Weighted Grouped Query Attention in Transformers

TL;DR

This paper introduces Weighted Grouped-Query Attention (WGQA), a novel variation of GQA that uses learnable weights for key and value heads, improving performance with minimal inference overhead in transformer models.

Contribution

The paper proposes WGQA, which incorporates learnable weights into GQA, enhancing performance during finetuning without increasing inference costs.

Findings

WGQA achieves 0.53% improvement over GQA.

Performance converges to traditional MHA with no extra inference overhead.

Scaling laws are demonstrated across T5-small and T5-base architectures.

Abstract

The attention mechanism forms the foundational blocks for transformer language models. Recent approaches show that scaling the model achieves human-level performance. However, with increasing demands for scaling and constraints on hardware memory, the inference costs of these models remain high. To reduce the inference time, Multi-Query Attention (MQA) and Grouped-Query Attention (GQA) were proposed in (Shazeer, 2019) and (Ainslieet al., 2023) respectively. In this paper, we propose a variation of Grouped-Query Attention, termed Weighted Grouped-Query Attention (WGQA). We introduced new learnable parameters for each key and value head in the T5 decoder attention blocks, enabling the model to take a weighted average during finetuning. Our model achieves an average of 0.53% improvement over GQA, and the performance converges to traditional Multi-head attention (MHA) with no additional overhead during inference. We evaluated the introduction of these parameters and subsequent finetuning informs the model about the grouping mechanism during training, thereby enhancing performance. Additionally, we demonstrate the scaling laws in our analysis by comparing the results between T5-small and T5-base architecture.