single mention · model architecture
CSA2 Full, Reindex, and Reuse modes
CSA2 assigns each decoder layer one of three statically fixed operating modes — Full, Reindex, or Reuse — that determine how it obtains its main KV and Top-K indices.1 Full Mode computes its own KV and indices, Reindex Mode recomputes indices over shared KV, and Reuse Mode attends with a preceding layer's KV and indices without any indexer computation.2
Trade-off Sharing main KV and indices across layers cuts cache storage and indexer computation, with Reindex Mode preserving cache sharing while letting selections change across layers.3
How this definition was written
Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.
The layer reuses the most recent available main KV and the latest Top-K indices computed against that main KV by a preceding layer in Full or Reindex Mode
— DeepSeek V41 Tech ReportIt performs attention using this selection without computing indexer Q or evaluating index scores
— DeepSeek V41 Tech ReportSharing main KV and indexer K reduces cache storage, while reusing Top-K indices avoids additional indexer computation
— DeepSeek V41 Tech Report
Also called CSA2 Full Mode, Full Mode, Full, Reindex, and Reuse.
Filed under model architecture › token mixer › sparse attention