implementation detail · filed under model architecture
Search-Based Sliding-Window Attention Pattern
Selects layers for conversion to sliding-window attention through a search-based adaptation, retaining full attention in the others.
Also called SWA Pattern (Search-Based).
- source
- 1
- model
- 1
- labs adopt it
- 0
- strongest
- evaluated
How sources treat it
One count per evidence span, weakest treatment to strongest.
evaluated 1
Documented in
Evidence
1 span quoted from the sources, strongest treatment first.
We introduce a search-based adaptation method that identifies the optimal subset of layers for SWA conversion while retaining full attention in the remaining layers.
evaluatedmodel architecturein GLM-5Z.ai
Filed alongside
Other methods under model architecture :: token mixer :: hybrid layer stacking.
Hybrid AttentionMamba-2, Mixture-of-Experts, and Selective Attention HybridHybrid Mamba-TransformerHybrid Mamba-AttentionGated DeltaNet and Gated AttentionHybrid Mamba-Transformer Mixture-of-Experts Layer LayoutDense Attention FallbackGated Attention and Sliding-Window Attention ConfigurationGated DeltaNet and Full AttentionGated DeltaNet and Qwen Sparse Attentionhybrid Gated DeltaNet + sparse MoE architectureHybrid sparse mixture-of-experts Transformer architecture