2026-04-10
KV Cache Offloading for Context-Intensive Tasks
针对长上下文任务KV缓存瓶颈，提出KV缓存卸载方法，降低内存与延迟同时保持精度
Authors: Andrey Bocharnikov, Ivan Ermakov, Denis Kuznedelev, Vyacheslav Zhdanovskiy, Yegor Yershov | Date: 2026-04-09
Link: http://arxiv.org/abs/2604.08426v1
Abstract
With the growing demand for long-context LLMs across a wide range of applications, the key-value (KV) cache has become a critical bottleneck for both latency and memory usage. Recently, KV-cache offloading has emerged as a promising approach to reduce memory footprint and inference latency while preserving accuracy. Prior evaluations have largely focused on tasks that do not require extracting large amounts of information from the context. In this work, we study KV-cache offloading on context-intensive tasks: problems where the solution requires looking up a lot of information from the input prompt. We create and release the Text2JSON benchmark, a highly context-intensive task that requires extracting structured knowledge from raw text. We evaluate modern KV offloading on Text2JSON and other context-intensive tasks and find significant performance degradation on both Llama 3 and Qwen 3 models. Our analysis identifies two key reasons for poor accuracy: low-rank projection of keys and unreliable landmarks, and proposes a simpler alternative strategy that significantly improves accuracy across multiple LLM families and benchmarks. These findings highlight the need for a comprehensive and rigorous evaluation of long-context compression techniques.DMax: Aggressive Parallel Decoding for dLLMs
解决扩散语言模型并行解码误差累积问题，提出DMax方法提升解码并行度与生成质量
Authors: Zigeng Chen, Gongfan Fang, Xinyin Ma, Ruonan Yu, Xinchao Wang | Date: 2026-04-09
Link: http://arxiv.org/abs/2604.08302v1
Abstract
We present DMax, a new paradigm for efficient diffusion language models (dLLMs). It mitigates error accumulation in parallel decoding, enabling aggressive decoding parallelism while preserving generation quality. Unlike conventional masked dLLMs that decode through a binary mask-to-token transition, DMax reformulates decoding as a progressive self-refinement from mask embeddings to token embeddings. At the core of our approach is On-Policy Uniform Training, a novel training strategy that efficiently unifies masked and uniform dLLMs, equipping the model to recover clean tokens from both masked inputs and its own erroneous predictions. Building on this foundation, we further propose Soft Parallel Decoding. We represent each intermediate decoding state as an interpolation between the predicted token embedding and the mask embedding, enabling iterative self-revising in embedding space. Extensive experiments across a variety of benchmarks demonstrate the effectiveness of DMax. Compared with the original LLaDA-2.0-mini, our method improves TPF on GSM8K from 2.04 to 5.47 while preserving accuracy. On MBPP, it increases TPF from 2.71 to 5.86 while maintaining comparable performance. On two H200 GPUs, our model achieves an average of 1,338 TPS at batch size 1. Code is available at: https://github.com/czg1225/DMaxAlloc-MoE: Budget-Aware Expert Activation Allocation for Efficient Mixture-of-Experts Inference
针对MoE推理延迟瓶颈，提出预算感知专家激活分配方法，优化资源受限部署性能
Authors: Baihui Liu, Kaiyuan Tian, Wei Wang, Zhaoning Zhang, Linbo Qiao, Dongsheng Li | Date: 2026-04-09
Link: http://arxiv.org/abs/2604.08133v1
Abstract
Mixture-of-Experts (MoE) has become a dominant architecture for scaling large language models due to their sparse activation mechanism. However, the substantial number of expert activations creates a critical latency bottleneck during inference, especially in resource-constrained deployment scenarios. Existing approaches that reduce expert activations potentially lead to severe model performance degradation. In this work, we introduce the concept of \emph{activation budget} as a constraint on the number of expert activations and propose Alloc-MoE, a unified framework that optimizes budget allocation coordinately at both the layer and token levels to minimize performance degradation. At the layer level, we introduce Alloc-L, which leverages sensitivity profiling and dynamic programming to determine the optimal allocation of expert activations across layers. At the token level, we propose Alloc-T, which dynamically redistributes activations based on routing scores, optimizing budget allocation without increasing latency. Extensive experiments across multiple MoE models demonstrate that Alloc-MoE maintains model performance under a constrained activation budget. Especially, Alloc-MoE achieves $1.15\times$ prefill and $1.34\times$ decode speedups on DeepSeek-V2-Lite at half of the original budget.Initialisation Determines the Basin: Efficient Codebook Optimisation for Extreme LLM Quantization
解决极端量化中代码book初始化瓶颈，提出高效优化方法提升2位量化稳定性
Authors: Ian W. Kennedy, Nafise Sadat Moosavi | Date: 2026-04-09
Link: http://arxiv.org/abs/2604.08118v1
Abstract
Additive quantization enables extreme LLM compression with O(1) lookup-table dequantization, making it attractive for edge deployment. Yet at 2-bit precision, it often fails catastrophically, even with extensive search and finetuning. We show that the dominant bottleneck is codebook initialisation. Greedy sequential initialisation frequently places the model in poor optimisation regions that subsequent beam search and PV-tuning struggle to overcome. We analyse this behaviour through the representational ratio \r{ho} = N/KM, which characterises the relationship between weight groups and codebook capacity, and propose OA-EM, an output-aware EM initialisation method using Hessian-weighted Mahalanobis distance. Across compression rates, search budgets, and three architectures (Llama 3.2 3B, Llama 3.1 8B, Qwen 2.5 3B), OA-EM consistently produces better solutions after PV-tuning and dominates the quality-compute frontier. The severity of the bottleneck scales with \r{ho}: moderate at 3 bpp but extreme at 2 bpp, where poor initialisation can degrade perplexity by orders of magnitude. More broadly, our results highlight the importance of optimisation geometry in compressed model spaces, where initialisation can dominate subsequent search and fine-tuning.Dual-Pool Token-Budget Routing for Cost-Efficient and Reliable LLM Serving
针对vLLM配置与流量不匹配问题，提出双池令牌预算路由方法，提升成本效率与可靠性
Authors: Xunzhuo Liu, Bowei He, Xue Liu, Andy Luo, Haichen Zhang, Huamin Chen | Date: 2026-04-09
Link: http://arxiv.org/abs/2604.08075v1
Abstract
Production vLLM fleets typically provision each instance for the worst-case context length, leading to substantial KV-cache over-allocation and under-utilized concurrency. In practice, 80-95% of requests are short, yet are served under configurations optimized for long contexts, wasting 4-8$\times$ throughput capacity and triggering reliability issues such as OOM crashes, preemption, and request rejections. We identify a common root cause for these inefficiencies: configuration-traffic mismatch. We propose dual-pool token-budget routing, a lightweight dispatch mechanism that partitions a homogeneous fleet into two specialized pools: a high-throughput short-context pool and a high-capacity long-context pool. Each request is routed based on its estimated total token budget, computed using a per-category bytes-to-token ratio that is learned online via exponential moving average from usage.prompt_tokens feedback, eliminating the need for a tokenizer. We also develop a simple analytical model that predicts fleet-level cost savings from workload characteristics and measured throughput differences, enabling practitioners to estimate benefits prior to deployment. Evaluations on real-world traces from the Azure LLM Inference Dataset and LMSYS-Chat-1M, serving Llama-3-70B on A100 GPUs, show that our approach reduces GPU-hours by 31-42%, corresponding to \$2.86M annual savings at fleet scale, while lowering preemption rates by 5.4$\times$ and improving P99 TTFT by 6%. A case study with Qwen3-235B-A22B on AMD MI300X at 10,000 req/s projects \$15.4M in annual savings. The method incurs only O(1) dispatch overhead, adapts automatically to heterogeneous workloads, and composes seamlessly with existing optimizations such as PagedAttention, continuous batching, and prefill-decode disaggregation.A Full-Stack Performance Evaluation Infrastructure for 3D-DRAM-based LLM Accelerators
针对3D-DRAM加速器评估工具封闭问题，提出全栈性能评估基础设施，提升分析可及性
Authors: Cong Li, Chenhao Xue, Yi Ren, Xiping Dong, Yu Cheng, Yinbo Hu, Fujun Bai, Yixin Guo, Xiping Jiang, Qiang Wu, Zhi Yang, Zhe Cheng, Yuan Xie, Guangyu Sun | Date: 2026-04-09
Link: http://arxiv.org/abs/2604.08044v1
Abstract
Large language models (LLMs) exhibit memory-intensive behavior during decoding, making it a key bottleneck in LLM inference. To accelerate decoding execution, hybrid-bonding-based 3D-DRAM has been adopted in LLM accelerators. While this emerging technology provides strong performance gains over existing hardware, current 3D-DRAM accelerators (3D-Accelerators) rely on closed-source evaluation tools, limiting access to publicly available performance analysis methods. Moreover, existing designs are highly customized for specific scenarios, lacking a general and reusable full-stack modeling for 3D-Accelerators across diverse usecases. To bridge this fundamental gap, we present ATLAS, the first silicon-proven Architectural Three-dimesional-DRAM-based LLM Accelerator Simulation framework. Built on commercially deployed multi-layer 3D-DRAM technology, ATLAS introduces unified abstractions for both 3D-Accelerator system architecture and programming primitives to support arbitrary LLM inference scenarios. Validation against real silicon shows that ATLAS achieves $\le$8.57% simulation error and 97.26-99.96\% correlation with measured performance. Through design space exploration with ATLAS, we demonstrate its ability to guide architecture design and distill key takeaways for both 3D-DRAM memory system and 3D-Accelerator microarchitecture across scenarios. ATLAS will be open-sourced upon publication, enabling further research on 3D-Accelerators.Rethinking Residual Errors in Compensation-based LLM Quantization
针对补偿量化中残差误差问题，提出重新思考方法，提升LLM量化效果
Authors: Shuaiting Li, Juncan Deng, Kedong Xu, Rongtao Deng, Hong Gu, Minghan Jiang, Haibin Shen, Kejie Huang | Date: 2026-04-09
Link: http://arxiv.org/abs/2604.07955v1
Abstract
Methods based on weight compensation, which iteratively apply quantization and weight compensation to minimize the output error, have recently demonstrated remarkable success in quantizing Large Language Models (LLMs). The representative work, GPTQ, introduces several key techniques that make such iterative methods practical for LLMs with billions of parameters. GPTAQ extends this approach by introducing an asymmetric calibration process that aligns the output of each quantized layer with its full-precision counterpart, incorporating a residual error into the weight compensation framework. In this work, we revisit the formulation of the residual error. We identify a sub-optimal calibration objective in existing methods: during the intra-layer calibration process, they align the quantized output with the output from compensated weights, rather than the true output from the original full-precision model. Therefore, we redefine the objective to precisely align the quantized model's output with the original output of the full-precision model at each step. We then reveal that the residual error originates not only from the output difference of the preceding layer but also from the discrepancy between the compensated and original weights within each layer, which we name the 'compensation-aware error'. By inheriting the neuron decomposition technique from GPTAQ, we can efficiently incorporate this compensation-aware error into the weight update process. Extensive experiments on various LLMs and quantization settings demonstrate that our proposed enhancements integrate seamlessly with both GPTQ and GPTAQ, significantly improving their quantization performance. Our code is publicly available at https://github.com/list0830/ResComp.Bit-by-Bit: Progressive QAT Strategy with Outlier Channel Splitting for Stable Low-Bit LLMs
解决低比特QAT收敛不稳定问题，提出渐进式训练与异常通道分割方法，提升训练稳定性
Authors: Binxing Xu, Hao Gu, Lujun Li, Hao Wang, Bei Liu, Jiacheng Liu, Qiyuan Zhu, Xintong Yang, Chao Li, Sirui Han, Yike Guo | Date: 2026-04-09
Link: http://arxiv.org/abs/2604.07888v1
Abstract
Training LLMs at ultra-low precision remains a formidable challenge. Direct low-bit QAT often suffers from convergence instability and substantial training costs, exacerbated by quantization noise from heavy-tailed outlier channels and error accumulation across layers. To address these issues, we present Bit-by-Bit, a progressive QAT framework with outlier channel splitting. Our approach integrates three key components: (1) block-wise progressive training that reduces precision stage by stage, ensuring stable initialization for low-bit optimization; (2) nested structure of integer quantization grids to enable a "train once, deploy any precision" paradigm, allowing a single model to support multiple bit-widths without retraining; (3) rounding-aware outlier channel splitting, which mitigates quantization error while acting as an identity transform that preserves the quantized outputs. Furthermore, we follow microscaling groups with E4M3 scales, capturing dynamic activation ranges in alignment with OCP/NVIDIA standards. To address the lack of efficient 2-bit kernels, we developed custom operators for both W2A2 and W2A16 configurations, achieving up to 11$\times$ speedup over BF16. Under W2A2 settings, Bit-by-Bit significantly outperforms baselines like BitDistiller and EfficientQAT on both Llama2/3, achieving a loss of only 2.25 WikiText2 PPL compared to full-precision models.QaRL: Rollout-Aligned Quantization-Aware RL for Fast and Stable Training under Training--Inference Mismatch
针对RL训练中量化-推理不匹配问题，提出对齐量化感知RL方法，提升训练速度与稳定性
Authors: Hao Gu, Hao Wang, Jiacheng Liu, Lujun Li, Qiyuan Zhu, Bei Liu, Binxing Xu, Lei Wang, Xintong Yang, Sida Lin, Sirui Han, Yike Guo | Date: 2026-04-09
Link: http://arxiv.org/abs/2604.07853v1
Abstract
Large language model (LLM) reinforcement learning (RL) pipelines are often bottlenecked by rollout generation, making end-to-end training slow. Recent work mitigates this by running rollouts with quantization to accelerate decoding, which is the most expensive stage of the RL loop. However, these setups destabilize optimization by amplifying the training-inference gap: rollouts are operated at low precision, while learning updates are computed at full precision. To address this challenge, we propose QaRL (Rollout Alignment Quantization-Aware RL), which aligns training-side forward with the quantized rollout to minimize mismatch. We further identify a failure mode in quantized rollouts: long-form responses tend to produce repetitive, garbled tokens (error tokens). To mitigate these problems, we introduce TBPO (Trust-Band Policy Optimization), a sequence-level objective with dual clipping for negative samples, aimed at keeping updates within the trust region. On Qwen3-30B-A3B MoE for math problems, QaRL outperforms quantized-rollout training by +5.5 while improving stability and preserving low-bit throughput benefits.AsyncTLS: Efficient Generative LLM Inference with Asynchronous Two-level Sparse Attention
解决长上下文注意力复杂度与内存问题，提出异步两级稀疏注意力系统，平衡精度与效率
Authors: Yuxuan Hu, Jianchao Tan, Jiaqi Zhang, Wen Zan, Pingwei Sun, Yifan Lu, Yerui Sun, Yuchen Xie, Xunliang Cai, Jing Zhang | Date: 2026-04-09
Link: http://arxiv.org/abs/2604.07815v1
Abstract
Long-context inference in LLMs faces the dual challenges of quadratic attention complexity and prohibitive KV cache memory. While token-level sparse attention offers superior accuracy, its indexing overhead is costly; block-level methods improve efficiency but sacrifice precision. We propose AsyncTLS, a hierarchical sparse attention system that combines coarse-grained block filtering with fine-grained token selection to balance accuracy and efficiency, coupled with an asynchronous offloading engine that overlaps KV cache transfers with computation via temporal locality exploitation. Evaluated on Qwen3 and GLM-4.7-Flash across GQA, and MLA architectures, AsyncTLS achieves accuracy comparable to full attention while delivering 1.2x - 10.0x operator speedups and 1.3x - 4.7x end-to-end throughput improvements on 48k - 96k contexts.HAWK: Head Importance-Aware Visual Token Pruning in Multimodal Models
针对多模态模型视觉令牌冗余问题，提出头部重要性感知剪枝方法，降低推理开销
Authors: Qihui Zhu, Tao Zhang, Yuchen Wang, Zijian Wen, Mengjie Zhang, Shuangwu Chen, Xiaobin Tan, Jian Yang, Yang Liu, Zhenhua Dong, Xianzhi Yu, Yinfei Pan | Date: 2026-04-09
Link: http://arxiv.org/abs/2604.07812v1
Abstract
In multimodal large language models (MLLMs), the surge of visual tokens significantly increases the inference time and computational overhead, making them impractical for real-time or resource-constrained applications. Visual token pruning is a promising strategy for reducing the cost of MLLM inference by removing redundant visual tokens. Existing research usually assumes that all attention heads contribute equally to the visual interpretation. However, our study reveals that different heads may capture distinct visual semantics and inherently play distinct roles in visual processing. In light of this observation, we propose HAWK, a head importance-aware visual token pruning method that perceives the varying importance of attention heads in visual tasks to maximize the retention of crucial tokens. By leveraging head importance weights and text-guided attention to assess visual token significance, HAWK effectively retains task-relevant visual tokens while removing redundant ones. The proposed HAWK is entirely training-free and can be seamlessly applied to various MLLMs. Extensive experiments on multiple mainstream vision-language benchmarks demonstrate that HAWK achieves state-of-the-art accuracy. When applied to Qwen2.5-VL, HAWK retains 96.0% of the original accuracy after pruning 80.2% of the visual tokens. Additionally, it reduces end-to-end latency to 74.4% of the original and further decreases GPU memory usage across the tested models. The code is available at https://github.com/peppery77/HAWK.git.GRASS: Gradient-based Adaptive Layer-wise Importance Sampling for Memory-efficient Large Language Model Fine-tuning
解决大模型微调内存瓶颈，提出梯度自适应层重要性采样方法，提升内存效率与性能
Authors: Kaiyuan Tian, Yu Tang, Gongqingjian Jiang, Baihui Liu, Yifu Gao, Xialin Su, Linbo Qiao, Dongsheng Li | Date: 2026-04-09
Link: http://arxiv.org/abs/2604.07808v1
Abstract
Full-parameter fine-tuning of large language models is constrained by substantial GPU memory requirements. Low-rank adaptation methods mitigate this challenge by updating only a subset of parameters. However, these approaches often limit model expressiveness and yield lower performance than full-parameter fine-tuning. Layer-wise fine-tuning methods have emerged as an alternative, enabling memory-efficient training through static layer importance sampling strategies. However, these methods overlook variations in layer importance across tasks and training stages, resulting in suboptimal performance on downstream tasks. To address these limitations, we propose GRASS, a gradient-based adaptive layer-wise importance sampling framework. GRASS utilizes mean gradient norms as a task-aware and training-stage-aware metric for estimating layer importance. Furthermore, GRASS adaptively adjusts layer sampling probabilities through an adaptive training strategy. We also introduce a layer-wise optimizer state offloading mechanism that overlaps computation and communication to further reduce memory usage while maintaining comparable training throughput. Extensive experiments across multiple models and benchmarks demonstrate that GRASS consistently outperforms state-of-the-art methods, achieving an average accuracy improvement of up to 4.38 points and reducing memory usage by up to 19.97\%.SepSeq: A Training-Free Framework for Long Numerical Sequence Processing in LLMs
解决LLM处理长数值序列性能退化问题，提出训练-free分离序列框架，提升注意力集中度
Authors: Jie Sun, Yu Liu, Lu Han, Qiwen Deng, Xiang Shu, Yang Xiao, Xingyu Lu, Jun Zhou, Pengfei Liu, Lintao Ma, Jiancan Wu, Xiang Wang | Date: 2026-04-09
Link: http://arxiv.org/abs/2604.07737v1
Abstract
While transformer-based Large Language Models (LLMs) theoretically support massive context windows, they suffer from severe performance degradation when processing long numerical sequences. We attribute this failure to the attention dispersion in the Softmax mechanism, which prevents the model from concentrating attention. To overcome this, we propose Separate Sequence (SepSeq), a training-free, plug-and-play framework to mitigate dispersion by strategically inserting separator tokens. Mechanistically, we demonstrate that separator tokens act as an attention sink, recalibrating attention to focus on local segments while preserving global context. Extensive evaluations on 9 widely-adopted LLMs confirm the effectiveness of our approach: SepSeq yields an average relative accuracy improvement of 35.6% across diverse domains while reducing total inference token consumption by 16.4% on average.MIPT-SSM: Scaling Language Models with Inference Cache via Phase Transitions
提出基于相变的序列架构，实现O(1)推理缓存，提升语言模型扩展效率
Authors: Yasong Fan | Date: 2026-04-09
Link: http://arxiv.org/abs/2604.07716v1