implementation detail · filed under model architecture
Reuse QSA index selection across speculative decoding steps
Avoids rerunning the indexer during speculative draft steps by reusing previously selected indices.
- sources
- 2
- model
- 1
- lab adopt it
- 1
- strongest
- core
How sources treat it
One count per evidence span, weakest treatment to strongest.
used 1core 1
Documented in
Evidence
2 spans quoted from the sources, strongest treatment first.
So the draft decode steps stop running the indexer altogether.
coreunclearin Qwen3.8-Flash-NextSGLang
we follow GLM (GLM-5-Team, 2026) and reuse the top- indices across speculative decoding steps
usedunclearin Qwen3.8-Flash-NextQwen
Filed alongside
Other methods under model architecture :: token mixer :: sparse attention :: sparse attention indexer.
Compressed Sparse Attention 2Hierarchical Sparse IndexerIndexCacheIndexShareLightning IndexerReindex ModeReuse ModeDense Warm-up StageFine-grained token selectionIndex BranchIndexer WarmupAverage poolingBlock-causal scoringCompressed lightweight indexerCross-stage shared-state management for attention reuseDetached indexer-input optimizationIndex Branch outputIndex Branch value headIndexPoolMQA indexerSingle-head index key