Full Attention Strikes Back: Transferring Full Attention into Sparse within Hundred Training Steps

• URL: http://arxiv.org/abs/2605.16928v1
• Date: Sat, 16 May 2026 10:51:58 GMT
• Title: Full Attention Strikes Back: Transferring Full Attention into Sparse within Hundred Training Steps
• Abstract summary: Long-context inference in large language models is bottlenecked by the quadratic cost of full attention.<n>We show that full-attention LLMs are already intrinsically sparse and can be transformed into highly sparse models with only minimal adaptation.<n>By exploiting the model's intrinsic sparsity, RTPurbo achieves sparsification with only a few hundred training steps.
• Score: 12.944531570933854
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Long-context inference in large language models is bottlenecked by the quadratic cost of full attention. Existing efficient alternatives often rely either on native sparse training or on heuristic token eviction, creating an undesirable trade-off among efficiency, training cost, and accuracy. In this work, we show that full-attention LLMs are already intrinsically sparse and can be transformed into highly sparse models with only minimal adaptation. Our approach is built on three observations: (1) only a small subset of attention heads truly requires full long-context processing; (2) long-range retrieval is governed primarily by a low-dimensional subspace, allowing relevant tokens to be retrieved efficiently with a 16-dimensional indexer; and (3) the useful token budget is strongly query-dependent, making dynamic top-$p$ selection more suitable than fixed top-$k$ sparsification. Based on these insights, we propose RTPurbo, which retains the full KV cache only for retrieval heads and introduces a lightweight token indexer for sparse attention. By exploiting the model's intrinsic sparsity, RTPurbo achieves sparsification with only a few hundred training steps. Experiments on long-context benchmarks and reasoning tasks show that RTPurbo preserves near-lossless accuracy while delivering substantial efficiency gains, including up to a 9.36$\times$ prefill speedup at 1M context and about a 2.01$\times$ decode speedup. These results suggest that strong sparse inference can be obtained from standard full-attention training without expensive native sparse pretraining.

Related papers

• CateKV: On Sequential Consistency for Long-Context LLM Inference Acceleration [67.17709877327232]
  Cate KV is a hybrid KV cache method that retains only critical token information for consistent heads.<n>Cate KV reduces memory usage by up to $2.72times$ and accelerates decoding by $2.18times$ in single-sample inputs.
  arXiv  Detail & Related papers  (2026-08-31T06:02:37Z)
• Scaling Attention via Feature Sparsity [50.64995497733461]
  Scaling Transformers to ultra-long contexts is bottlenecked by the $O(n2 d)$ cost of self-attention.<n>We propose Sparse Feature Attention, where queries and keys are represented as $k$-sparse codes that preserve high-dimensional expressivity.<n>Across GPT-2 and Qwen3 pretraining, SFA matches dense baselines while improving speed by up to $2.5times$ and reducing FLOPs and KV-cache by nearly 50%.
  arXiv  Detail & Related papers  (2026-03-17T08:41:50Z)
• OmniSparse: Training-Aware Fine-Grained Sparse Attention for Long-Video MLLMs [43.78743496579736]
  We introduce OmniSparse, a training-aware fine-grained sparse attention framework for long-video MLLMs.<n>Experiment results show that OmniSparse matches the performance of full attention while achieving up to 2.7x speedup during prefill and 2.4x memory reduction during decoding.
  arXiv  Detail & Related papers  (2025-11-15T13:14:17Z)
• DELTA: Dynamic Layer-Aware Token Attention for Efficient Long-Context Reasoning [6.468843780300177]
  We present textbfDELTA, a training-free sparse attention mechanism that achieves computational efficiency without sacrificing model accuracy.<n>Our results show that selective reuse of intermediate attention maps offers a robust path toward efficient long-context reasoning.
  arXiv  Detail & Related papers  (2025-10-10T21:37:49Z)
• vAttention: Verified Sparse Attention [100.98210818821688]
  vAttention is a practical sparse attention mechanism with user-specified $(epsilon, delta)$ guarantees on approximation accuracy (thus, verified)<n>We show that vAttention significantly improves the quality of sparse attention across datasets.<n>It can be deployed in reasoning scenarios to achieve fast decoding without compromising model quality.
  arXiv  Detail & Related papers  (2025-10-07T08:46:08Z)
• Judge Q: Trainable Queries for Optimized Information Retention in KV Cache Eviction [53.83828564664595]
  Large language models (LLMs) utilize key-value ( KV) cache to store historical information during sequence processing.<n>Current methods for KV cache eviction typically utilize the last window from the pre-filling phase as queries to compute the KV importance scores for eviction.<n>We propose Judge Q, a novel training method which incorporates a soft token list.
  arXiv  Detail & Related papers  (2025-09-13T03:34:12Z)
• Delta Attention: Fast and Accurate Sparse Attention Inference by Delta Correction [52.14200610448542]
  A transformer has a quadratic complexity, leading to high inference costs and latency for long sequences.<n>We propose a simple, novel, and effective procedure for correcting this distributional shift.<n>Our method can maintain approximately 98.5% sparsity over full quadratic attention, making our model 32 times faster than Flash Attention 2 when processing 1M token prefills.
  arXiv  Detail & Related papers  (2025-05-16T13:48:33Z)
• Sparsity Forcing: Reinforcing Token Sparsity of MLLMs [40.93786579652003]
  We explicitly reinforce token sparsity in well-posed multimodal large language models (MLLMs) through a simple RL-based post-training framework named textitSparsity Forcing.<n>Our method explores the efficiency-accuracy trade-off by running multiple rollouts with different token budgets, where both efficiency (token reduction ratio) and performance (answer correctness) are formulated as joint rewards.
  arXiv  Detail & Related papers  (2025-04-23T01:45:55Z)
• Twilight: Adaptive Attention Sparsity with Hierarchical Top-$p$ Pruning [36.25003932215851]
  Twilight is a framework to bring adaptive sparsity to any existing sparse attention algorithm.<n>We show that Twilight can adaptively prune at most 98% of redundant tokens, leading to $15.4times$ acceleration in self-attention operations and $3.9times$ acceleration in end-to-end per token latency.
  arXiv  Detail & Related papers  (2025-02-04T23:26:10Z)
• Squeezed Attention: Accelerating Long Context Length LLM Inference [61.787865959140994]
  We propose Squeezed Attention to accelerate applications where a large portion of the input context is fixed.<n>During inference, we compare query tokens from the user input with the centroids to predict which keys from the fixed context are semantically relevant.<n>We also present a hierarchical version of our algorithm which can reduce the complexity of attention from linear to logarithmic with respect to the fixed context length.
  arXiv  Detail & Related papers  (2024-11-14T18:54:19Z)
• Post-Training Sparse Attention with Double Sparsity [44.772593893621085]
  "Double Sparsity" is a novel post-training sparse attention technique designed to alleviate this bottleneck by reducing KV cache access. Double Sparsity combines token sparsity, which focuses on utilizing only the important tokens for computing self-attention, with channel sparsity, an approach that uses important feature channels for identifying important tokens. With offloading, it achieves a decoding speed acceleration of 16.3$times$ compared to state-of-the-art solutions at a sequence length of 256K.
  arXiv  Detail & Related papers  (2024-08-11T18:40:36Z)
• S2-Attention: Hardware-Aware Context Sharding Among Attention Heads [49.1454481007861]
  Sparse attention selectively attends to a subset of tokens in the context.<n>It remains unclear whether sparse attention can maintain the model's quality at a scale of today's large language models.<n>This paper presents Sparsely-Sharded(S2) Attention, a Triton library that provides kernel optimization for sparse attention customizable at both per-head and per-context-range levels.
  arXiv  Detail & Related papers  (2024-07-25T00:27:07Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.