implementation detail · filed under model architecture
From-scratch sparse attention training without dense warmup
Training sparse attention from the start at a long sequence length without a preceding dense-attention warmup.
- source
- 1
- model
- 1
- lab adopt it
- 1
- strongest
- used
How sources treat it
One count per evidence span, weakest treatment to strongest.
used 1
Documented in
Evidence
1 span quoted from the sources, strongest treatment first.
Sparse attention is trained from scratch at a sequence length of 64K, without any dense attention warmup stages.
usedotherin DeepSeek-V4.1-FlashDeepSeek
Filed alongside
Other methods under model architecture :: token mixer :: sparse attention.
DeepSeek Sparse AttentionCompressed Sparse AttentionQwen Sparse AttentionSparse attentionHeavily Compressed AttentionGated DeepSeek Sparse AttentionCSA2 Full ModeKV-outer sparse attentionProgressive sequence-length extension for sparse attentionSparse-attention continued pre-training with joint model and indexer optimizationCross-layer KV and index reuse with statically assigned CSA2 modesFixed-budget sparse-attention selectionJoint backbone and indexer training under sparse attentionNative Sparse AttentionNatively trained sparsityNoPE sparse multi-head latent attentionQSA micro-block compression at ratio 4Sequential block processingSparse retrieval over long contextsSparse softmax attentionToken-wise compressionTwo-stage introduction of sparse attentionTwo-stage sparse attention