implementation detail · filed under model architecture
Dense Warm-up Stage
A training stage that retains dense attention and freezes all parameters except the lightning indexer.
- sources
- 2
- model
- 1
- lab adopt it
- 1
- strongest
- used
How sources treat it
One count per evidence span, weakest treatment to strongest.
used 2
Documented in
Evidence
2 spans quoted from the sources, strongest treatment first.
In this stage, we keep dense attention and freeze all model parameters except for the lightning indexer.
usedoptimizationin DeepSeek-V3.2DeepSeek
In this stage, we keep dense attention and freeze all model parameters except for the lightning indexer.
usedoptimizationin DeepSeek-V3.2DeepSeek
Filed alongside
Other methods under model architecture :: token mixer :: sparse attention :: sparse attention indexer.
Compressed Sparse Attention 2Hierarchical Sparse IndexerIndexCacheIndexShareLightning IndexerReindex ModeReuse ModeFine-grained token selectionIndex BranchIndexer WarmupReuse QSA index selection across speculative decoding stepsAverage poolingBlock-causal scoringCompressed lightweight indexerCross-stage shared-state management for attention reuseDetached indexer-input optimizationIndex Branch outputIndex Branch value headIndexPoolMQA indexerSingle-head index key