🤖 AI Summary
To address memory and throughput bottlenecks in large language models (LLMs) during long-context reasoning—caused by uniform sliding-window attention—this paper proposes the first automated sparse attention compression framework. It identifies a bimodal, heterogeneous behavior across attention heads in long contexts: “expansion” (broad context coverage) and “local fixation” (narrow, position-stable focus), and designs a hybrid sparse mechanism enabling dynamic, head-wise, layer-wise, and sequence-length–adaptive sparsity patterns. Leveraging performance profiling–driven search space construction, joint optimization of sparsity patterns and scaling laws, and automatic policy generation, the framework integrates seamlessly with FlashAttention2 and vLLM. Experiments demonstrate: 3.9× longer supported context length; 1.5–7.1× higher retrieval accuracy; maximum performance degradation reduced from 9%–36% to ≤5%; 1.2–1.4× lower GPU memory consumption; and 6.6–8.2× higher decoding throughput.
📝 Abstract
Sparse attention can effectively mitigate the significant memory and throughput demands of Large Language Models (LLMs) in long contexts. Existing methods typically employ a uniform sparse attention mask, applying the same sparse pattern across different attention heads and input lengths. However, this uniform approach fails to capture the diverse attention patterns inherent in LLMs, ignoring their distinct accuracy-latency trade-offs. To address this challenge, we propose the Mixture of Attention (MoA), which automatically tailors distinct sparse attention configurations to different heads and layers. MoA constructs and navigates a search space of various attention patterns and their scaling rules relative to input sequence lengths. It profiles the model, evaluates potential configurations, and pinpoints the optimal sparse attention compression plan. MoA adapts to varying input sizes, revealing that some attention heads expand their focus to accommodate longer sequences, while other heads consistently concentrate on fixed-length local contexts. Experiments show that MoA increases the effective context length by $3.9 imes$ with the same average attention span, boosting retrieval accuracy by $1.5-7.1 imes$ over the uniform-attention baseline across Vicuna-{7B,13B}, and Llama3-{8B,70B} models. Moreover, MoA narrows the capability gaps between sparse and dense models, reducing the maximum relative performance drop from $9%-36%$ to within $5%$ across two long-context understanding benchmarks. MoA achieves a $1.2-1.4 imes$ GPU memory reduction, boosting decode throughput by $6.6-8.2 imes$ and $1.7-1.9 imes$ compared to FlashAttention2 and vLLM, with minimal impact on performance. Our code is available at url{https://github.com/thu-nics/MoA}.