SPIN predicts KV-block importance from past indexer scores, skipping 30–40% of indexer input blocks on DeepSeek-V4 and lifting vLLM throughput by 14.9%
SPIN (Shadow Predictive Indexer) uses per-layer statistics of indexer scores from previous decoding iterations, maintaining EMAs along vertical and diagonal patterns to predict KV-block importance and skip context blocks before the indexer; on DeepSeek-V4 Flash and Pro across long-context and agentic benchmarks it preserves task quality at 30–40% block sparsity, and integrated into vLLM it improves output throughput by up to 14.9% and reduces median inter-token latency by up to 13.2%.