Hardware-Efficient Attention for Fast Decoding
This addresses latency and throughput issues in large language model serving, offering incremental improvements for faster online inference.
The paper tackles the bottleneck in LLM decoding caused by loading the key-value cache from memory, which increases latency, by redesigning attention to improve hardware efficiency. It introduces Grouped-Tied Attention (GTA) and Grouped Latent Attention (GLA), showing that GTA uses roughly half the KV cache while matching quality, and GLA achieves up to 2× faster decoding and higher throughput in benchmarks.
LLM decoding is bottlenecked for large batches and long contexts by loading the key-value (KV) cache from high-bandwidth memory, which inflates per-token latency, while the sequential nature of decoding limits parallelism. We analyze the interplay among arithmetic intensity, parallelization, and model quality and question whether current architectures fully exploit modern hardware. This work redesigns attention to perform more computation per byte loaded from memory to maximize hardware efficiency without trading off parallel scalability. We first propose Grouped-Tied Attention (GTA), a simple variant that combines and reuses key and value states, reducing memory transfers without compromising model quality. We then introduce Grouped Latent Attention (GLA), a parallel-friendly latent attention paired with low-level optimizations for fast decoding while maintaining high model quality. Experiments show that GTA matches Grouped-Query Attention (GQA) quality while using roughly half the KV cache and that GLA matches Multi-head Latent Attention (MLA) and is easier to shard. Our optimized GLA kernel is up to 2$\times$ faster than FlashMLA, for example, in a speculative decoding setting when the query length exceeds one. Furthermore, by fetching a smaller KV cache per device, GLA reduces end-to-end latency and increases throughput in online serving benchmarks by up to 2$\times$.