SPIN (Shadow Predictive Indexer) is a new method for sparse attention that reduces the overhead of scoring the entire key-value (KV) cache during decoding. By using lightweight, history-based predictions to identify important KV blocks, SPIN avoids full cache scoring at each step and treats KV blocks and speculative decoding as core design elements. Experiments on long-context and agentic benchmarks show 30‑40% sparsity with maintained task quality, and in vLLM serving it boosts throughput by up to 14.9% while cutting median inter‑token latency by up to 13.2%.
arXiv:2602. 07223v2 Announce Type: replace Abstract: Long-context large language model (LLM) inference has become the norm for today's AI applications.
arXiv:2606. 31519v1 Announce Type: new Abstract: Long-context Large Language Model inference is severely bottlenecked by the massive Key-Value (KV) cache, yet existing sparse attention methods often suffer from static fixed-budget (Top-k) retrieval or rely on proxy scores that are computationally expensive and biased.
arXiv:2606. 06467v1 Announce Type: cross Abstract: Long-context inference in modern LLMs is increasingly constrained by decoding efficiency, especially in reasoning-heavy settings where models generate long intermediate chains of thought.
arXiv:2609.26368v1 Announce Type: new Abstract: Long-horizon and multi-turn agents typically generate short actions and process long observations from tools and environments. This growing context dem...
arXiv:2610.06927v1 Announce Type: cross Abstract: The key-value (KV) cache of autoregressive transformers grows linearly with context length and dominates memory at long context. Most training-free r...
arXiv:2610.20392v1 Announce Type: cross Abstract: Speculative decoding accelerates LLM generation through lightweight drafting and parallel verification. It is particularly promising for long inputs,...
arXiv:2606. 04511v1 Announce Type: cross Abstract: Sparse attention reduces compute and memory bandwidth for long-context LLM inference.
arXiv:2608. 19920v1 Announce Type: new Abstract: A lot of prior work addressed key-value (KV) cache selection and compression by sparse attention to enable long-context inference for transformer language models without excessive hardware budgets.
KVBuffer is an IO‑aware serving mechanism designed to improve the efficiency of linear attention models. By buffering recent keys and values, it enables chunkwise decoding, parallel speculative decoding, and direct attention computation for short contexts, thereby reducing memory access and latency. Implemented in SGLang for Qwen3.5 and Nemotron‑3, KVBuffer achieves up to 1.47× faster batch decoding and a five‑fold increase in speculative decoding throughput when verifying four draft tokens.
The paper introduces QUILT, a new sparse‑attention execution mechanism that jointly processes neighboring queries to reuse shared key‑value (KV) entries, thereby reducing redundant memory traffic and computation. QUILT employs Shift‑and‑Compare Set Decomposition (SCSD) to convert irregular set operations into regular data‑parallel primitives and pipelines this with attention computation to hide overhead. The approach includes cascaded sharing across multiple granularities and a tile‑aware strategy that balances sharing with hardware utilization, achieving up to 55.1% lower kernel latency and 36.8% faster time‑to‑first‑token on LongBench benchmarks with negligible accuracy loss.
The paper introduces On‑Demand Attention (ODA), a local‑first decoding strategy that predicts when a pretrained language model would benefit from global attention. By training only a lightweight recall head, ODA selectively triggers global attention during generation, keeping pretrained weights unchanged and preserving the full key‑value cache for future recall. Experiments on Qwen, Gemma, and hybrid‑attention models show that ODA largely recovers performance lost with local attention while significantly cutting global reads, enabling faster long‑context inference.
arXiv:2606. 09079v1 Announce Type: cross Abstract: Conventional LLMs keep the full KV cache loaded during decoding, causing a severe GPU memory bottleneck for ultra-long context serving.