Attn-QAT: 4-Bit Attention With Quantization-Aware Training

TL;DR

This paper introduces Attn-QAT, a novel quantization-aware training method for 4-bit attention that stabilizes training and maintains model quality, enabling faster inference on GPUs with minimal accuracy loss.

Contribution

It presents the first systematic approach to 4-bit attention QAT, identifying key principles for stability and implementing efficient kernels for training and inference.

Findings

Attn-QAT recovers FP4 attention quality without outlier heuristics.

Achieves up to 1.5x speedup on RTX 5090.

Stabilizes 4-bit attention training with new principles.

Abstract

Achieving reliable 4-bit attention is a prerequisite for end-to-end FP4 computation on emerging FP4-capable GPUs, yet attention remains the main obstacle due to FP4's tiny dynamic range and attention's heavy-tailed activations. This paper presents the first systematic study of 4-bit quantization-aware training (QAT) for attention. We find that "drop-in" QAT, which naively combines an FP4 forward pass with a high-precision Flash Attention (FA)-style backward pass, leads to training instability. We identify two key principles for stable FP4 attention: (1) matching low-precision recomputation of attention scores in the backward pass, and (2) resolving implicit precision assumptions in FA's gradient calculation. Based on these insights, we propose Attn-QAT and implement fused Triton kernels for training as well as FP4 inference kernels. Across diffusion and language models, Attn-QAT recovers the quality drop from FP4 attention without explicit outlier-mitigation heuristics used in prior FP4 attention, and delivers up to a 1.5x speedup on an RTX 5090. Video demos can be found at https://drive.google.com/drive/folders/190F6xbBDUF2kGQYIcXBt3ehSYij5jlim?usp=sharing.