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

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 ``drop-in'' QAT -- naively combining an FP4 forward pass with 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 plus 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.

Modern AI models can generate text, images, and videos, but running them is expensive because they require large amounts of computation and memory. Using 4-bit numbers can reduce this cost, but an important part of these models, attention, often loses quality under such aggressive compression. In this work, we show how to make 4-bit attention reliable. Instead of simply compressing a trained model, we let the model adapt to 4-bit computation through a short additional training stage. We also identify two changes needed to keep this training stable and implement efficient GPU kernels for the method. Our method preserves the quality of full-precision attention across video generation and language models while making attention faster on modern GPUs. This can help reduce the cost and energy needed to serve large AI models.