Star Attention: Efficient LLM Inference over Long Sequences
This addresses efficiency bottlenecks for users deploying LLMs on long sequences, though it is an incremental improvement over existing attention mechanisms.
The paper tackles the problem of costly and slow inference with Transformer-based LLMs on long sequences by introducing Star Attention, a two-phase block-sparse approximation that reduces memory requirements and inference time by up to 11x while preserving 97-100% of accuracy.
Inference with Transformer-based Large Language Models (LLMs) on long sequences is both costly and slow due to the quadratic complexity of the self-attention mechanism. We introduce Star Attention, a two-phase block-sparse approximation that improves computational efficiency by sharding attention across multiple hosts while minimizing communication overhead. In the first phase, the context is processed using blockwise-local attention across hosts, in parallel. In the second phase, query and response tokens attend to all prior cached tokens through sequence-global attention. Star Attention integrates seamlessly with most Transformer-based LLMs trained with global attention, reducing memory requirements and inference time by up to 11x while preserving 97-100% of accuracy.