LLM Inference Acceleration

LLM: Large Language Model

Ask about this topic

Overview

This field studies how to make large language models answer faster and use less memory and computing power, especially during prompt processing and token-by-token generation. Common approaches include speculative decoding, which drafts and verifies tokens; sparse attention and KV-cache compression for long contexts; low-bit quantization; and adaptive computation that skips or reallocates work. Researchers test these methods on long-context and retrieval benchmarks such as LongBench and RULER, as well as reasoning and coding tasks, measuring both quality and serving metrics such as time to first token and throughput. A central challenge is turning theoretical reductions in computation or memory into real speedups on particular hardware without sacrificing accuracy or changing the model’s output distribution.

Latest papers

CardsList

1,101 papers202 papers in the last four weeks, up 51% on the four weeks before. 1.2% of all new papers.

Apr 7, 2026

ForkKV: Scaling Multi-LoRA Agent Serving via Copy-on-Write Disaggregated KV Cache

The serving paradigm of large language models (LLMs) is rapidly shifting towards complex multi-agent workflows where specialized agents collaborate over massive shared contexts. While Low-Rank Adaptation (LoRA) enables the efficient co-hosting of these specialized agents on a single base model, it introduces a critical memory footprint bottleneck during serving. Specifically, unique LoRA activations cause Key-Value (KV) cache divergence across agents, rendering traditional prefix caching ineffective for shared contexts. This forces redundant KV cache maintenance, rapidly saturating GPU capacity and degrading throughput. To address this challenge, we introduce ForkKV, a serving system for multi-LoRA agent workflows centered around a novel memory management paradigm in OS: fork with copy-on-write (CoW). By exploiting the structural properties of LoRA, ForkKV physically decouples the KV cache into a massive shared component (analogous to the parent process's memory pages) and lightweight agent-specific components (the child process's pages). To support this mechanism, we propose a DualRadixTree architecture that allows newly forked agents to inherit the massive shared cache and apply CoW semantics for their lightweight unique cache. Furthermore, to guarantee efficient execution, we design ResidualAttention, a specialized kernel that reconstructs the disaggregated KV cache directly within on-chip SRAM. Comprehensive evaluations across diverse language models and practical datasets of different tasks demonstrate that ForkKV achieves up to 3.0x the throughput of state-of-the-art multi-LoRA serving systems with a negligible impact on generation quality.

Shao Wang, Rui Ren, Lin Gui

Shanghai Jiao Tong University, Shanghai, China

LLM ServingLLM Inference AccelerationDisaggregated LLM ServingKV-Cache Reuse
AskPDF
Apr 7, 2026

STDec: Spatio-Temporal Stability Guided Decoding for dLLMs

Diffusion Large Language Models (dLLMs) have achieved rapid progress, viewed as a promising alternative to the autoregressive paradigm. However, most dLLM decoders still adopt a global confidence threshold, and do not explicitly model local context from neighboring decoded states or temporal consistency of predicted token IDs across steps. To address this issue, we propose a simple spatio-temporal stability guided decoding approach, named STDec. We observe strong spatio-temporal stability in dLLM decoding: newly decoded tokens tend to lie near decoded neighbors, and their predicted IDs often remain consistent across several denoising steps. Inspired by this stability, our STDec includes spatial-aware decoding and temporal-aware decoding. The spatial-aware decoding dynamically generates the token-adaptive threshold by aggregating the decoded states of nearby tokens. The temporal-aware decoding relaxes the decoding thresholds for tokens whose predicted token IDs remain consistent over denoising steps. Our STDec is training-free and remains compatible with cache-based acceleration methods. Across textual reasoning and multimodal understanding benchmarks, STDec substantially improves throughput while maintaining comparable task performance score. Notably, on MBPP with LLaDA, STDec achieves up to 14.17x speedup with a comparable score. Homepage: https://yzchen02.github.io/STDec.

Yuzhe Chen, Jiale Cao, Xuyang Liu +3

1Tianjin University · 2Sichuan University · 3Chongqing University

Diffusion Language Model DecodingDiffusion Language ModelsLLM Inference Acceleration
AskPDF
Apr 7, 2026

BOSCH: Black-Box Binary Optimization for Short-Context Attention-Head Selection in LLMs

Post-training hybridization of large language models (LLMs) often replaces quadratic self-attention with sliding-window attention (SWA) to reduce KV cache usage and improve latency. Existing hybridization schemes are typically defined either at the layer level (e.g., interleaving) or at the head level via static rankings from local to global. Layer-level schemes ignore that local and global dependencies are routed through heads within the same layer, while static head-level rankings suffer from entanglement: a head's local/global behavior can change after hybridization. We propose BOSCH, Black-box Binary Optimization for Short-context Head Selection, a training-free method that formulates the problem as a Large Neighborhood Search and decomposes it into three subproblems: (i) layer-importance detection via small-budget black-box probes, (ii) adaptive per-layer SWA-ratio assignment based on these sensitivities, and (iii) grouped head-level optimization within ratio buckets. Extensive experiments on 4 LLMs ranging from 1.7B to 30B parameters, across 4 SWA ratios, show that BOSCH consistently outperforms layer-level heuristics and 6 strong static head-level methods, with larger gains at higher SWA ratios. Under continual pretraining, BOSCH recover original long-context performance faster and to a higher level. Analysis of the selected heads reveals substantial turnover for BOSCH across different SWA ratios, underscoring the importance of performing head-level selection for each target ratio rather than relying on fixed locality rankings.

Abbas Ghaddar, Ivan Kobyzev, Boxing Chen +1

Huawei Noah’s Ark Lab, Montreal Research Center, Canada

Hybrid AttentionLLM Inference AccelerationSliding-Window Attention
AskPDF
Apr 7, 2026

Attention Editing: A Versatile Framework for Cross-Architecture Attention Conversion

Key-Value (KV) cache memory and bandwidth increasingly dominate large language model inference cost in long-context and long-generation regimes. Architectures such as multi-head latent attention (MLA) and hybrid sliding-window attention (SWA) can alleviate this bound, but integrating them into existing models remains difficult. Prior methods impose fine-grained structural requirements on both source and target attention modules, which cannot meet the feasible requirement in practical deployment. We present Attention Editing, a practical framework for converting already-trained large language models (LLMs) with new attention architectures without re-pretraining from scratch. Attention editing replaces the original attention with a learnable target module and trains it using progressive distillation, consisting of (1) layer-wise teacher-forced optimization with intermediate activation supervision to prevent cold-start error accumulation, and (2) model-level distillation on next-token distributions, optionally regularized by weak feature matching. We instantiate the framework on two different target--MLA and GateSWA, a gated hybrid SWA design, and apply it to Qwen3-8B and Qwen3-30B-A3B. The resulting models maintain competitive performance while delivering substantial efficiency improvements, demonstrating that large-scale attention conversion is both feasible and robust. Notably, experiments are conducted on an Ascend 910B clusters, offering a practical training case study on domestic hardware.

Zhen Cheng, Hao-Bo Yang, Wan-Yi Huang +1

China Merchants Bank Artificial Intelligence Laboratory

LLM Inference Acceleration
AskPDF
Apr 7, 2026

See the Forest for the Trees: Loosely Speculative Decoding via Visual-Semantic Guidance for Efficient Inference of Video LLMs

Video Large Language Models (Video-LLMs) excel in video understanding but suffer from high inference latency during autoregressive generation. Speculative Decoding (SD) mitigates this by applying a draft-and-verify paradigm, yet existing methods are constrained by rigid exact-match rules, severely limiting the acceleration potential. To bridge this gap, we propose LVSpec, the first training-free loosely SD framework tailored for Video-LLMs. Grounded in the insight that generation is governed by sparse visual-relevant anchors (mandating strictness) amidst abundant visual-irrelevant fillers (permitting loose verification), LVSpec employs a lightweight visual-relevant token identification scheme to accurately pinpoint the former. To further maximize acceptance, we augment this with a position-shift tolerant mechanism that effectively salvages positionally mismatched but semantically equivalent tokens. Experiments demonstrate that LVSpec achieves high fidelity and speed: it preserves >99.8 of target performance while accelerating Qwen2.5-VL-32B by 2.70x and LLaVA-OneVision-72B by 2.94x. Notably, it boosts the mean accepted length and speedup ratio by 136% and 35% compared to SOTA training-free SD methods for Video-LLMs.

Yicheng Ji, Jun Zhang, Jinpeng Chen +4

ZJU · BUPT

Speculative DecodingLLM Inference AccelerationVideo-Language Models
AskPDF
Apr 7, 2026

Residual-Mass Accounting for Partial-KV Decoding

We study a controlled partial-KV decoding setting in which exact unnormalized softmax contributions are computed for sink/tail anchors and a retrieved token set, while the remaining prefill tokens are represented by a residual estimate. We focus on the accounting rule after the query-dependent exact support has been selected, and use exhaustive Top-K only as an oracle selector, not as a deployable retrieval system. The proposed rule leaves the backbone language model and the exact-branch KV tensors unchanged. It builds fixed-size summary states (S,u)(S,u) from learned positive feature maps φφ, subtracts retrieved-token feature contributions to keep the exact and residual sets non-overlapping, and merges the estimated residual numerator and denominator with the exact branch under one normalization. At a 1% exact-support budget, our residual-completion method improves over the selection-only Top-K baseline on RULER and BABILong across frozen 1B and 3B Llama-3.2-Instruct backbones at all reported context lengths. In the 0.5-4% exact-support budget sweeps, this trend largely persists. On LongBench, summarization results are mostly favorable, while multi-document QA is mixed. Attention-output diagnostics support retrieved-token subtraction as the partition-consistent accounting rule, while indicating that the main remaining error is imperfect learned-φφ approximation of the unretrieved residual mass.

Yasuto Hoshi, Daisuke Miyashita, Jun Deguchi

Kioxia Corporation

Softmax AttentionLong-Context Language ModelingLLM Inference Acceleration
AskPDF
Apr 7, 2026

Multi-Drafter Speculative Decoding with Alignment Feedback

Speculative decoding (SD) accelerates large language model (LLM) inference by using a smaller model to draft future tokens, which are then verified by the target LLM. This preserves generation quality by accepting only aligned tokens. However, individual drafters, often trained for specific tasks or domains, exhibit limited effectiveness across diverse applications. To address this, we introduce MetaSD, a unified framework that integrates multiple drafters into the SD process. MetaSD dynamically allocates computational resources to heterogeneous drafters by leveraging alignment feedback and framing drafter selection as a multi-armed bandit problem. Extensive experiments show MetaSD consistently outperforms single-drafter approaches.

Taehyeon Kim, Hojung Jung, Se-Young Yun

LG AI Research · KAIST AI

Speculative DecodingLLM Inference Acceleration
AskPDF
Apr 6, 2026

Rethinking Model Efficiency: Multi-Agent Inference with Large Models

Most vision-language models (VLMs) apply a large language model (LLM) as the decoder, where the response tokens are generated sequentially through autoregression. Therefore, the number of output tokens can be the bottleneck of the end-to-end latency. However, different models may require vastly different numbers of output tokens to achieve comparable performance. In this work, we conduct a comprehensive analysis of the latency across different components of VLMs on simulated data. The experiment shows that a large model with fewer output tokens can be more efficient than a small model with a long output sequence. The empirical study on diverse real-world benchmarks confirms the observation that a large model can achieve better or comparable performance as a small model with significantly fewer output tokens. To leverage the efficiency of large models, we propose a multi-agent inference framework that keeps large models with short responses but transfers the key reasoning tokens from the small model when necessary. The comparison on benchmark tasks demonstrates that by reusing the reasoning tokens from small models, it can help approach the performance of a large model with its own reasoning, which confirms the effectiveness of our proposal.

Sixun Dong, Juhua Hu, Steven Li +2

Independent Researcher · University of Washington · Meta Reality Labs

Efficient VLM InferenceLLM Inference AccelerationMulti-Agent LLMsVLM Reasoning
AskPDF
Apr 6, 2026

TriAttention: Efficient Long Reasoning with Trigonometric KV Compression

Extended reasoning in large language models (LLMs) creates severe KV cache memory bottlenecks. Leading KV cache compression methods estimate KV importance using attention scores from recent post-RoPE queries. However, queries rotate with position during RoPE, making representative queries very few, leading to poor top-key selection and unstable reasoning. To avoid this issue, we turn to the pre-RoPE space, where we observe that Q and K vectors are highly concentrated around fixed non-zero centers and remain stable across positions -- Q/K concentration. We show that this concentration causes queries to preferentially attend to keys at specific distances (e.g., nearest keys), with the centers determining which distances are preferred via a trigonometric series. Based on this, we propose TriAttention to estimate key importance by leveraging these centers. Via the trigonometric series, we use the distance preference characterized by these centers to score keys according to their positions, and also leverage Q/K norms as an additional signal for importance estimation. On AIME25 with 32K-token generation, TriAttention matches Full Attention reasoning accuracy while achieving 2.5x higher throughput or 10.7x KV memory reduction, whereas leading baselines achieve only about half the accuracy at the same efficiency. TriAttention enables OpenClaw deployment on a single consumer GPU, where long context would otherwise cause out-of-memory with Full Attention.

Weian Mao, Xi Lin, Wei Huang +5

MIT · ZJU · NVIDIA

Long-Context Reasoning in Language ModelsKV-Cache CompressionLLM Inference AccelerationAttention Mechanisms
AskPDF
Apr 6, 2026

Don't Waste Bits! Adaptive KV-Cache Quantization for Lightweight On-Device LLMs

Large Language Models (LLMs) have achieved remarkable progress across reasoning, generation, and decision-making tasks, yet deploying them on mobile, embedded, and edge devices remains particularly challenging. On-device LLM inference is heavily constrained by the memory and bandwidth overhead of the key-value (KV) cache, which grows linearly with context length and often dominates decoding cost. Existing KV-cache quantization schemes typically rely on fixed precision or hand-crafted heuristics, thereby wasting bits on low-impact tokens while over-compressing informative ones, leading to avoidable accuracy degradation. Inspired by Huffman coding's principle of variable-length allocation, we propose adaptive KV-cache quantization, a learned policy that assigns bit-width proportional to token importance, minimizing expected memory and latency without sacrificing competitive accuracy. Our framework extracts lightweight token-level features, including token frequency, quality score, attention variance, and entropy-based uncertainty, and feeds them into a compact data-driven controller that dynamically selects KV precision from {2-bit, 4-bit, 8-bit, FP16} during decoding. This adaptive precision policy reduces KV memory footprint and latency while improving accuracy compared to static KV quantization and rule-based baselines, and maintaining competitive accuracy close to FP16 inference across standard LLM benchmarks. Extensive experiments across multiple commonsense reasoning benchmarks using SmolLM-135M, SmolLM-360M, and SmolLM-1.7B demonstrate that our controller consistently improves the accuracy-latency trade-off. For instance, with SmolLM-360M on HellaSwag, our method reduces decoding latency (ms/token) by 17.75% relative to static KV quantization, improves accuracy by 7.60 points, and remains within only 0.30 points of FP16 inference.

Sayed Pedram Haeri Boroujeni, Niloufar Mehrabi, Patrick Woods +2

Clemson University, USA

KV-Cache QuantizationOn-Device Language Model InferenceLLM Inference AccelerationMixed-Precision Quantization
AskPDF
Apr 5, 2026

BWTA: Accurate and Efficient Binarized Transformer by Algorithm-Hardware Co-design

Ultra low-bit quantization brings substantial efficiency for Transformer-based models, but the accuracy degradation and limited GPU support hinder its wide usage. In this paper, we analyze zero-point distortion in binarization and propose a Binary Weights & Ternary Activations (BWTA) quantization scheme, which projects tiny values to zero and preserves the accuracy of extremely low-bit models. For training, we propose Smooth Multi-Stage Quantization, combining a Levelwise Degradation Strategy and a Magnitude-Alignment Projection Factor to enable stable and fast convergence. For inference, we develop a BWTA MatMul CUDA kernel with instruction-level parallel bit-packing and comprehensive binary/ternary MatMul implementations for both linear and attention operators, allowing seamless integration across Transformer architectures. Experiments show that BWTA approaches full-precision performance for BERT, with an average 3.5% drop on GLUE and less than 2% drop on five tasks, and achieves comparable perplexity and accuracy for LLMs. In efficiency, it delivers 16 to 24 times kernel-level speedup over FP16 on NVIDIA GPUs, and 216 to 330 tokens/s end-to-end prefill speedup with lower memory footprint on LLMs. As an algorithm-hardware co-design, BWTA demonstrates practical, low-latency ultra-low-bit inference without sacrificing model quality.

Yifu Ding, Xianglong Liu, Shenghao Jin +2

School of Computer Science and Engineering, Beihang University, Beijing, China. · Institute of Artificial Intelligence, Beihang University, Beijing, China. · Department of Automation, Tsinghua University, Beijing, China.

Binary QuantizationLow-Bit QuantizationTransformer InferenceLLM Inference AccelerationQuantization-Aware TrainingActivation Quantization
AskPDF
Apr 5, 2026

Diagonal-Tiled Mixed-Precision Attention for Efficient Low-Bit MXFP Inference

Transformer-based large language models (LLMs) have demonstrated remarkable performance across a wide range of real-world tasks, but their inference cost remains prohibitively high due to the quadratic complexity of attention and the memory bandwidth limitations of high-precision operations. In this work, we present a low-bit mixed-precision attention kernel using the microscaling floating-point (MXFP) data format, utilizing the computing capability on next-generation GPU architectures. Our Diagonal-Tiled Mixed-Precision Attention (DMA) incorporates two kinds of low-bit computation at the tiling-level, and is a delicate fused kernel implemented using Triton, exploiting hardware-level parallelism and memory efficiency to enable fast and efficient inference without compromising model performance. Extensive empirical evaluations on NVIDIA B200 GPUs show that our kernel maintains generation quality with negligible degradation, and meanwhile achieves significant speedup by kernel fusion. We release our code at https://github.com/yifu-ding/MP-Sparse-Attn.

Yifu Ding, Xinhao Zhang, Jinyang Guo

State Key Laboratory of Complex & Critical Software Environment (SKLCCSE), Beihang University · School of Computer Science and Engineering, Beihang University · Beijing Jiaotong University +1

LLM Inference AccelerationGPU Kernel OptimizationAttention QuantizationMixed-Precision Quantization
AskPDF
Apr 5, 2026

Cactus: Accelerating Auto-Regressive Decoding with Constrained Acceptance Speculative Sampling

Speculative sampling (SpS) has been successful in accelerating the decoding throughput of auto-regressive large language models by leveraging smaller draft models. SpS strictly enforces the generated distribution to match that of the verifier LLM. This is unnecessarily restrictive as slight variations of the verifier's distribution, such as sampling with top-kk or temperature, would also be acceptable. Typical acceptance sampling (TAS) alleviates this issue by accepting more tokens using entropy-based heuristics. However, this approach distorts the verifier distribution, potentially degrading output quality when the verifier encodes critical information. In this work, we formalize the speculative sampling algorithm through the lens of constrained optimization. Based on this formulation, we propose Cactus (constrained acceptance speculative sampling), a method that guarantees controlled divergence from the verifier distribution and increasing acceptance rates. Empirical results across a wide range of benchmarks confirm the effectiveness of our approach.

Yongchang Hao, Lili Mou

Dept. Computing Science & Alberta Machine Intelligence Institute (Amii), University of Alberta · Canada CIFAR AI Chair

Speculative DecodingLLM Inference AccelerationAutoregressive Decoding
AskPDF
Apr 4, 2026

Researchers waste 80% of LLM annotation costs by classifying one text at a time

Large language models (LLMs) are increasingly being used for text classification across the social sciences, yet researchers overwhelmingly classify one text per variable per prompt. Coding 100,000 texts on four variables requires 400,000 API calls. Batching 25 items and stacking all variables into a single prompt reduces this to 4,000 calls, cutting token costs by over 80%. Whether this degrades coding quality is unknown. We tested eight production LLMs from four providers on 3,962 expert-coded tweets across four tasks, varying batch size from 1 to 1,000 items and stacking up to 25 coding dimensions per prompt. Six of eight models maintained accuracy within 2 pp of the single-item baseline through batch sizes of 100. Variable stacking with up to 10 dimensions produced results comparable to single-variable coding, with degradation driven by task complexity rather than prompt length. Within this safe operating range, the measurement error from batching and stacking is smaller than typical inter-coder disagreement in the ground-truth data.

Christian Pipal, Eva-Maria Vogel, Morgan Wack +1

Department of Communication and Media Research (IKMZ), University of Zurich

LLM-Assisted AnnotationLarge Language Model-Based Text ClassificationLLM Inference EfficiencyLLM EvaluationLLM Inference Acceleration
AskPDF
Apr 4, 2026

LightThinker++: From Reasoning Compression to Memory Management

Large language models (LLMs) excel at complex reasoning, yet their efficiency is limited by the surging cognitive overhead of long thought traces. In this paper, we propose LightThinker, a method that enables LLMs to dynamically compress intermediate thoughts into compact semantic representations. However, static compression often struggles with complex reasoning where the irreversible loss of intermediate details can lead to logical bottlenecks. To address this, we evolve the framework into LightThinker++, introducing Explicit Adaptive Memory Management. This paradigm shifts to behavioral-level management by incorporating explicit memory primitives, supported by a specialized trajectory synthesis pipeline to train purposeful memory scheduling. Extensive experiments demonstrate the framework's versatility across three dimensions. (1) LightThinker reduces peak token usage by 70% and inference time by 26% with minimal accuracy loss. (2) In standard reasoning, LightThinker++ slashes peak token usage by 69.9% while yielding a +2.42% accuracy gain under the same context budget for maximum performance. (3) Most notably, in long-horizon agentic tasks, it maintains a stable footprint beyond 80 rounds (a 60%-70% reduction), achieving an average performance gain of 14.8% across different complex scenarios. Overall, our work provides a scalable direction for sustaining deep LLM reasoning over extended horizons with minimal overhead.

Yuqi Zhu, Jintian Zhang, Zhenjie Wan +7

Zhejiang University · Ant Group

Memory Management for LLM AgentsEfficient Language Model ReasoningCoT CompressionLLM Inference AccelerationAgent Memory ManagementMemory-Augmented Language Model Agents
AskPDF
Apr 3, 2026

JoyAI-LLM Flash: Advancing Mid-Scale LLMs with Token Efficiency

We introduce JoyAI-LLM Flash, an efficient Mixture-of-Experts (MoE) language model designed to redefine the trade-off between strong performance and token efficiency in the sub-50B parameter regime. JoyAI-LLM Flash is pretrained on a massive corpus of 20 trillion tokens and further optimized through a rigorous post-training pipeline, including supervised fine-tuning (SFT), Direct Preference Optimization (DPO), and large-scale reinforcement learning (RL) across diverse environments. To improve token efficiency, JoyAI-LLM Flash strategically balances thinking and non-thinking cognitive modes and introduces FiberPO, a novel RL algorithm inspired by fibration theory that decomposes trust-region maintenance into global and local components, providing unified multi-scale stability control for LLM policy optimization. To enhance architectural sparsity, the model comprises 48B total parameters while activating only 2.7B parameters per forward pass, achieving a substantially higher sparsity ratio than contemporary industry leading models of comparable scale. To further improve inference throughput, we adopt a joint training-inference co-design that incorporates dense Multi-Token Prediction (MTP) and Quantization-Aware Training (QAT). We release the checkpoints for both JoyAI-LLM-48B-A3B Base and its post-trained variants on Hugging Face to support the open-source community.

Aichen Cai, Anmeng Zhang, Anyu Li +66

Language ModelingMixture-of-Experts Language ModelsRL for Language ModelsLLM Inference Acceleration
AskPDF
Apr 3, 2026

Prompt Compression in the Wild: Measuring Latency, Rate Adherence, and Quality for Faster LLM Inference

With the wide adoption of language models for IR -- and specifically RAG systems -- the latency of the underlying LLM becomes a crucial bottleneck, since the long contexts of retrieved passages lead large prompts and therefore, compute increase. Prompt compression, which reduces the size of input prompts while aiming to preserve performance on downstream tasks, has established itself as a cost-effective and low-latency method for accelerating inference in large language models. However, its usefulness depends on whether the additional preprocessing time during generation is offset by faster decoding. We present the first systematic, large-scale study of this trade-off, with thousands of runs and 30,000 queries across several open-source LLMs and three GPU classes. Our evaluation separates compression overhead from decoding latency while tracking output quality and memory usage. LLMLingua achieves up to 18% end-to-end speed-ups, when prompt length, compression ratio, and hardware capacity are well matched, with response quality remaining statistically unchanged across summarization, code generation, and question answering tasks. Outside this operating window, however, the compression step dominates and cancels out the gains. We also show that effective compression can reduce memory usage enough to offload workloads from data center GPUs to commodity cards, with only a 0.3s increase in latency. Our open-source profiler predicts the latency break-even point for each model-hardware setup, providing practical guidance on when prompt compression delivers real-world benefits.

Cornelius Kummer, Lena Jurkschat, Michael Färber +1

ScaDS.AI Dresden/Leipzig, CIDS, TU Dresden

Prompt CompressionLLM Inference EfficiencyLLM Inference Acceleration
AskPDF
Apr 3, 2026

FluxMoE: Decoupling Expert Residency for High-Performance MoE Serving

Mixture-of-Experts (MoE) models have become mainstream for scaling language models to hundreds of billions of expert parameters. Despite sparse expert activation, existing inference engines keep all experts GPU-resident, crowding out the key-value cache in large-batch, long-output offline workloads. We present FluxMoE, which decouples experts from physical GPU residency and adapts their footprint to available memory through a new expert paging abstraction. FluxMoE combines PagedTensor for transparent remapping, a bandwidth-balanced hierarchy spanning losslessly compressed GPU memory and host DRAM, and a budget-aware residency planner. Unlike CPU-GPU co-inference and whole-layer offloading, FluxMoE streams weights on demand while keeping expert computation on GPUs. We implement FluxMoE atop vLLM and evaluate it on three MoE models. For GLM-4.5 on 8×\timesH20 GPUs, FluxMoE delivers up to 7.2×\times vLLM's throughput and 79.0% lower average Time-Per-Output-Token (TPOT), without measurable model-quality loss using lossless compression. For Mixtral-8×\times7B-Instruct on 2×\timesL40S GPUs, where weight-resident vLLM cannot fit, FluxMoE delivers 4.3×\times KTransformers's throughput and 29.1% lower average TPOT.

Qingxiu Liu, Yongchao He, Runhan Jiang +4

CUHK · Unaffiliated · SCITIX +1

Expert OffloadingGPU AccelerationMixture of ExpertsMixture-of-Experts InferenceMemory-Efficient InferenceLLM Inference Acceleration
AskPDF
Apr 3, 2026

AXELRAM: Quantize Once, Never Dequantize

We propose AXELRAM, a smart SRAM macro architecture that computes attention scores directly from quantized KV cache indices without dequantization. The key enabler is a design-time fixed codebook: orthogonal-transform-based quantization concentrates each coordinate's distribution to N(0,1/d), so the optimal quantizer depends only on dimension d and bit-width b, not on input data. The asymmetric path design -- transform on write, table-lookup on read with no inverse transform -- reduces per-query multiplications by 102.4x (a mathematical identity). Through multi-seed evaluation (10 seeds x 3 models), we discover that sign pattern sensitivity causes catastrophic PPL spikes (Delta > 50) on certain models (Qwen2.5-3B), while others (LLaMA-3.1-8B) are fully stable. This phenomenon extends SpinQuant's observation of rotation variance in weight quantization to the KV cache domain, where the effect is qualitatively more severe. We trace the root cause to layer-wise norm heterogeneity and propose a gradient-free sign pattern selection (200 candidates, 8 calibration samples, one-time) that eliminates catastrophic spikes with zero additional hardware cost. All source code is available at https://github.com/Axelidea/AXELRAM.

Yasushi Nishida

Axelidea Inc.

Attention QuantizationKV-Cache QuantizationLLM Inference AccelerationCompute-in-MemoryRotation-Based Quantization
AskPDF
Apr 2, 2026

Fast NF4 Dequantization Kernels for Large Language Model Inference

Large language models (LLMs) have grown beyond the memory capacity of single GPU devices, necessitating quantization techniques for practical deployment. While NF4 (4-bit NormalFloat) quantization enables 4×\times memory reduction, inference on current NVIDIA GPUs (e.g., Ampere A100) requires expensive dequantization back to FP16 format, creating a critical performance bottleneck. This paper presents a lightweight shared memory optimization that addresses this gap through principled memory hierarchy exploitation while maintaining full ecosystem compatibility. We compare our technique against the open-source BitsAndBytes implementation, achieving 2.0--2.2×\times kernel speedup across three models (Gemma 27B, Qwen3 32B, and Llama3.3 70B) and up to 1.54×\times end-to-end improvement by leveraging the 12--15×\times latency advantage of shared memory over global memory access. Our optimization reduces instruction counts through simplified indexing logic while using only 64 bytes of shared memory per thread block, demonstrating that lightweight optimizations can deliver substantial performance gains with minimal engineering effort. This work provides a plug-and-play solution for the HuggingFace ecosystem that democratizes access to advanced models on existing GPU infrastructure.

Xiangbo Qi, Chaoyi Jiang, Murali Annavaram

University of Southern California Los Angeles, USA

LLM Inference AccelerationGPU Kernel OptimizationLLM QuantizationNeural Network Quantization
AskPDF
Apr 2, 2026

Goose: Anisotropic Speculation Trees for Training-Free Speculative Decoding

Speculative decoding accelerates large language model inference by drafting multiple candidate tokens and verifying them in a single forward pass. Candidates are organized as a tree: deeper trees accept more tokens per step, but adding depth requires sacrificing breadth (fallback options) under a fixed verification budget. Existing training-free methods draft from a single token source and shape their trees without distinguishing candidate quality across origins. We observe that two common training-free token sources -- n-gram matches copied from the input context, and statistical predictions from prior forward passes -- differ sharply in acceptance rate (~6x median gap, range 2-18x across five models and five benchmarks). We prove that when such a quality gap exists, the optimal tree is anisotropic (asymmetric): reliable tokens should form a deep chain while unreliable tokens spread as wide branches, raising the depth ceiling of balanced trees. We realize this structure in GOOSE, a training-free framework that builds an adaptive spine tree: a deep chain of high-acceptance context-matched tokens with wide branches of low-acceptance alternatives at each node. The resulting tree provably accepts at least as many tokens per step as either source alone. On five LLMs (7B-33B) and five benchmarks, GOOSE achieves 1.9-4.3x lossless speedup, outperforming balanced-tree baselines by 12-33% under the same budget.

Tao Jin, Phuong Minh Nguyen, Naoya Inoue

Japan Advanced Institute of Science and Technology (JAIST)

Speculative DecodingLLM Inference Acceleration
AskPDF
Apr 2, 2026

DWDP: Distributed Weight Data Parallelism for High-Performance LLM Inference on NVL72

Large language model (LLM) inference increasingly depends on multi-GPU execution, yet existing inference parallelization strategies require layer-wise inter-rank synchronization, making end-to-end performance sensitive to workload imbalance. We present DWDP (Distributed Weight Data Parallelism), an inference parallelization strategy that preserves data-parallel execution while offloading MoE weights across peer GPUs and fetching missing experts on demand. By removing collective inter-rank synchronization, DWDP allows each GPU to progress independently. We further address the practical overheads of this design with two optimizations for split-weight management and asynchronous remote-weight prefetch. Implemented in TensorRT-LLM and evaluated with DeepSeek-R1 on GB200 NVL72, DWDP improves end-to-end output TPS/GPU by 8.8% at comparable TPS/user in the 20-100 TPS/user serving range under 8K input sequence length and 1K output sequence length.

Wanqian Li, Jintao Peng, Zongfei Jing +7

NVIDIA

LLM Inference AccelerationDistributed InferenceMixture-of-Experts Inference
AskPDF
Apr 1, 2026

Screening Is Enough

We call query--key relevance absolute when its values lie on a fixed bounded scale, depend on neither competing keys nor sequence length, require no sequence-length-dependent calibration, and can all be zero. To realize this notion, we introduce screening, whose explicit threshold transforms bounded query--key similarities into relevance values, enabling exact rejection, empty selection, and direct inspection on a common scale. In a controlled comparison of 12 attention mechanisms on a matched Transformer backbone, only screening maintains both low long-context perplexity and robust retrieval beyond the training context; notably, it does so without inference-time scaling. Building on screening, we introduce Multiscreen, a language-model architecture composed of parallel gated screening tiles. Multiscreen retains these long-context gains while achieving greater parameter efficiency, stronger general zero-shot downstream performance, lower training cost at larger scales, and lower model-side time to first token than Transformer baselines. We further develop a normalization design that keeps Multiscreen training stable even at a learning rate of 11 and show that an adapted version likewise stabilizes Transformer at the same learning rate.

Ken M. Nakanishi

Center for Emergent Matter Science (CEMS), RIKEN · Graduate School of Science, The University of Tokyo

Self-AttentionLong-Context Language ModelingTransformer AttentionLLM Inference AccelerationAttention Mechanisms
AskPDF
Apr 1, 2026

Optimal Brain Decomposition for Accurate LLM Low-Rank Approximation

Low-rank decomposition has emerged as an important problem in Large Language Model (LLM) fine-tuning and inference. Through Singular Value Decomposition (SVD), the weight matrix can be factorized into low-rank spaces optimally. Previously, a common practice was to decompose the weight in the activation-whitened space, and then achieve satisfying results. In this work, we propose Optimal Brain Decomposition LLM (OBD-LLM), which studies the decomposition problem in the model space by utilizing second-order Hessian information. Through a rigorous Kronecker-factorization of the Hessian, we show that the decomposition needs to consider both input and output information of the layer, and achieves much better decomposition results compared to input only method. Our loss-aware decomposition method involves a bi-directional whitening on the weight matrix. As a result, OBD-LLM is a closed-form solution for the optimal decomposition of weights in the language model. Remarkably, we achieve ~20-40% better results than previous state-of-the-art decomposition methods, the SVD-LLM.

Yuhang Li, Donghyun Lee, Ruokai Yin +1

Yale University · University of Sounthern California

Low-Rank ApproximationLLM CompressionLLM Inference Acceleration
AskPDF
Apr 1, 2026

SpecTr-GBV: Multi-Draft Block Verification Accelerating Speculative Decoding

Autoregressive language models suffer from high inference latency due to their sequential decoding nature. Speculative decoding (SD) mitigates this by employing a lightweight draft model to propose candidate tokens, which are selectively verified by a larger target model. While existing methods either adopt multi-draft strategies to increase acceptance rates or block verification techniques to jointly verify multiple tokens, they remain limited by treating these improvements in isolation. In this work, we propose SpecTr-GBV, a novel SD method that unifies multi-draft and greedy block verification (GBV) into a single framework. By formulating the verification step as an optimal transport problem over draft and target token blocks, SpecTr-GBV improves both theoretical efficiency and empirical performance. We theoretically prove that SpecTr-GBV achieves the optimal expected acceptance length physically attainable within the framework of i.i.d. draft generation, and this bound improves as the number of drafts increases. Empirically, we evaluate SpecTr-GBV across five datasets and four baselines. Our method achieves superior speedup and significantly higher block efficiency while preserving output quality. In addition, we perform comprehensive ablation studies to evaluate the impact of various hyperparameters in the model.

Yijun Lin, Jinhao Sheng, Qingyue Cai +1

Center for Applied Statistics and School of Statistics, Renmin University of China · School of Intelligent Medicine, China Medical University · Beijing Advanced Innovation Center for Future Blockchain and Privacy Computing

Speculative DecodingLLM Inference Acceleration
AskPDF
Apr 1, 2026

Adaptive Parallel Monte Carlo Tree Search for Efficient Test-time Compute Scaling

Monte Carlo Tree Search (MCTS) is an effective test-time compute scaling (TTCS) method for improving the reasoning performance of large language models, but its highly variable execution time leads to severe long-tail latency in practice. Existing optimizations such as positive early exit, reduce latency in favorable cases but are less effective when search continues without meaningful progress. We introduce {\it negative early exit}, which prunes unproductive MCTS trajectories, and an {\it adaptive boosting mechanism} that reallocates reclaimed computation to reduce resource contention among concurrent searches. Integrated into vLLM, these techniques substantially reduce p99 end-to-end latency while improving throughput and maintaining reasoning accuracy.

Hongbeen Kim, Juhyun Lee, Sanghyeon Lee +2

School of Computing, Korea Advanced Institute of Science and Technology, Daejeon, South Korea.

Monte Carlo Tree SearchLLM Inference AccelerationInference-Time ScalingLLM Inference Scheduling
AskPDF
Apr 1, 2026

Executing as You Generate: Hiding Execution Latency in LLM Code Interpreters

Current LLM systems are increasingly equipped with a code interpreter that executes generated code to obtain results. This works serially: the model first generates the complete code, then an interpreter executes it. This sequential workflow leaves the executor idle during generation and the generator idle during execution, resulting in unnecessary end-to-end latency. Our key observation is that an LLM, unlike a human developer, emits code tokens left to right and does not backtrack over what it has already written. This makes it possible to start executing a piece of code while later tokens are still being generated. We formalize this parallel execution paradigm, modeling it as a three-stage pipeline of generation, detection, and execution, and derive closed-form latency bounds that characterize its speedup potential and operating regimes. We then present EAGER, a concrete implementation featuring AST-based chunking, dynamic batching with gated execution, and early error interruption. We evaluate EAGER across four benchmarks, seven LLMs, and three execution environments. The overlap mechanism hides almost all execution behind generation, reducing the non-overlapped portion of execution time by up to 99.8% and cutting end-to-end latency by up to 37.3% on error-free runs.

Zhensu Sun, Zhihao Lin, Zhi Chen +4

Singapore Management University · Beihang University

LLM Inference AccelerationPipeline ParallelismLLM Code Generation
AskPDF
Mar 31, 2026

MAC-Attention: a Match-Amend-Complete Scheme for Fast and Accurate Attention Computation

Long-context decoding in LLMs is IO-bound: each token re-reads an ever-growing KV cache. Prior accelerations cut bytes via compression, which lowers fidelity, or selection/eviction, which restricts what remains accessible, and both can degrade delayed recall and long-form generation. We introduce MAC-Attention, a fidelity- and access-preserving alternative that accelerates decoding by reusing prior attention computations for semantically similar recent queries. It starts with a match stage that performs pre-RoPE L2 matching over a short local window; an amend stage rectifies the reused attention by recomputing a small band near the match boundary; and a complete stage fuses the rectified results with fresh attention computed on the KV tail through a numerically stable merge. On a match hit, the compute and bandwidth complexity is constant regardless of context length. The method is model-agnostic and composes with IO-aware kernels, paged-KV managers, and MQA/GQA. Across LongBench v2 (120K), RULER (120K), and LongGenBench (16K continuous generation), compared to the latest FlashInfer library, MAC-Attention reduces KV accesses by up to 99%, cuts token generation latency by over 60% at 128K, and achieves over 14.3x attention-phase speedups, up to 2.6x end-to-end, while maintaining full-attention quality. By reusing computation, MAC-Attention delivers long-context inference that is both fast and faithful. Code is available here: https://github.com/YJHMITWEB/MAC-Attention.git

Jinghan Yao, Sam Adé Jacobs, Walid Krichene +2

The Ohio State University · Microsoft, WA, USA · Anyscale, CA, USA

LLM Inference AccelerationEfficient AttentionLong-Context Language Model InferenceSelf-Attention
AskPDF
Mar 31, 2026

SISA: A Scale-In Systolic Array for GEMM Acceleration

The currently dominant AI/ML workloads, such as Large Language Models (LLMs), rely on the efficient execution of General Matrix-Matrix Multiplication (GEMM) operations. Thus, most systems are equipped with dedicated matrix hardware accelerators based on square Systolic Arrays (SAs) of Processing Elements (PEs). While this organization was effective for traditional Deep Neural Networks (DNNs), LLMs introduce input-dependent and highly skewed matrices, leading to underutilized SA resources. To address this challenge, we propose SISA (Scale-In Systolic Array), a novel SA architecture that partitions the traditional square array into horizontal rectangular slabs. With minimal overhead, SISA exposes parallelism through independently scheduled slabs for efficient execution of small or skewed matrix shapes, while retaining full-array operation for large GEMMs. SISA achieves up to 8.52x speedup and 93% energy-delay-product (EDP) reduction for representative LLMs compared to a state-of-the-art monolithic SA with the same number of PEs.

Luigi Altamura, Alessio Cicero, Mateo Vázquez Maceiras +2

Department of Computer Science and Engineering, Chalmers University of Technology and University of Gothenburg, Gothenburg, Sweden

LLM Inference Acceleration
AskPDF
Mar 30, 2026

Understand and Accelerate Memory Processing Pipeline for Large Language Model Inference

Modern large language models (LLMs) increasingly depends on efficient long-context processing and generation mechanisms, including sparse attention, retrieval-augmented generation (RAG), and compressed contextual memory, to support complex reasoning. We show that these optimizations can be unified into a four-step memory processing pipeline: Prepare Memory, Compute Relevancy, Retrieval, and Apply to Inference. Through systematic profiling, we identify a 22%-97% memory processing overhead in LLM inference and strong heterogeneity in its computational characteristics. Motivated by this insight, we argue that heterogeneous systems are well-suited to accelerate memory processing and thus end-to-end inference. We demonstrate this approach on a GPU-FPGA system by offloading sparse, irregular, and memory-bounded operations to FPGAs while retaining compute-intensive operations on GPUs. Evaluated on an AMD MI210 GPU and an Alveo U55C FPGA, our system is up to 2.2×2.2\times faster and achieves up to 4.7×4.7\times less energy across multiple LLM inference optimizations than the GPU baseline (similar results hold on NVIDIA A100). These results establish heterogeneous systems as a practical direction for efficient LLM memory processing and inform future heterogeneous hardware design.

Zifan He, Rui Ma, Yizhou Sun +1

Department of Computer Science, University of California, Los Angeles · Microsoft Research Asia

LLM InferenceLLM Inference AccelerationFPGA Acceleration
AskPDF