Tucker Attention: A generalization of approximate attention mechanisms
This work addresses the memory efficiency challenge in transformer models for AI researchers and practitioners, offering a novel generalization that unifies existing methods like GQA and MLA, though it appears incremental as it builds on low-rank approximation techniques.
The paper tackles the problem of reducing the memory footprint of self-attention mechanisms in transformers by proposing Tucker Attention, a generalized factorization strategy that requires an order of magnitude fewer parameters while achieving comparable validation metrics in LLM and ViT test cases.
The pursuit of reducing the memory footprint of the self-attention mechanism in multi-headed self attention (MHA) spawned a rich portfolio of methods, e.g., group-query attention (GQA) and multi-head latent attention (MLA). The methods leverage specialized low-rank factorizations across embedding dimensions or attention heads. From the point of view of classical low-rank approximation, these methods are unconventional and raise questions of which objects they really approximate and how to interpret the low-rank behavior of the resulting representations. To answer these questions, this work proposes a generalized view on the weight objects in the self-attention layer and a factorization strategy, which allows us to construct a parameter efficient scheme, called Tucker Attention. Tucker Attention requires an order of magnitude fewer parameters for comparable validation metrics, compared to GQA and MLA, as evaluated in LLM and ViT test cases. Additionally, Tucker Attention~encompasses GQA, MLA, MHA as special cases and is fully compatible with flash-attention and rotary position embeddings (RoPE). This generalization strategy yields insights of the actual ranks achieved by MHA, GQA, and MLA, and further enables simplifications for MLA.