The paper introduces INTRA, a hardware‑aware static sparse attention mechanism that interleaves non‑contiguous token groups to enable efficient long‑range communication while reducing quadratic cost. INTRA uses Scatter and Gather layers to create short mediated paths across layers, and a FlashAttention‑style kernel computes structured indices at runtime for dense packing of selected tokens. Experiments on an A100 GPU show significant speedups—up to 8.4× over MoBA and 10.4× over density‑constrained NSA—while maintaining or improving image quality and long‑context performance on models like FLUX.1‑dev and LLaMA‑3.1‑8B.
Sparse attention reduces the KV-cache traffic that dominates long-context LLM decoding, but practical methods largely operate at coarse granularity. Token-level selection offers finer control, yet pre...
Joint Subspace Sparse Attention (JSSA) is a new method that reduces KV‑cache traffic in long‑context LLM decoding by selecting tokens at the token level while preserving query–key structure. It learns a shared low‑dimensional query–key subspace through a tractable spectral solution and implements a gather‑free indexed‑attention kernel that supports per‑KV‑head selections. Across various attention architectures, JSSA maintains near full‑KV accuracy, avoids generation‑length inflation, and can achieve up to 6.43× decoding throughput on a single GPU for 128K context, even at 700B‑parameter scale.
arXiv:2606. 13392v1 Announce Type: new Abstract: Ultra-long-context capability is becoming indispensable for frontier LLMs: agentic workflows, repository-scale code reasoning, and persistent memory all require the model to jointly attend over hundreds of thousands to millions of tokens, yet the quadratic cost of softmax attention makes this untenable at deployment scale.
arXiv:2606. 04511v1 Announce Type: cross Abstract: Sparse attention reduces compute and memory bandwidth for long-context LLM inference.
arXiv:2508. 18224v3 Announce Type: replace-cross Abstract: Recent advances in sparse attention mechanisms have demonstrated strong potential for reducing the computational cost of long-context training and inference in large language models (LLMs).
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.
SPARTA is a sparse ternary LLM accelerator designed for long‑context inference on edge devices. It combines a phase‑aware sparse dataflow, a unified compute fabric that maps ternary SpMM/SpMV and bit‑serial attention onto shared processing elements, and an adaptive two‑stage attention engine that reduces KV‑cache traffic. The design achieves significant reductions in KV traffic (4.1‑4.8× over full attention) and outperforms NVIDIA Jetson AGX Orin and AMD Ryzen 7 9700X in generation throughput and energy efficiency, while also surpassing other accelerators in throughput, energy, and area efficiency.
DeltaLLM is a training‑free framework that accelerates on‑device large language model inference by introducing delta‑key sparse attention and a vectorized structured sparse online softmax. It reuses delta keys across query heads, preserves exact attention near the diagonal while approximating non‑local regions, and exploits hardware‑friendly sparsity with a specialized vector‑processor architecture. Experiments on BitNet‑b1.58‑2B‑4T and LLaMA3.2‑1B‑Instruct show up to 60% delta‑key sparsity, 63% structured attention‑score sparsity, and a 1.62× speedup on the KV260 FPGA with minimal accuracy loss.
BF1 is a deterministic block‑aligned dyadic sparse‑attention retrofit designed to reduce the cost of causal attention in long‑context transformers. It combines a small exact local neighborhood, a global first block, and logarithmically spaced historical blocks, achieving O(n log n) token interactions per layer with O(log n) communication depth. On an NVIDIA RTX PRO 6000 Blackwell GPU, BF1 outperforms dense attention for 2K–4K tokens and delivers up to a 10.91× prefill speedup at 32K tokens, while retrofitting eight of 28 Qwen3‑0.6B layers reduces first‑token latency by up to 15.3% at 32K tokens and yields the lowest perplexity among compared sparse and dense training protocols.
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.
iCATS introduces a training‑free sparse attention mechanism for Diffusion Transformers that accelerates video generation by reducing computations. It improves importance estimation by clustering query‑key dot‑product interactions and reformulates the objective as a quadratic form for low‑cost computation. The method also adapts sparsity dynamically across denoising timesteps using an SNR‑guided schedule and employs a tail‑merging strategy to enhance GPU kernel utilization, achieving significant speedups while maintaining high PSNR scores.
arXiv:2608. 01662v1 Announce Type: cross Abstract: DeepSeek Sparse Attention (DSA) enables efficient long-context modeling through its Lightning Indexer.