Power Law Guided Dynamic Sifting for Efficient Attention

• URL: http://arxiv.org/abs/2506.05300v1
• Date: Thu, 05 Jun 2025 17:50:32 GMT
• Title: Power Law Guided Dynamic Sifting for Efficient Attention
• Authors: Nirav Koley, Prajwal Singhania, Abhinav Bhatele,
• Abstract summary: We propose SiftAttention, a novel approximate attention method that replaces the top-k$ step with a computationally efficient element-wise filtering operation based on a threshold value.<n>Our evaluation demonstrates that SiftAttention preserves model quality better than existing approximate attention methods while reducing memory usage when loading vectors.
• Score: 0.8198759882422455
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Efficient inference on GPUs using large language models remains challenging due to memory bandwidth limitations, particularly during data transfers between High Bandwidth Memory (HBM) and SRAM in attention computations. Approximate attention methods address this issue by reducing computational and memory overhead but often rely on expensive top-$k$ operations, which perform poorly on GPUs. We propose SiftAttention, a novel approximate attention method that replaces the top-$k$ step with a computationally efficient element-wise filtering operation based on a threshold value. Our intuition for doing this is based on our empirical observation that the $\tau$-th quantile of attention scores follows a predictable power-law over sequential generation steps. Exploiting this insight, our approach dynamically estimates a threshold value per prompt at each generation step. Only attention scores above this threshold and their corresponding value vectors are loaded/used to compute the attention output, reducing data movement between HBM and SRAM. Our evaluation demonstrates that SiftAttention preserves model quality better than existing approximate attention methods while reducing memory bandwidth usage when loading value vectors.

Related papers

• RetroInfer: A Vector-Storage Approach for Scalable Long-Context LLM Inference [27.69137902678418]
  RetroInfer is a novel system that exploits the inherent attention sparsity to accelerate long-context inference.<n>We show up to 4.5X speedup over full attention within GPU memory limits and up to 10.5X over sparse attention baselines when KV cache is extended to CPU memory.
  arXiv  Detail & Related papers  (2025-05-05T18:01:17Z)
• FlashAttention on a Napkin: A Diagrammatic Approach to Deep Learning IO-Awareness [0.0]
  Methods like FlashAttention have achieved a x6 performance improvement over native PyTorch by avoiding unnecessary data transfers.<n>This paper extends Neural Circuit Diagrams for deep learning models to consider resource usage and the distribution of tasks across a GPU hierarchy.<n>We develop a methodology for representing intermediate-level pseudocode with diagrams, allowing hardware-aware algorithms to be derived step-by-step.
  arXiv  Detail & Related papers  (2024-12-04T13:52:04Z)
• Continual Low-Rank Scaled Dot-product Attention [67.11704350478475]
  We introduce a new formulation of the Scaled-product Attention based on the Nystr"om approximation that is suitable for Continual Inference.<n>In experiments on Online Audio Classification and Online Action Detection tasks, the proposed Continual Scaled Dot-product Attention can lower the number of operations by up to three orders of magnitude.
  arXiv  Detail & Related papers  (2024-12-04T11:05:01Z)
• MAS-Attention: Memory-Aware Stream Processing for Attention Acceleration on Resource-Constrained Edge Devices [24.1144641404561]
  We propose a scheme for exact attention inference acceleration on memory-constrained edge accelerators.<n>We show up to 2.75x speedup and 54% reduction in energy consumption as compared to the state-of-the-art attention fusion method (FLAT) in the edge computing scenario.
  arXiv  Detail & Related papers  (2024-11-20T19:44:26Z)
• Hybrid Dynamic Pruning: A Pathway to Efficient Transformer Inference [1.0919012968294923]
  We introduce a novel algorithm-architecture co-design approach that accelerates transformers using head sparsity, block sparsity and approximation opportunities to reduce computations in attention and reduce memory access. With the observation of the huge redundancy in attention scores and attention heads, we propose a novel integer-based row-balanced block pruning to prune unimportant blocks in the attention matrix at run time. Also propose integer-based head pruning to detect and prune unimportant heads at an early stage at run time.
  arXiv  Detail & Related papers  (2024-07-17T11:15:16Z)
• Lean Attention: Hardware-Aware Scalable Attention Mechanism for the Decode-Phase of Transformers [4.674454841332859]
  Transformer-based models have emerged as one of the most widely used architectures for natural language processing.<n>These huge models are memory hungry and incur significant inference latency even on cutting edge AI-accelerators.<n>We propose LeanAttention, a scalable technique of computing self-attention for the token-generation phase.
  arXiv  Detail & Related papers  (2024-05-17T00:52:39Z)
• CORM: Cache Optimization with Recent Message for Large Language Model Inference [57.109354287786154]
  We introduce an innovative method for optimizing the KV cache, which considerably minimizes its memory footprint. CORM, a KV cache eviction policy, dynamically retains essential key-value pairs for inference without the need for model fine-tuning. Our validation shows that CORM reduces the inference memory usage of KV cache by up to 70% with negligible performance degradation across six tasks in LongBench.
  arXiv  Detail & Related papers  (2024-04-24T16:11:54Z)
• Simple linear attention language models balance the recall-throughput tradeoff [60.06020449520365]
  We propose BASED, a simple architecture combining linear and sliding window attention.<n>We train language models up to 1.3b parameters and show that BASED matches the strongest sub-quadratic models in perplexity and outperforms them on real-world recall-intensive tasks by 6.22 accuracy points.
  arXiv  Detail & Related papers  (2024-02-28T19:28:27Z)
• Constant Memory Attention Block [74.38724530521277]
  Constant Memory Attention Block (CMAB) is a novel general-purpose attention block that computes its output in constant memory and performs updates in constant computation. We show our proposed methods achieve results competitive with state-of-the-art while being significantly more memory efficient.
  arXiv  Detail & Related papers  (2023-06-21T22:41:58Z)
• Skip-Attention: Improving Vision Transformers by Paying Less Attention [55.47058516775423]
  Vision computation transformers (ViTs) use expensive self-attention operations in every layer. We propose SkipAt, a method to reuse self-attention from preceding layers to approximate attention at one or more subsequent layers. We show the effectiveness of our method in image classification and self-supervised learning on ImageNet-1K, semantic segmentation on ADE20K, image denoising on SIDD, and video denoising on DAVIS.
  arXiv  Detail & Related papers  (2023-01-05T18:59:52Z)
• Sparse Attention Acceleration with Synergistic In-Memory Pruning and On-Chip Recomputation [6.303594714446706]
  Self-attention mechanism gauges pairwise correlations across entire input sequence. Despite favorable performance, calculating pairwise correlations is prohibitively costly. This work addresses these constraints by architecting an accelerator, called SPRINT, which computes attention scores in an approximate manner.
  arXiv  Detail & Related papers  (2022-09-01T17:18:19Z)
• Mesa: A Memory-saving Training Framework for Transformers [58.78933015299703]
  We present Mesa, a memory-saving training framework for Transformers. Mesa uses exact activations during forward pass while storing a low-precision version of activations to reduce memory consumption during training. Experiments on ImageNet, CIFAR-100 and ADE20K demonstrate that Mesa can reduce half of the memory footprints during training.
  arXiv  Detail & Related papers  (2021-11-22T11:23:01Z)
• Fast Graph Attention Networks Using Effective Resistance Based Graph Sparsification [70.50751397870972]
  FastGAT is a method to make attention based GNNs lightweight by using spectral sparsification to generate an optimal pruning of the input graph. We experimentally evaluate FastGAT on several large real world graph datasets for node classification tasks.
  arXiv  Detail & Related papers  (2020-06-15T22:07:54Z)

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.