SageBwd: A Trainable Low-bit Attention

TL;DR

SageBwd introduces a trainable low-bit attention method that effectively matches full-precision attention during pre-training by addressing quantization errors and stability issues, enabling faster inference without performance loss.

Contribution

This work demonstrates that SageBwd can match full-precision attention during pre-training by analyzing quantization errors and stability factors, improving low-bit attention training methods.

Findings

QK-norm is essential for stable training at large tokens per step.

Quantization errors mainly come from the backward-pass score gradient dS.

Reducing tokens per step helps SageBwd match full-precision performance.

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.