Native Hybrid Attention for Efficient Sequence Modeling
This addresses efficiency and accuracy bottlenecks in sequence modeling for AI applications, offering a tunable solution that is incremental but with strong specific gains.
The paper tackles the quadratic complexity of Transformers and recall accuracy issues of linear attention by introducing Native Hybrid Attention (NHA), a hybrid architecture that integrates linear and full attention, achieving superior performance on recall-intensive and commonsense reasoning tasks while enabling efficient pretraining of LLMs.
Transformers excel at sequence modeling but face quadratic complexity, while linear attention offers improved efficiency but often compromises recall accuracy over long contexts. In this work, we introduce Native Hybrid Attention (NHA), a novel hybrid architecture of linear and full attention that integrates both intra \& inter-layer hybridization into a unified layer design. NHA maintains long-term context in key-value slots updated by a linear RNN, and augments them with short-term tokens from a sliding window. A single \texttt{softmax attention} operation is then applied over all keys and values, enabling per-token and per-head context-dependent weighting without requiring additional fusion parameters. The inter-layer behavior is controlled through a single hyperparameter, the sliding window size, which allows smooth adjustment between purely linear and full attention while keeping all layers structurally uniform. Experimental results show that NHA surpasses Transformers and other hybrid baselines on recall-intensive and commonsense reasoning tasks. Furthermore, pretrained LLMs can be structurally hybridized with NHA, achieving competitive accuracy while delivering significant efficiency gains. Code is available at https://github.com/JusenD/NHA.