poorly understood. We present the first systematic characterization of SDC vulnerability across major computation interfaces in both the forward and backward passes of Transformer training. Our analysis reveals two distinct error propagation mechanisms: forward-pass vulnerability is highly location dependent, with faults on the Q/K path producing persistent training deviations, whereas backward-pass vulnerability is largely governed by gradient exponent distributions rather than computation locations. Motivated by these observations, we propose TrainSDC, a characterization-guided protection framework consisting of Q/K-path recomputation, residual-gain monitoring, and exponent-aware gradient scaling. Experiments on Llama 3.2-1B and Qwen3-0.6B show that TrainSDC maintains training behavior close to fault-free execution under both sparse and dense fault injection while introducing only 1.65%-6.76% runtime overhead.