On-the-Fly Adaptive Distillation of Transformer to Dual-State Linear Attention

• URL: http://arxiv.org/abs/2506.09316v3
• Date: Tue, 17 Jun 2025 04:56:46 GMT
• Title: On-the-Fly Adaptive Distillation of Transformer to Dual-State Linear Attention
• Authors: Yeonju Ro, Zhenyu Zhang, Souvik Kundu, Zhangyang Wang, Aditya Akella,
• Abstract summary: Large language models (LLMs) excel at capturing global token dependencies via self-attention but face prohibitive compute and memory costs on lengthy inputs.<n>We first propose dual-state linear attention (A), a novel design that maintains two hidden states-one for preserving historical context and one for tracking recencythereby mitigating the short-range bias typical of linear-attention architectures.<n>We introduce DSLA-Serve, an online adaptive distillation framework that progressively replaces Transformer layers DSLA layers at inference time, guided by a sensitivity-based layer ordering.
• Score: 53.22963042513293
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Large language models (LLMs) excel at capturing global token dependencies via self-attention but face prohibitive compute and memory costs on lengthy inputs. While sub-quadratic methods (e.g., linear attention) can reduce these costs, they often degrade accuracy due to overemphasizing recent tokens. In this work, we first propose dual-state linear attention (DSLA), a novel design that maintains two specialized hidden states-one for preserving historical context and one for tracking recency-thereby mitigating the short-range bias typical of linear-attention architectures. To further balance efficiency and accuracy under dynamic workload conditions, we introduce DSLA-Serve, an online adaptive distillation framework that progressively replaces Transformer layers with DSLA layers at inference time, guided by a sensitivity-based layer ordering. DSLA-Serve uses a chained fine-tuning strategy to ensure that each newly converted DSLA layer remains consistent with previously replaced layers, preserving the overall quality. Extensive evaluations on commonsense reasoning, long-context QA, and text summarization demonstrate that DSLA-Serve yields 2.3x faster inference than Llama2-7B and 3.0x faster than the hybrid Zamba-7B, while retaining comparable performance across downstream tasks. Our ablation studies show that DSLA's dual states capture both global and local dependencies, addressing the historical-token underrepresentation seen in prior linear attentions. Codes are available at https://github.com/utnslab/DSLA-Serve.

Related papers

• Beyond Scattered Acceptance: Fast and Coherent Inference for DLMs via Longest Stable Prefixes [10.877713536966601]
  Longestahead Prefix (LSP) scheduler is a training-free and model-agnostic inference paradigm based on monolithic prefix absorption.<n>LSP evaluates token stability via a single forward pass, dynamically identifies a contiguous left-aligned block of stable predictions.<n>It snaps its boundary to natural linguistic or structural acceptances before an atomic commitment.
  arXiv  Detail & Related papers  (2026-03-05T18:25:26Z)
• Rethinking Multi-Condition DiTs: Eliminating Redundant Attention via Position-Alignment and Keyword-Scoping [61.459927600301654]
  Multi-condition control is bottlenecked by the conventional concatenate-and-attend'' strategy.<n>Our analysis reveals that much of this cross-modal interaction is spatially or semantically redundant.<n>We propose Position-aligned and Keyword-scoped Attention (PKA), a highly efficient framework designed to eliminate these redundancies.
  arXiv  Detail & Related papers  (2026-02-06T16:39:10Z)
• VidLaDA: Bidirectional Diffusion Large Language Models for Efficient Video Understanding [52.69880888587866]
  Current Video Large Language Models (Video LLMs) typically encode frames via a encoder vision and employ an autoregressive (AR) LLM for understanding and generation.<n>We propose VidLaDA, a Diffusion Video LLM based on Language Models (DLMs) that leverages bidirectional attention to unlock comprehensive modeling and decode tokens in parallel.<n>Experiments show VidLaDA rivals state-of-the-art AR baselines and outperforms DLM baselines, with MARS-Cache delivering over 12x speedup without compromising accuracy.
  arXiv  Detail & Related papers  (2026-01-25T15:02:01Z)
• Training-free Context-adaptive Attention for Efficient Long Context Modeling [57.703159205740185]
  Training-free Context-adaptive Attention (TCA-Attention) is a training-free sparse attention mechanism that selectively attends to only the informative tokens for efficient long-context inference.<n>TCA-Attention achieves a 2.8$times$ speedup and reduces KV cache by 61% at 128K context length while maintaining performance comparable to full attention.
  arXiv  Detail & Related papers  (2025-12-10T01:54:57Z)
• LMAR: Language Model Augmented Retriever for Domain-specific Knowledge Indexing [42.51773265892766]
  Retrieval Augmented Generation (RAG) systems often struggle with domain-specific knowledge due to performance deterioration of pre-trained embeddings.<n>We propose LMAR (Language Model Augmented Retriever), a model-agnostic framework that addresses these challenges by combining LLM-guided data synthesis with contrastive embedding adaptation and efficient text clustering.
  arXiv  Detail & Related papers  (2025-08-04T16:59:43Z)
• Beyond Fixed: Variable-Length Denoising for Diffusion Large Language Models [74.15250326312179]
  Diffusion Large Language Models offer efficient parallel generation and capable global modeling.<n>The dominant application ofDLLMs is hindered by the need for a statically predefined generation length.<n>We introduce DAEDAL, a novel training-free denoising strategy that enables Dynamic Adaptive Length Expansion.
  arXiv  Detail & Related papers  (2025-08-01T17:56:07Z)
• Mitigating Hallucination of Large Vision-Language Models via Dynamic Logits Calibration [8.192590936983347]
  Large Vision-Language Models (LVLMs) have demonstrated significant advancements in multimodal understanding.<n>They are frequently hampered by hallucination-the generation of text that contradicts visual input.<n>Existing training-free decoding strategies exhibit critical limitations.<n>This paper introduces Dynamic Logits (DLC), a novel training-free decoding framework designed to align text generation with visual evidence at inference time.
  arXiv  Detail & Related papers  (2025-06-26T17:35:40Z)
• SP-VLA: A Joint Model Scheduling and Token Pruning Approach for VLA Model Acceleration [69.54069477520534]
  Vision-Language-Action (VLA) models have attracted increasing attention for their strong control capabilities.<n>Their high computational cost and low execution frequency hinder their suitability for real-time tasks such as robotic manipulation and autonomous navigation.<n>We propose SP-VLA, a unified framework that accelerates VLA models by jointly scheduling models and pruning tokens.
  arXiv  Detail & Related papers  (2025-06-15T05:04:17Z)
• Constrained Entropic Unlearning: A Primal-Dual Framework for Large Language Models [7.566515311806724]
  Large Language Models (LLMs) deployed in real-world settings increasingly face the need to unlearn sensitive, outdated, or proprietary information.<n>Existing unlearning methods formulate forgetting and retention as a regularized trade-off, combining both objectives into a single scalarized loss.<n>We propose a new formulation of LLM unlearning as a constrained optimization problem: forgetting is enforced via a novel logit-margin flattening loss.
  arXiv  Detail & Related papers  (2025-06-05T17:55:23Z)
• Rectified Sparse Attention [61.7702154360081]
  Efficient long-sequence generation is a critical challenge for Large Language Models.<n>We propose Rectified Sparse Attention (ReSA), a simple yet effective method that combines block-sparse attention with periodic dense rectification.<n> Experiments across math reasoning, language modeling, and retrieval tasks demonstrate that ReSA achieves near-lossless generation quality.
  arXiv  Detail & Related papers  (2025-06-04T16:01:48Z)
• Fast-dLLM: Training-free Acceleration of Diffusion LLM by Enabling KV Cache and Parallel Decoding [51.711605076319216]
  Diffusion-based large language models (Diffusion LLMs) have shown promise for non-autoregressive text generation with parallel decoding capabilities.<n>We introduce a novel block-wise approximate KV Cache mechanism tailored for bidirectional diffusion models, enabling cache reuse with negligible performance drop.<n>We propose a confidence-aware parallel decoding strategy that selectively decodes tokens exceeding a confidence threshold, mitigating dependency violations and maintaining generation quality.
  arXiv  Detail & Related papers  (2025-05-28T17:39:15Z)
• LONGER: Scaling Up Long Sequence Modeling in Industrial Recommenders [23.70714095931094]
  Long-sequence optimized traNsformer for GPU-Efficient Recommenders.<n>Longer consistently outperforms strong baselines in offline metrics and online A/B testing.
  arXiv  Detail & Related papers  (2025-05-07T13:54:26Z)
• Efficient Pretraining Length Scaling [21.4715211093876]
  We present the Parallel Hidden Decoding Transformer (textitPHD-Transformer), a novel framework that enables efficient length scaling during pre-training.<n>textitPHD-Transformer achieves this through an innovative KV cache management strategy that distinguishes between original tokens and hidden decoding tokens.
  arXiv  Detail & Related papers  (2025-04-21T09:41:26Z)
• Rethinking Addressing in Language Models via Contexualized Equivariant Positional Encoding [89.52931576290976]
  Transformers rely on both content-based and position-based addressing mechanisms to make predictions.<n>TAPE is a novel framework that enhances positional embeddings by incorporating sequence content across layers.<n>Our method can be easily integrated into pre-trained transformers, offering parameter-efficient fine-tuning with minimal overhead.
  arXiv  Detail & Related papers  (2025-01-01T03:23:00Z)
• FiRST: Finetuning Router-Selective Transformers for Input-Adaptive Latency Reduction [11.146015814220858]
  We propose FiRST, an algorithm that reduces latency by using layer-specific routers to select a subset of transformer layers adaptively for each input sequence.<n>FiRST preserves compatibility with KV caching enabling faster inference while being quality-aware.<n>Our approach reveals that input adaptivity is critical - indeed, different task-specific middle layers play a crucial role in evolving hidden representations depending on tasks.
  arXiv  Detail & Related papers  (2024-10-16T12:45:35Z)
• SHERL: Synthesizing High Accuracy and Efficient Memory for Resource-Limited Transfer Learning [63.93193829913252]
  We propose an innovative METL strategy called SHERL for resource-limited scenarios. In the early route, intermediate outputs are consolidated via an anti-redundancy operation. In the late route, utilizing minimal late pre-trained layers could alleviate the peak demand on memory overhead.
  arXiv  Detail & Related papers  (2024-07-10T10:22:35Z)
• Short-Long Convolutions Help Hardware-Efficient Linear Attention to Focus on Long Sequences [60.489682735061415]
  We propose CHELA, which replaces state space models with short-long convolutions and implements linear attention in a divide-and-conquer manner. Our experiments on the Long Range Arena benchmark and language modeling tasks demonstrate the effectiveness of the proposed method.
  arXiv  Detail & Related papers  (2024-06-12T12:12:38Z)
• Efficient Long Sequence Modeling via State Space Augmented Transformer [92.74707853711374]
  We propose SPADE, short for $underlinetextbfS$tate sunderlinetextbfP$ace. We augment a SSM into the bottom layer of SPADE, and we employ efficient local attention methods for the other layers. Experimental results on the Long Range Arena benchmark and language modeling tasks demonstrate the effectiveness of the proposed method.
  arXiv  Detail & Related papers  (2022-12-15T20:51:27Z)

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.