Is Flash Attention Stable?

TL;DR

This paper investigates the numerical stability of Flash Attention, revealing it causes more deviation than baseline methods but has a limited impact on training stability compared to low-precision training.

Contribution

It introduces a framework for analyzing numeric deviation effects and applies it to assess Flash Attention's stability in large-scale training.

Findings

Flash Attention exhibits roughly ten times more numeric deviation than Baseline Attention at BF16.

The deviation in Flash Attention is 2-5 times less impactful than low-precision training.

The framework helps quantify the impact of numeric deviation on training stability.

Abstract

Training large-scale machine learning models poses distinct system challenges, given both the size and complexity of today's workloads. Recently, many organizations training state-of-the-art Generative AI models have reported cases of instability during training, often taking the form of loss spikes. Numeric deviation has emerged as a potential cause of this training instability, although quantifying this is especially challenging given the costly nature of training runs. In this work, we develop a principled approach to understanding the effects of numeric deviation, and construct proxies to put observations into context when downstream effects are difficult to quantify. As a case study, we apply this framework to analyze the widely-adopted Flash Attention optimization. We find that Flash Attention sees roughly an order of magnitude more numeric deviation as compared to Baseline Attention at BF16 when measured during an isolated forward pass. We then use a data-driven analysis based on the Wasserstein Distance to provide upper bounds on how this numeric deviation impacts model weights during training, finding that the numerical deviation present in Flash Attention is 2-5 times less significant than low-precision training.