Co-Designing Sparse Attention for Efficient Long-Context Inference
Co-Designing Sparse Attention for Efficient Long-Context Inference
NVIDIA co-designed sparse attention mechanisms to achieve efficient long-context reasoning, which significantly reduces the attention computation cost in 128K token contexts. The attention computation cost grows linearly with the context length N, and dense attention mechanisms may consume most of the pre-filled time at 128K tokens.
Agentic workloads, long documents, and large codebases are pushing context lengths from tens of thousands of tokens to hundreds of thousands. At those lengths,...
Agentic workloads, long documents, and large codebases are pushing context lengths from tens of thousands of tokens to hundreds of thousands. At those lengths, attention becomes the most expensive part of a transformer layer. Dense attention scores every query against every key, so its cost per query grows with the context length N, and at 128K tokens it can take most of the prefill time in a…
Source: NVIDIA Technical Blog · developer.nvidia.com