implementation detail · filed under model architecture
Progressive sequence-length extension for sparse attention
A training schedule that starts sparse attention at 64K sequence length and extends context to 1M tokens.
- 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.
We train the model from scratch with sparse attention at a sequence length of 64K and extend the sequence length to 1M at 34T tokens.
usedunclearin DeepSeek-V4.1-FlashDeepSeek
with sparse attention trained at a sequence length of 64K and context extended to 1M tokens at 34T tokens
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 attentionSparse-attention continued pre-training with joint model and indexer optimizationCross-layer KV and index reuse with statically assigned CSA2 modesFixed-budget sparse-attention selectionFrom-scratch sparse attention training without dense warmupJoint 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