[2603.02170] SageBwd: A Trainable Low-bit Attention

Computer Science > Machine Learning arXiv:2603.02170 (cs) [Submitted on 2 Mar 2026] Title:SageBwd: A Trainable Low-bit Attention Authors:Jintao Zhang, Marco Chen, Haoxu Wang, Kai Jiang, Ion Stoica, Joseph E. Gonzalez, Jianfei Chen, Jun Zhu View a PDF of the paper titled SageBwd: A Trainable Low-bit Attention, by Jintao Zhang and 7 other authors View PDF HTML (experimental) Abstract:Low-bit attention, such as SageAttention, has emerged as an effective approach for accelerating model inference, but its applicability to training remains poorly understood. In prior work, we introduced SageBwd, a trainable INT8 attention that quantizes six of seven attention matrix multiplications while preserving fine-tuning performance. However, SageBwd exhibited a persistent performance gap to full-precision attention (FPA) during pre-training. In this work, we investigate why this gap occurs and demonstrate that SageBwd matches full-precision attention during pretraining. Through experiments and theoretical analysis, we reach a few important insights and conclusions: (i) QK-norm is necessary for stable training at large tokens per step, (ii) quantization errors primarily arise from the backward-pass score gradient dS, (iii) reducing tokens per step enables SageBwd to match FPA performance in pre-training, and (iv) K-smoothing remains essential for training stability, while Q-smoothing provides limited benefit during pre-training. Subjects: Machine Learning (cs.LG); Artificial Intelligence (...