The authors first analyzed the temporal patterns of indexer scores in DeepSeek-V4's sparse attention (DSA).
Temporal indexer prediction speeds up long-context sparse attention by 14.9%
SPIN uses per-iteration statistics from previous decoding steps to predict which KV blocks to skip, reducing indexer overhead while preserving task quality.
Independent
Yao Fu · Cyrus Chang · Ritchie Zhao · Bryce Long · Yueying Li · Mahdi Kamani · +8 more
Research Digest··2 min read
The authors introduce SPIN, a training-free method that sparsifies the indexer in sparse attention by predicting block importance from past scores.
Why this paper
Independent
In one line
SPIN predicts which KV cache blocks are important using past scores, reducing indexer overhead by 30-40% and improving throughput by up to 14.9%.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§