Memory-Efficient LLM Training with Dynamic Sparsity: From Stability to Practical Scaling
For LLM practitioners, SMET addresses a key instability bottleneck in DST, making sparse training a practical alternative to dense training for memory efficiency.
Dynamic Sparse Training (DST) for LLMs suffers from optimization instability (loss spikes) due to cold-start issues for regrown parameters. The proposed SMET method stabilizes training with optimizer warm-up and density-aware learning-rate scaling, reducing memory by storing gradients/states only for active parameters, enabling stable and scalable sparse pre-training.
Dynamic Sparse Training (DST) offers a promising paradigm for improving the training and inference efficiency of deep neural networks; however, we find that in large language model training, DST can suffer from optimization instability, manifested as loss spikes after topology updates. In this work, we show that the naive use of standard Adam-based optimizers leads to a cold-start issue for newly regrown parameters, resulting in excessively large updates and disrupted training dynamics. To address this issue, we propose Sparse Memory-Efficient Training (SMET), which stabilizes DST with optimizer warm-up and improves training progress through density-aware learning-rate scaling. SMET further reduces memory consumption by storing gradients and optimizer states only for active parameters. We provide a theoretical analysis of the update behaviors under SMET, showing improved optimization stability. Extensive experiments demonstrate that SMET enables stable, scalable, and memory-efficient sparse pre-training of LLMs, paving the way for sparse training as a practical alternative to dense training. Our code is publicly available at: https://github.com/QiaoXiao7282/SMET.