Zhihang Yuan

Yuzhang Shang, Zhihang Yuan, Qiang Wu et al. · berkeley

This paper explores network binarization, a radical form of quantization, compressing model weights to a single bit, specifically for Large Language Models (LLMs) compression. Due to previous binarization methods collapsing LLMs, we propose a novel approach, Partially-Binarized LLM (PB-LLM), which can achieve extreme low-bit quantization while maintaining the linguistic reasoning capacity of quantized LLMs. Specifically, our exploration first uncovers the ineffectiveness of naive applications of existing binarization algorithms and highlights the imperative role of salient weights in achieving low-bit quantization. Thus, PB-LLM filters a small ratio of salient weights during binarization, allocating them to higher-bit storage, i.e., partially-binarization. PB-LLM is extended to recover the capacities of quantized LMMs, by analyzing from the perspective of post-training quantization (PTQ) and quantization-aware training (QAT). Under PTQ, combining the concepts from GPTQ, we reconstruct the binarized weight matrix guided by the Hessian matrix and successfully recover the reasoning capacity of PB-LLM in low-bit. Under QAT, we freeze the salient weights during training, explore the derivation of optimal scaling factors crucial for minimizing the quantization error, and propose a scaling mechanism based on this derived scaling strategy for residual binarized weights. Those explorations and the developed methodologies significantly contribute to rejuvenating the performance of low-bit quantized LLMs and present substantial advancements in the field of network binarization for LLMs.The code is available at https://github.com/hahnyuan/BinaryLLM.

Jiawei Liu, Lin Niu, Zhihang Yuan et al.

Post-training quantization (PTQ) is a neural network compression technique that converts a full-precision model into a quantized model using lower-precision data types. Although it can help reduce the size and computational cost of deep neural networks, it can also introduce quantization noise and reduce prediction accuracy, especially in extremely low-bit settings. How to determine the appropriate quantization parameters (e.g., scaling factors and rounding of weights) is the main problem facing now. Existing methods attempt to determine these parameters by minimize the distance between features before and after quantization, but such an approach only considers local information and may not result in the most optimal quantization parameters. We analyze this issue and ropose PD-Quant, a method that addresses this limitation by considering global information. It determines the quantization parameters by using the information of differences between network prediction before and after quantization. In addition, PD-Quant can alleviate the overfitting problem in PTQ caused by the small number of calibration sets by adjusting the distribution of activations. Experiments show that PD-Quant leads to better quantization parameters and improves the prediction accuracy of quantized models, especially in low-bit settings. For example, PD-Quant pushes the accuracy of ResNet-18 up to 53.14% and RegNetX-600MF up to 40.67% in weight 2-bit activation 2-bit. The code is released at https://github.com/hustvl/PD-Quant.

Yuzhang Shang, Zhihang Yuan, Bin Xie et al.

Denoising diffusion (score-based) generative models have recently achieved significant accomplishments in generating realistic and diverse data. These approaches define a forward diffusion process for transforming data into noise and a backward denoising process for sampling data from noise. Unfortunately, the generation process of current denoising diffusion models is notoriously slow due to the lengthy iterative noise estimations, which rely on cumbersome neural networks. It prevents the diffusion models from being widely deployed, especially on edge devices. Previous works accelerate the generation process of diffusion model (DM) via finding shorter yet effective sampling trajectories. However, they overlook the cost of noise estimation with a heavy network in every iteration. In this work, we accelerate generation from the perspective of compressing the noise estimation network. Due to the difficulty of retraining DMs, we exclude mainstream training-aware compression paradigms and introduce post-training quantization (PTQ) into DM acceleration. However, the output distributions of noise estimation networks change with time-step, making previous PTQ methods fail in DMs since they are designed for single-time step scenarios. To devise a DM-specific PTQ method, we explore PTQ on DM in three aspects: quantized operations, calibration dataset, and calibration metric. We summarize and use several observations derived from all-inclusive investigations to formulate our method, which especially targets the unique multi-time-step structure of DMs. Experimentally, our method can directly quantize full-precision DMs into 8-bit models while maintaining or even improving their performance in a training-free manner. Importantly, our method can serve as a plug-and-play module on other fast-sampling methods, e.g., DDIM. The code is available at https://github.com/42Shawn/PTQ4DM .

Yizeng Han, Zhihang Yuan, Yifan Pu et al.

Spatial-wise dynamic convolution has become a promising approach to improving the inference efficiency of deep networks. By allocating more computation to the most informative pixels, such an adaptive inference paradigm reduces the spatial redundancy in image features and saves a considerable amount of unnecessary computation. However, the theoretical efficiency achieved by previous methods can hardly translate into a realistic speedup, especially on the multi-core processors (e.g. GPUs). The key challenge is that the existing literature has only focused on designing algorithms with minimal computation, ignoring the fact that the practical latency can also be influenced by scheduling strategies and hardware properties. To bridge the gap between theoretical computation and practical efficiency, we propose a latency-aware spatial-wise dynamic network (LASNet), which performs coarse-grained spatially adaptive inference under the guidance of a novel latency prediction model. The latency prediction model can efficiently estimate the inference latency of dynamic networks by simultaneously considering algorithms, scheduling strategies, and hardware properties. We use the latency predictor to guide both the algorithm design and the scheduling optimization on various hardware platforms. Experiments on image classification, object detection and instance segmentation demonstrate that the proposed framework significantly improves the practical inference efficiency of deep networks. For example, the average latency of a ResNet-101 on the ImageNet validation set could be reduced by 36% and 46% on a server GPU (Nvidia Tesla-V100) and an edge device (Nvidia Jetson TX2 GPU) respectively without sacrificing the accuracy. Code is available at https://github.com/LeapLabTHU/LASNet.

Yizeng Han, Zeyu Liu, Zhihang Yuan et al.

Dynamic computation has emerged as a promising avenue to enhance the inference efficiency of deep networks. It allows selective activation of computational units, leading to a reduction in unnecessary computations for each input sample. However, the actual efficiency of these dynamic models can deviate from theoretical predictions. This mismatch arises from: 1) the lack of a unified approach due to fragmented research; 2) the focus on algorithm design over critical scheduling strategies, especially in CUDA-enabled GPU contexts; and 3) challenges in measuring practical latency, given that most libraries cater to static operations. Addressing these issues, we unveil the Latency-Aware Unified Dynamic Networks (LAUDNet), a framework that integrates three primary dynamic paradigms-spatially adaptive computation, dynamic layer skipping, and dynamic channel skipping. To bridge the theoretical and practical efficiency gap, LAUDNet merges algorithmic design with scheduling optimization, guided by a latency predictor that accurately gauges dynamic operator latency. We've tested LAUDNet across multiple vision tasks, demonstrating its capacity to notably reduce the latency of models like ResNet-101 by over 50% on platforms such as V100, RTX3090, and TX2 GPUs. Notably, LAUDNet stands out in balancing accuracy and efficiency. Code is available at: https://www.github.com/LeapLabTHU/LAUDNet.

Zhihang Yuan, Lin Niu, Jiawei Liu et al.

Large-scale language models (LLMs) have demonstrated impressive performance, but their deployment presents challenges due to their significant memory usage. This issue can be alleviated through quantization. In this paper, we identify that the challenge in quantizing activations in LLMs arises from varying ranges across channels, rather than solely the presence of outliers. To address this challenge, we introduce a quantization method called RPTQ, which utilizes a reorder-based approach. By rearranging the channels and quantizing them in clusters, RPTQ effectively mitigates the impact of range differences between channels. To minimize the overhead of the reorder operation, we fuse it into the layer norm operation and weights in linear layers. In our experiments, RPTQ achieved a significant breakthrough by utilizing 3-bit activation in LLMs for the first time, resulting in a substantial reduction in memory usage. For instance, quantizing OPT-175b can lead to a memory consumption reduction of up to 80%.

Zhihang Yuan, Jiawei Liu, Jiaxiang Wu et al.

Post-training quantization (PTQ) is a popular method for compressing deep neural networks (DNNs) without modifying their original architecture or training procedures. Despite its effectiveness and convenience, the reliability of PTQ methods in the presence of some extrem cases such as distribution shift and data noise remains largely unexplored. This paper first investigates this problem on various commonly-used PTQ methods. We aim to answer several research questions related to the influence of calibration set distribution variations, calibration paradigm selection, and data augmentation or sampling strategies on PTQ reliability. A systematic evaluation process is conducted across a wide range of tasks and commonly-used PTQ paradigms. The results show that most existing PTQ methods are not reliable enough in term of the worst-case group performance, highlighting the need for more robust methods. Our findings provide insights for developing PTQ methods that can effectively handle distribution shift scenarios and enable the deployment of quantized DNNs in real-world applications.

Luning Wang, Shiyao Li, Xuefei Ning et al.

Large Language Models (LLMs) have been widely adopted to process long-context tasks. However, the large memory overhead of the key-value (KV) cache poses significant challenges in long-context scenarios. Existing training-free KV cache compression methods typically focus on quantization and token pruning, which have compression limits, and excessive sparsity can lead to severe performance degradation. Other methods design new architectures with less KV overhead but require significant training overhead. To address the above two drawbacks, we further explore the redundancy in the channel dimension and apply an architecture-level design with minor training costs. Therefore, we introduce CSKV, a training-efficient Channel Shrinking technique for KV cache compression: (1) We first analyze the singular value distribution of the KV cache, revealing significant redundancy and compression potential along the channel dimension. Based on this observation, we propose using low-rank decomposition for key and value layers and storing the low-dimension features. (2) To preserve model performance, we introduce a bi-branch KV cache, including a window-based full-precision KV cache and a low-precision compressed KV cache. (3) To reduce the training costs, we minimize the layer-wise reconstruction loss for the compressed KV cache instead of retraining the entire LLMs. Extensive experiments show that CSKV can reduce the memory overhead of the KV cache by 80% while maintaining the model's long-context capability. Moreover, we show that our method can be seamlessly combined with quantization to further reduce the memory overhead, achieving a compression ratio of up to 95%. Code is available at https://github.com/wln20/CSKV.

Lin Niu, Jiawei Liu, Zhihang Yuan et al.

Efficient inference for object detection networks is a major challenge on edge devices. Post-Training Quantization (PTQ), which transforms a full-precision model into low bit-width directly, is an effective and convenient approach to reduce model inference complexity. But it suffers severe accuracy drop when applied to complex tasks such as object detection. PTQ optimizes the quantization parameters by different metrics to minimize the perturbation of quantization. The p-norm distance of feature maps before and after quantization, Lp, is widely used as the metric to evaluate perturbation. For the specialty of object detection network, we observe that the parameter p in Lp metric will significantly influence its quantization performance. We indicate that using a fixed hyper-parameter p does not achieve optimal quantization performance. To mitigate this problem, we propose a framework, DetPTQ, to assign different p values for quantizing different layers using an Object Detection Output Loss (ODOL), which represents the task loss of object detection. DetPTQ employs the ODOL-based adaptive Lp metric to select the optimal quantization parameters. Experiments show that our DetPTQ outperforms the state-of-the-art PTQ methods by a significant margin on both 2D and 3D object detectors. For example, we achieve 31.1/31.7(quantization/full-precision) mAP on RetinaNet-ResNet18 with 4-bit weight and 4-bit activation.

Rundong Su, Jintao Zhang, Zhihang Yuan et al. · tsinghua

Diffusion transformers have demonstrated remarkable capabilities in generating videos. However, their practical deployment is severely constrained by high memory usage and computational cost. Post-Training Quantization provides a practical way to reduce memory usage and boost computation speed. Existing quantization methods typically apply a static bit-width allocation, overlooking the quantization difficulty of activations across diffusion timesteps, leading to a suboptimal trade-off between efficiency and quality. In this paper, we propose a inference time NVFP4/INT8 Mixed-Precision Quantization framework. We find a strong linear correlation between a block's input-output difference and the quantization sensitivity of its internal linear layers. Based on this insight, we design a lightweight predictor that dynamically allocates NVFP4 to temporally stable layers to maximize memory compression, while selectively preserving INT8 for volatile layers to ensure robustness. This adaptive precision strategy enables aggressive quantization without compromising generation quality. Beside this, we observe that the residual between the input and output of a Transformer block exhibits high temporal consistency across timesteps. Leveraging this temporal redundancy, we introduce Temporal Delta Cache (TDC) to skip computations for these invariant blocks, further reducing the computational cost. Extensive experiments demonstrate that our method achieves 1.92$\times$ end-to-end acceleration and 3.32$\times$ memory reduction, setting a new baseline for efficient inference in Video DiTs.

Zhihang Yuan, Yuzhang Shang, Yang Zhou et al.

The field of efficient Large Language Model (LLM) inference is rapidly evolving, presenting a unique blend of opportunities and challenges. Although the field has expanded and is vibrant, there hasn't been a concise framework that analyzes the various methods of LLM Inference to provide a clear understanding of this domain. Our survey stands out from traditional literature reviews by not only summarizing the current state of research but also by introducing a framework based on roofline model for systematic analysis of LLM inference techniques. This framework identifies the bottlenecks when deploying LLMs on hardware devices and provides a clear understanding of practical problems, such as why LLMs are memory-bound, how much memory and computation they need, and how to choose the right hardware. We systematically collate the latest advancements in efficient LLM inference, covering crucial areas such as model compression (e.g., Knowledge Distillation and Quantization), algorithm improvements (e.g., Early Exit and Mixture-of-Expert), and both hardware and system-level enhancements. Our survey stands out by analyzing these methods with roofline model, helping us understand their impact on memory access and computation. This distinctive approach not only showcases the current research landscape but also delivers valuable insights for practical implementation, positioning our work as an indispensable resource for researchers new to the field as well as for those seeking to deepen their understanding of efficient LLM deployment. The analyze tool, LLM-Viewer, is open-sourced.

Xing Hu, Yuan Cheng, Dawei Yang et al.

Post-training quantization (PTQ) has emerged as a widely adopted technique for compressing and accelerating Large Language Models (LLMs). The major challenge in LLM quantization is that uneven and heavy-tailed data distributions can expand the quantization range, thereby reducing bit precision for most values. Recent methods attempt to eliminate outliers and balance inter-channel differences by employing linear transformations; however, they remain heuristic and are often overlook optimizing the data distribution across the entire quantization space.In this paper, we introduce Quantization Space Utilization Rate (QSUR), a novel metric that effectively assesses the quantizability of transformed data by measuring the space utilization of the data in the quantization space. We complement QSUR with mathematical derivations that examine the effects and limitations of various transformations, guiding our development of Orthogonal and Scaling Transformation-based Quantization (OSTQuant). OSQuant employs a learnable equivalent transformation, consisting of an orthogonal transformation and a scaling transformation, to optimize the distributions of weights and activations across the entire quantization space. Futhermore, we propose the KL-Top loss function, designed to mitigate noise during optimization while retaining richer semantic information within the limited calibration data imposed by PTQ. OSTQuant outperforms existing work on various LLMs and benchmarks. In the W4-only setting, it retains 99.5\% of the floating-point accuracy. In the more challenging W4A4KV4 configuration, OSTQuant reduces the performance gap by 32\% on the LLaMA-3-8B model compared to state-of-the-art methods. \href{https://github.com/BrotherHappy/OSTQuant}{https://github.com/BrotherHappy/OSTQuant}.

Dawei Yang, Ning He, Xing Hu et al.

Although neural networks have made remarkable advancements in various applications, they require substantial computational and memory resources. Network quantization is a powerful technique to compress neural networks, allowing for more efficient and scalable AI deployments. Recently, Re-parameterization has emerged as a promising technique to enhance model performance while simultaneously alleviating the computational burden in various computer vision tasks. However, the accuracy drops significantly when applying quantization on the re-parameterized networks. We identify that the primary challenge arises from the large variation in weight distribution across the original branches. To address this issue, we propose a coarse & fine weight splitting (CFWS) method to reduce quantization error of weight, and develop an improved KL metric to determine optimal quantization scales for activation. To the best of our knowledge, our approach is the first work that enables post-training quantization applicable on re-parameterized networks. For example, the quantized RepVGG-A1 model exhibits a mere 0.3% accuracy loss. The code is in https://github.com/NeonHo/Coarse-Fine-Weight-Split.git

Chuanhao Sun, Zhihang Yuan, Kai Xu et al.

Fourier features based positional encoding (PE) is commonly used in machine learning tasks that involve learning high-frequency features from low-dimensional inputs, such as 3D view synthesis and time series regression with neural tangent kernels. Despite their effectiveness, existing PEs require manual, empirical adjustment of crucial hyperparameters, specifically the Fourier features, tailored to each unique task. Further, PEs face challenges in efficiently learning high-frequency functions, particularly in tasks with limited data. In this paper, we introduce sinusoidal PE (SPE), designed to efficiently learn adaptive frequency features closely aligned with the true underlying function. Our experiments demonstrate that SPE, without hyperparameter tuning, consistently achieves enhanced fidelity and faster training across various tasks, including 3D view synthesis, Text-to-Speech generation, and 1D regression. SPE is implemented as a direct replacement for existing PEs. Its plug-and-play nature lets numerous tasks easily adopt and benefit from SPE.

Tianyu Fu, Yi Ge, Yichen You et al. · tsinghua

Large Language Models (LLMs) achieve impressive reasoning capabilities at the cost of substantial inference overhead, posing substantial deployment challenges. Although distilled Small Language Models (SLMs) significantly enhance efficiency, their performance suffers as they fail to follow LLMs' reasoning paths. Luckily, we reveal that only a small fraction of tokens genuinely diverge reasoning paths between LLMs and SLMs. Most generated tokens are either identical or exhibit neutral differences, such as minor variations in abbreviations or expressions. Leveraging this insight, we introduce **Roads to Rome (R2R)**, a neural token routing method that selectively utilizes LLMs only for these critical, path-divergent tokens, while leaving the majority of token generation to the SLM. We also develop an automatic data generation pipeline that identifies divergent tokens and generates token-level routing labels to train the lightweight router. We apply R2R to combine R1-1.5B and R1-32B models from the DeepSeek family, and evaluate on challenging math, coding, and QA benchmarks. With an average activated parameter size of 5.6B, R2R surpasses the average accuracy of R1-7B by 1.6x, outperforming even the R1-14B model. Compared to R1-32B, it delivers a 2.8x wall-clock speedup with comparable performance, advancing the Pareto frontier of test-time scaling efficiency. Our code is available at https://github.com/thu-nics/R2R.

Haojie Duanmu, Xiuhong Li, Zhihang Yuan et al.

Mixture-of-Experts (MoE) models face deployment challenges due to their large parameter counts and computational demands. We explore quantization for MoE models and highlight two key insights: 1) linear blocks exhibit varying quantization sensitivity, and 2) divergent expert activation frequencies create heterogeneous computational characteristics. Based on these observations, we introduce MxMoE, a mixed-precision optimization framework for MoE models that considers both algorithmic and system perspectives. MxMoE navigates the design space defined by parameter sensitivity, expert activation dynamics, and hardware resources to derive efficient mixed-precision configurations. Additionally, MxMoE automatically generates optimized mixed-precision GroupGEMM kernels, enabling parallel execution of GEMMs with different precisions. Evaluations show that MxMoE outperforms existing methods, achieving 2.4 lower Wikitext-2 perplexity than GPTQ at 2.25-bit and delivering up to 3.4x speedup over full precision, as well as up to 29.4% speedup over uniform quantization at equivalent accuracy with 5-bit weight-activation quantization. Our code is available at https://github.com/cat538/MxMoE.

JiangYong Yu, Sifan Zhou, Dawei Yang et al.

Multimodal large language models (MLLMs) have garnered widespread attention due to their ability to understand multimodal input. However, their large parameter sizes and substantial computational demands severely hinder their practical deployment and application.While quantization is an effective way to reduce model size and inference latency, its application to MLLMs remains underexplored. In this paper, we propose MQuant, a post-training quantization (PTQ) framework designed to tackle the unique challenges of multimodal large language models (MLLMs). Conventional quantization often struggles with MLLMs because of (a) high inference latency from large visual token counts, (b) distributional disparities between visual and textual tokens, and (c) extreme outliers introduced by Hadamard-based transformations. To address these issues, MQuant introduces: Modality-Specific Static Quantization (MSQ), assigning distinct static scales for visual vs. textual tokens; Attention-Invariant Flexible Switching (AIFS), reordering tokens to preserve casual attention while eliminating expensive token-wise scale computations; Rotation Magnitude Suppression (RMS), mitigating weight outliers arising from online Hadamard rotations. On five mainstream MLLMs (including Qwen-VL, MiniCPM-V, CogVLM2), MQuant under W4A8 achieves near-floating-point accuracy (<1% degradation) while reducing inference latency by up to 30%, significantly outperforming existing PTQ baselines. Our MQuant effectively bridges the gap for efficient and accurate MLLMs inference in resource-constrained devices. Code has been released in https://github.com/StiphyJay/MQuant.

Yuanpeng Zhang, Xing Hu, Xi Chen et al.

SRAM Processing-in-Memory (PIM) has emerged as the most promising implementation for high-performance PIM, delivering superior computing density, energy efficiency, and computational precision. However, the pursuit of higher performance necessitates more complex circuit designs and increased operating frequencies, which exacerbate IR-drop issues. Severe IR-drop can significantly degrade chip performance and even threaten reliability. Conventional circuit-level IR-drop mitigation methods, such as back-end optimizations, are resource-intensive and often compromise power, performance, and area (PPA). To address these challenges, we propose AIM, comprehensive software and hardware co-design for architecture-level IR-drop mitigation in high-performance PIM. Initially, leveraging the bit-serial and in-situ dataflow processing properties of PIM, we introduce Rtog and HR, which establish a direct correlation between PIM workloads and IR-drop. Building on this foundation, we propose LHR and WDS, enabling extensive exploration of architecture-level IR-drop mitigation while maintaining computational accuracy through software optimization. Subsequently, we develop IR-Booster, a dynamic adjustment mechanism that integrates software-level HR information with hardware-based IR-drop monitoring to adapt the V-f pairs of the PIM macro, achieving enhanced energy efficiency and performance. Finally, we propose the HR-aware task mapping method, bridging software and hardware designs to achieve optimal improvement. Post-layout simulation results on a 7nm 256-TOPS PIM chip demonstrate that AIM achieves up to 69.2% IR-drop mitigation, resulting in 2.29x energy efficiency improvement and 1.152x speedup.

Zhihang Yuan, Chengyu Yue, Long Huang et al.

Instruction tuning is a standard paradigm for adapting large language models (LLMs), but modern instruction datasets are large, noisy, and redundant, making full-data fine-tuning costly and often unnecessary. Existing data selection methods either build expensive gradient datastores or assign static scores from a weak proxy, largely ignoring evolving uncertainty, and thus missing a key source of LLM interpretability. We propose GRADFILTERING, an objective-agnostic, uncertainty-aware data selection framework that utilizes a small GPT-2 proxy with a LoRA ensemble and aggregates per-example gradients into a Gradient Signal-to-Noise Ratio (G-SNR) utility. Our method matches or surpasses random subsets and strong baselines in most LLM-as-a-judge evaluations as well as in human assessment. Moreover, GRADFILTERING-selected subsets converge faster than competitive filters under the same compute budget, reflecting the benefit of uncertainty-aware scoring.

Zixuan Zhou, Xuefei Ning, Ke Hong et al. · tsinghua

Large Language Models (LLMs) have attracted extensive attention due to their remarkable performance across various tasks. However, the substantial computational and memory requirements of LLM inference pose challenges for deployment in resource-constrained scenarios. Efforts within the field have been directed towards developing techniques aimed at enhancing the efficiency of LLM inference. This paper presents a comprehensive survey of the existing literature on efficient LLM inference. We start by analyzing the primary causes of the inefficient LLM inference, i.e., the large model size, the quadratic-complexity attention operation, and the auto-regressive decoding approach. Then, we introduce a comprehensive taxonomy that organizes the current literature into data-level, model-level, and system-level optimization. Moreover, the paper includes comparative experiments on representative methods within critical sub-fields to provide quantitative insights. Last but not least, we provide some knowledge summary and discuss future research directions.

Chen Zhu, Wangbo Zhao, Huiwen Zhang et al.

Vision Transformers (ViTs) have emerged as a foundational model in computer vision, excelling in generalization and adaptation to downstream tasks. However, deploying ViTs to support diverse resource constraints typically requires retraining multiple, size-specific ViTs, which is both time-consuming and energy-intensive. To address this issue, we propose an efficient ViT adaptation framework that enables a single adaptation process to generate multiple models of varying sizes for deployment on platforms with various resource constraints. Our approach comprises two stages. In the first stage, we enhance a pre-trained ViT with a nested elastic architecture that enables structural flexibility across MLP expansion ratio, number of attention heads, embedding dimension, and network depth. To preserve pre-trained knowledge and ensure stable adaptation, we adopt a curriculum-based training strategy that progressively increases elasticity. In the second stage, we design a lightweight router to select submodels according to computational budgets and downstream task demands. Initialized with Pareto-optimal configurations derived via a customized NSGA-II algorithm, the router is then jointly optimized with the backbone. Extensive experiments on multiple benchmarks demonstrate the effectiveness and versatility of EA-ViT. The code is available at https://github.com/zcxcf/EA-ViT.

Zhihang Yuan, Leyang Xue, Waleed Ahsan et al.

Traffic forecasting based network operation optimization and management offers enormous promise but also presents significant challenges from traffic forecasting perspective. While deep learning models have proven to be relatively more effective than traditional statistical methods for time series forecasting, their reliability is not satisfactory due to their inability to effectively handle unique characteristics of network traffic. In particular, the burst and complex traffic patterns makes the existing models less reliable, as each type of deep learning model has limited capability in capturing traffic patterns. To address this issue, we introduce TUBO, a novel machine learning framework custom designed for reliable network traffic forecasting. TUBO features two key components: burst processing for handling significant traffic fluctuations and model selection for adapting to varying traffic patterns using a pool of models. A standout feature of TUBO is its ability to provide deterministic predictions along with quantified uncertainty, which serves as a cue for identifying the most reliable forecasts. Evaluations on three real-world network demand matrix (DM) datasets (Abilene, GEANT, and CERNET) show that TUBO significantly outperforms existing methods on forecasting accuracy (by 4 times), and also achieves up to 94% accuracy in burst occurrence forecasting. Furthermore, we also consider traffic demand forecasting based proactive traffic engineering (TE) as a downstream use case. Our results show that compared to reactive approaches and proactive TE using the best existing DM forecasting methods, proactive TE powered by TUBO improves aggregated throughput by 9 times and 3 times, respectively.

Yuxuan Yue, Zhihang Yuan, Haojie Duanmu et al.

Large Language Models (LLMs) face significant deployment challenges due to their substantial memory requirements and the computational demands of auto-regressive text generation process. This paper addresses these challenges by focusing on the quantization of LLMs, a technique that reduces memory consumption by converting model parameters and activations into low-bit integers. We critically analyze the existing quantization approaches, identifying their limitations in balancing the accuracy and efficiency of the quantized LLMs. To advance beyond these limitations, we propose WKVQuant, a PTQ framework especially designed for quantizing weights and the key/value (KV) cache of LLMs. Specifically, we incorporates past-only quantization to improve the computation of attention. Additionally, we introduce two-dimensional quantization strategy to handle the distribution of KV cache, along with a cross-block reconstruction regularization for parameter optimization. Experiments show that WKVQuant achieves almost comparable memory savings to weight-activation quantization, while also approaching the performance of weight-only quantization.

Haoxuan Wang, Yuzhang Shang, Zhihang Yuan et al.

The practical deployment of diffusion models is still hindered by the high memory and computational overhead. Although quantization paves a way for model compression and acceleration, existing methods face challenges in achieving low-bit quantization efficiently. In this paper, we identify imbalanced activation distributions as a primary source of quantization difficulty, and propose to adjust these distributions through weight finetuning to be more quantization-friendly. We provide both theoretical and empirical evidence supporting finetuning as a practical and reliable solution. Building on this approach, we further distinguish two critical types of quantized layers: those responsible for retaining essential temporal information and those particularly sensitive to bit-width reduction. By selectively finetuning these layers under both local and global supervision, we mitigate performance degradation while enhancing quantization efficiency. Our method demonstrates its efficacy across three high-resolution image generation tasks, obtaining state-of-the-art performance across multiple bit-width settings.

Haojie Duanmu, Zhihang Yuan, Xiuhong Li et al.

Large language models (LLMs) can now handle longer sequences of tokens, enabling complex tasks like book understanding and generating lengthy novels. However, the key-value (KV) cache required for LLMs consumes substantial memory as context length increasing, becoming the bottleneck for deployment. In this paper, we present a strategy called SKVQ, which stands for sliding-window KV cache quantization, to address the issue of extremely low bitwidth KV cache quantization. To achieve this, SKVQ rearranges the channels of the KV cache in order to improve the similarity of channels in quantization groups, and applies clipped dynamic quantization at the group level. Additionally, SKVQ ensures that the most recent window tokens in the KV cache are preserved with high precision. This helps maintain the accuracy of a small but important portion of the KV cache.SKVQ achieves high compression ratios while maintaining accuracy. Our evaluation on LLMs demonstrates that SKVQ surpasses previous quantization approaches, allowing for quantization of the KV cache to 2-bit keys and 1.5-bit values with minimal loss of accuracy. With SKVQ, it is possible to process context lengths of up to 1M on an 80GB memory GPU for a 7b model and up to 7 times faster decoding.

Yuzhang Shang, Zhihang Yuan, Yan Yan

Dataset distillation (DD) aims to synthesize a small dataset whose test performance is comparable to a full dataset using the same model. State-of-the-art (SoTA) methods optimize synthetic datasets primarily by matching heuristic indicators extracted from two networks: one from real data and one from synthetic data (see Fig.1, Left), such as gradients and training trajectories. DD is essentially a compression problem that emphasizes maximizing the preservation of information contained in the data. We argue that well-defined metrics which measure the amount of shared information between variables in information theory are necessary for success measurement but are never considered by previous works. Thus, we introduce mutual information (MI) as the metric to quantify the shared information between the synthetic and the real datasets, and devise MIM4DD numerically maximizing the MI via a newly designed optimizable objective within a contrastive learning framework to update the synthetic dataset. Specifically, we designate the samples in different datasets that share the same labels as positive pairs and vice versa negative pairs. Then we respectively pull and push those samples in positive and negative pairs into contrastive space via minimizing NCE loss. As a result, the targeted MI can be transformed into a lower bound represented by feature maps of samples, which is numerically feasible. Experiment results show that MIM4DD can be implemented as an add-on module to existing SoTA DD methods.

Zukang Xu, Yuxuan Yue, Xing Hu et al.

Mamba is an efficient sequence model that rivals Transformers and demonstrates significant potential as a foundational architecture for various tasks. Quantization is commonly used in neural networks to reduce model size and computational latency. However, applying quantization to Mamba remains underexplored, and existing quantization methods, which have been effective for CNN and Transformer models, appear inadequate for Mamba models (e.g., Quarot suffers a 21% accuracy drop on Vim-T$^\dagger$ even under W8A8). We have pioneered the exploration of this issue and identified several key challenges. First, significant outliers are present in gate projections, output projections, and matrix multiplications. Second, Mamba's unique parallel scan further amplifies these outliers, leading to uneven and heavy-tailed data distributions. Third, even with the application of the Hadamard transform, the variance across channels in weights and activations still remains inconsistent. To these ends, we propose MambaQuant, a post-training quantization (PTQ) framework consisting of: 1) Karhunen-Loeve Transformation (KLT) enhanced rotation, rendering the rotation matrix adaptable to diverse channel distributions. 2) Smooth-Fused rotation, which equalizes channel variances and can merge additional parameters into model weights. Experiments show that MambaQuant can quantize both weights and activations into 8-bit with less than 1% accuracy loss for Mamba-based vision and language tasks. To the best of our knowledge, MambaQuant is the first comprehensive PTQ design for the Mamba family, paving the way for further advancements in its application.

Sifan Zhou, Shuo Wang, Zhihang Yuan et al.

Large Language Models (LLMs) fine-tuning technologies have achieved remarkable results. However, traditional LLM fine-tuning approaches face significant challenges: they require large Floating Point (FP) computation, raising privacy concerns when handling sensitive data, and are impractical for resource-constrained edge devices. While Parameter-Efficient Fine-Tuning (PEFT) techniques reduce trainable parameters, their reliance on floating-point arithmetic creates fundamental incompatibilities with edge hardware. In this work, we introduce a novel framework for on-device LLM fine-tuning that eliminates the need for floating-point operations in both inference and training, named GSQ-Tuning. At its core is the Group-Shared Exponents Integer format, which efficiently represents model parameters in integer format using shared exponents among parameter groups. When combined with LoRA-like adapters, this enables fully integer-based fine-tuning that is both memory and compute efficient. We demonstrate that our approach achieves accuracy comparable to BF16-based fine-tuning while significantly reducing 1.85x memory usage. Moreover, compared to FP8, our method can reduce 5x power consumption and 11x chip area with same performance, making large-scale model adaptation feasible on edge devices.

Xing Hu, Zhixuan Chen, Dawei Yang et al.

Mixture-of-Experts (MoE) large language models (LLMs), which leverage dynamic routing and sparse activation to enhance efficiency and scalability, have achieved higher performance while reducing computational costs. However, these models face significant memory overheads, limiting their practical deployment and broader adoption. Post-training quantization (PTQ), a widely used method for compressing LLMs, encounters severe accuracy degradation and diminished generalization performance when applied to MoE models. This paper investigates the impact of MoE's sparse and dynamic characteristics on quantization and identifies two primary challenges: (1) Inter-expert imbalance, referring to the uneven distribution of samples across experts, which leads to insufficient and biased calibration for less frequently utilized experts; (2) Intra-expert imbalance, arising from MoE's unique aggregation mechanism, which leads to varying degrees of correlation between different samples and their assigned experts. To address these challenges, we propose MoEQuant, a novel quantization framework tailored for MoE LLMs. MoE-Quant includes two novel techniques: 1) Expert-Balanced Self-Sampling (EBSS) is an efficient sampling method that efficiently constructs a calibration set with balanced expert distributions by leveraging the cumulative probabilities of tokens and expert balance metrics as guiding factors. 2) Affinity-Guided Quantization (AGQ), which incorporates affinities between experts and samples into the quantization process, thereby accurately assessing the impact of individual samples on different experts within the MoE layer. Experiments demonstrate that MoEQuant achieves substantial performance gains (more than 10 points accuracy gain in the HumanEval for DeepSeekMoE-16B under 4-bit quantization) and boosts efficiency.

Hanling Zhang, Rundong Su, Zhihang Yuan et al.

Text-to-image generation models, especially Multimodal Diffusion Transformers (MMDiT), have shown remarkable progress in generating high-quality images. However, these models often face significant computational bottlenecks, particularly in attention mechanisms, which hinder their scalability and efficiency. In this paper, we introduce DiTFastAttnV2, a post-training compression method designed to accelerate attention in MMDiT. Through an in-depth analysis of MMDiT's attention patterns, we identify key differences from prior DiT-based methods and propose head-wise arrow attention and caching mechanisms to dynamically adjust attention heads, effectively bridging this gap. We also design an Efficient Fused Kernel for further acceleration. By leveraging local metric methods and optimization techniques, our approach significantly reduces the search time for optimal compression schemes to just minutes while maintaining generation quality. Furthermore, with the customized kernel, DiTFastAttnV2 achieves a 68% reduction in attention FLOPs and 1.5x end-to-end speedup on 2K image generation without compromising visual fidelity.

Zhaofeng Hu, Sifan Zhou, Zhihang Yuan et al.

3D single object tracking is essential in autonomous driving and robotics. Existing methods often struggle with sparse and incomplete point cloud scenarios. To address these limitations, we propose a Multimodal-guided Virtual Cues Projection (MVCP) scheme that generates virtual cues to enrich sparse point clouds. Additionally, we introduce an enhanced tracker MVCTrack based on the generated virtual cues. Specifically, the MVCP scheme seamlessly integrates RGB sensors into LiDAR-based systems, leveraging a set of 2D detections to create dense 3D virtual cues that significantly improve the sparsity of point clouds. These virtual cues can naturally integrate with existing LiDAR-based 3D trackers, yielding substantial performance gains. Extensive experiments demonstrate that our method achieves competitive performance on the NuScenes dataset.

Rui Xie, Tianchen Zhao, Zhihang Yuan et al. · tsinghua

Visual Autoregressive (VAR) has emerged as a promising approach in image generation, offering competitive potential and performance comparable to diffusion-based models. However, current AR-based visual generation models require substantial computational resources, limiting their applicability on resource-constrained devices. To address this issue, we conducted analysis and identified significant redundancy in three dimensions of the VAR model: (1) the attention map, (2) the attention outputs when using classifier free guidance, and (3) the data precision. Correspondingly, we proposed efficient attention mechanism and low-bit quantization method to enhance the efficiency of VAR models while maintaining performance. With negligible performance lost (less than 0.056 FID increase), we could achieve 85.2% reduction in attention computation, 50% reduction in overall memory and 1.5x latency reduction. To ensure deployment feasibility, we developed efficient training-free compression techniques and analyze the deployment feasibility and efficiency gain of each technique.

Zhihang Yuan, Zhiyuan Liu, Mahesh K. Marina

Multivariate time series forecasting involves two qualitatively distinct factors: (i) stable within-series autoregressive (AR) dynamics, and (ii) intermittent cross-dimension interactions that can become spurious over long horizons. We argue that fitting a single model to capture both effects creates an optimization conflict: the high-variance updates needed for cross-dimension modeling can corrupt the gradients that support autoregression, resulting in brittle training and degraded long-horizon accuracy. To address this, we propose ALTTS, a dual-path framework that explicitly decouples autoregression and cross-relation (CR) modeling. In ALTTS, the AR path is instantiated with a linear predictor, while the CR path uses a Transformer equipped with Cross-Relation Self-Attention (CRSA); the two branches are coordinated via alternating optimization to isolate gradient noise and reduce cross-block interference. Extensive experiments on multiple benchmarks show that ALTTS consistently outperforms prior methods, with the most pronounced improvements on long-horizon forecasting. Overall, our results suggest that carefully designed optimization strategies, rather than ever more complex architectures, can be a key driver of progress in multivariate time series forecasting.

Songsheng Wang, Rucheng Yu, Zhihang Yuan et al.

Vision-Language-Action (VLA) models have made substantial progress by leveraging the robust capabilities of Visual Language Models (VLMs). However, VLMs' significant parameter size and autoregressive (AR) decoding nature impose considerable computational demands on VLA models. While Speculative Decoding (SD) has shown efficacy in accelerating Large Language Models (LLMs) by incorporating efficient drafting and parallel verification, allowing multiple tokens to be generated in one forward pass, its application to VLA models remains unexplored. This work introduces Spec-VLA, an SD framework designed to accelerate VLA models. Due to the difficulty of the action prediction task and the greedy decoding mechanism of the VLA models, the direct application of the advanced SD framework to the VLA prediction task yields a minor speed improvement. To boost the generation speed, we propose an effective mechanism to relax acceptance utilizing the relative distances represented by the action tokens of the VLA model. Empirical results across diverse test scenarios affirm the effectiveness of the Spec-VLA framework, and further analysis substantiates the impact of our proposed strategies, which enhance the acceptance length by 44%, achieving 1.42 times speedup compared with the OpenVLA baseline, without compromising the success rate. The success of the Spec-VLA framework highlights the potential for broader application of speculative execution in VLA prediction scenarios.

Zhihang Yuan, Yuzhang Shang, Hanling Zhang et al.

Recent advances in autoregressive (AR) models with continuous tokens for image generation show promising results by eliminating the need for discrete tokenization. However, these models face efficiency challenges due to their sequential token generation nature and reliance on computationally intensive diffusion-based sampling. We present ECAR (Efficient Continuous Auto-Regressive Image Generation via Multistage Modeling), an approach that addresses these limitations through two intertwined innovations: (1) a stage-wise continuous token generation strategy that reduces computational complexity and provides progressively refined token maps as hierarchical conditions, and (2) a multistage flow-based distribution modeling method that transforms only partial-denoised distributions at each stage comparing to complete denoising in normal diffusion models. Holistically, ECAR operates by generating tokens at increasing resolutions while simultaneously denoising the image at each stage. This design not only reduces token-to-image transformation cost by a factor of the stage number but also enables parallel processing at the token level. Our approach not only enhances computational efficiency but also aligns naturally with image generation principles by operating in continuous token space and following a hierarchical generation process from coarse to fine details. Experimental results demonstrate that ECAR achieves comparable image quality to DiT Peebles & Xie [2023] while requiring 10$\times$ FLOPs reduction and 5$\times$ speedup to generate a 256$\times$256 image.

Ang Li, Zhihang Yuan, Yang Zhang et al.

Reinforcement Learning with Verifiable Feedback (RLVF) has become a key technique for enhancing the reasoning abilities of Large Language Models (LLMs). However, its reliance on sparse, outcome based rewards, which only indicate if a final answer is correct or not, fails to provide granular guidance on the reasoning process itself. This limitation hinders efficient learning, as the model cannot distinguish between high quality and inefficient solutions, nor can it learn effectively from different types of failures. To address this, we observe that an LLMs self-certainty often correlates with task difficulty and solution quality. We introduce Difficulty Aware Certainty guided Exploration (DACE), a novel RL algorithm that leverages this insight to dynamically balance the exploration exploitation trade-off. DACE assesses task difficulty online based on the policys success rate. It then uses this signal to modulate an intrinsic reward: for difficult tasks where the model is struggling, DACE encourages exploration by penalizing high certainty; for easier tasks, it encourages learning efficiency by rewarding high certainty. Experiments on challenging mathematical reasoning benchmarks (AIME, MATH) show that DACE significantly outperforms strong baselines. The DACE-trained models not only achieve higher accuracy but also demonstrate more robust performance when scaling test-time compute, validating that our adaptive approach fosters effective exploration without sacrificing precision.

Chen Xu, Yuxuan Yue, Zukang Xu et al.

RWKV is a modern RNN architecture with comparable performance to Transformer, but still faces challenges when deployed to resource-constrained devices. Post Training Quantization (PTQ), which is a an essential technique to reduce model size and inference latency, has been widely used in Transformer models. However, it suffers significant degradation of performance when applied to RWKV. This paper investigates and identifies two key constraints inherent in the properties of RWKV: (1) Non-linear operators hinder the parameter-fusion of both smooth- and rotation-based quantization, introducing extra computation overhead. (2) The larger amount of uniformly distributed weights poses challenges for cluster-based quantization, leading to reduced accuracy. To this end, we propose RWKVQuant, a PTQ framework tailored for RWKV models, consisting of two novel techniques: (1) a coarse-to-fine proxy capable of adaptively selecting different quantization approaches by assessing the uniformity and identifying outliers in the weights, and (2) a codebook optimization algorithm that enhances the performance of cluster-based quantization methods for element-wise multiplication in RWKV. Experiments show that RWKVQuant can quantize RWKV-6-14B into about 3-bit with less than 1% accuracy loss and 2.14x speed up.

Yi Guo, Wei Wang, Zhihang Yuan et al.

Generative models like Flow Matching have achieved state-of-the-art performance but are often hindered by a computationally expensive iterative sampling process. To address this, recent work has focused on few-step or one-step generation by learning the average velocity field, which directly maps noise to data. MeanFlow, a leading method in this area, learns this field by enforcing a differential identity that connects the average and instantaneous velocities. In this work, we argue that this differential formulation is a limiting special case of a more fundamental principle. We return to the first principles of average velocity and leverage the additivity property of definite integrals. This leads us to derive a novel, purely algebraic identity we term Interval Splitting Consistency. This identity establishes a self-referential relationship for the average velocity field across different time intervals without resorting to any differential operators. Based on this principle, we introduce SplitMeanFlow, a new training framework that enforces this algebraic consistency directly as a learning objective. We formally prove that the differential identity at the core of MeanFlow is recovered by taking the limit of our algebraic consistency as the interval split becomes infinitesimal. This establishes SplitMeanFlow as a direct and more general foundation for learning average velocity fields. From a practical standpoint, our algebraic approach is significantly more efficient, as it eliminates the need for JVP computations, resulting in simpler implementation, more stable training, and broader hardware compatibility. One-step and two-step SplitMeanFlow models have been successfully deployed in large-scale speech synthesis products (such as Doubao), achieving speedups of 20x.

Zhenglin Lai, Mengyao Liao, Bingzhe Wu et al.

Large language models with Mixture-of-Experts (MoE) architectures achieve efficiency and scalability, yet their routing mechanisms introduce safety alignment challenges insufficiently addressed by techniques developed for dense models. In this work, the MoE-specific safety risk of positional vulnerability-that safety-aligned behaviors rely on specific expert modules-is formalized and systematically analyzed. An analytical framework, SAFEx, is presented to robustly identify, characterize, and validate safety-critical experts via a stability-based expert selection procedure, and to decompose them into two functional groups: the Harmful Content Detection Group (HCDG), which specializes in identifying and recognizing harmful content within user inputs, and the Harmful Response Control Group (HRCG), which specializes in controlling and enforcing model behaviors to generate appropriate safety responses. Expert-level interventions are conducted to probe causality and to test mitigation. Targeted masking of SAFEx-selected experts reveals that safety behavior is highly concentrated. On Qwen3-30B-A3B, configured with 48 MoE-FFN layers and 128 experts per layer under top-8 routing (48x128=6,144 experts in total), disabling 12 selected experts reduces the refusal rate by 22%. In addition, lightweight adaptation is performed using LoRA under three configurations-the HRCG, the union of HCDG and HRCG, and all experts-and the resulting updates are composed through negative weight merging targeted at the HRCG, leading to improved refusal under adversarial prompts without full-model retraining. These results establish positional vulnerability as a distinct MoE-specific safety challenge and provide a practical, compute-efficient pathway for expert-level safety interventions within routed architectures.

Yuxuan Yue, Zukang Xu, Zhihang Yuan et al.

Large Language Models (LLMs) face significant challenges in edge deployment due to their massive parameter scale. Vector Quantization (VQ), a clustering-based quantization method, serves as a prevalent solution to this issue for its extremely low-bit (even at 2-bit) and considerable accuracy. Since a vector is a quantity in mathematics and physics that has both direction and magnitude, existing VQ works typically quantize them in a coupled manner. However, we find that direction exhibits significantly greater sensitivity to quantization compared to the magnitude. For instance, when separately clustering the directions and magnitudes of weight vectors in LLaMA-2-7B, the accuracy drop of zero-shot tasks are 46.5\% and 2.3\%, respectively. This gap even increases with the reduction of clustering centers. Further, Euclidean distance, a common metric to access vector similarities in current VQ works, places greater emphasis on reducing the magnitude error. This property is contrary to the above finding, unavoidably leading to larger quantization errors. To these ends, this paper proposes Polar Coordinate Decoupled Vector Quantization (PCDVQ), an effective and efficient VQ framework consisting of two key modules: 1) Polar Coordinate Decoupling (PCD), which transforms vectors into their polar coordinate representations and perform independent quantization of the direction and magnitude parameters.2) Distribution Aligned Codebook Construction (DACC), which optimizes the direction and magnitude codebooks in accordance with the source distribution. Experimental results show that PCDVQ outperforms baseline methods at 2-bit level by at least 1.5\% zero-shot accuracy, establishing a novel paradigm for accurate and highly compressed LLMs.

Zhihang Yuan, Siyuan Wang, Rui Xie et al.

In this paper, we propose the Dynamic Latent Frame Rate VAE (DLFR-VAE), a training-free paradigm that can make use of adaptive temporal compression in latent space. While existing video generative models apply fixed compression rates via pretrained VAE, we observe that real-world video content exhibits substantial temporal non-uniformity, with high-motion segments containing more information than static scenes. Based on this insight, DLFR-VAE dynamically adjusts the latent frame rate according to the content complexity. Specifically, DLFR-VAE comprises two core innovations: (1) A Dynamic Latent Frame Rate Scheduler that partitions videos into temporal chunks and adaptively determines optimal frame rates based on information-theoretic content complexity, and (2) A training-free adaptation mechanism that transforms pretrained VAE architectures into a dynamic VAE that can process features with variable frame rates. Our simple but effective DLFR-VAE can function as a plug-and-play module, seamlessly integrating with existing video generation models and accelerating the video generation process.

Weisheng Xu, Sifan Zhou, Jiaqi Xiong et al.

LiDAR-based 3D single object tracking (3D SOT) is a critical issue in robotics and autonomous driving. Existing 3D SOT methods typically adhere to a point-based processing pipeline, wherein the re-sampling operation invariably leads to either redundant or missing information, thereby impacting performance. To address these issues, we propose PillarTrack, a novel pillar-based 3D SOT framework. First, we transform sparse point clouds into dense pillars to preserve the local and global geometrics. Second, we propose a Pyramid-Encoded Pillar Feature Encoder (PE-PFE) design to enhance the robustness of pillar feature for translation/rotation/scale. Third, we present an efficient Transformer-based backbone from the perspective of modality differences. Finally, we construct our PillarTrack based on above designs. Extensive experiments show that our method achieves comparable performance on the KITTI and NuScenes datasets, significantly enhancing the performance of the baseline.

Yushi Huang, Zining Wang, Zhihang Yuan et al.

Mixture-of-Experts (MoE) Multimodal large language models (MLLMs) excel at vision-language tasks, but they suffer from high computational inefficiency. To reduce inference overhead, expert skipping methods have been proposed to deactivate redundant experts based on the current input tokens. However, we find that applying these methods-originally designed for unimodal large language models (LLMs)-to MLLMs results in considerable performance degradation. This is primarily because such methods fail to account for the heterogeneous contributions of experts across MoE layers and modality-specific behaviors of tokens within these layers. Motivated by these findings, we propose MoDES, the first training-free framework that adaptively skips experts to enable efficient and accurate MoE MLLM inference. It incorporates a globally-modulated local gating (GMLG) mechanism that integrates global layer-wise importance into local routing probabilities to accurately estimate per-token expert importance. A dual-modality thresholding (DMT) method is then applied, which processes tokens from each modality separately, to derive the skipping schedule. To set the optimal thresholds, we introduce a frontier search algorithm that exploits monotonicity properties, cutting convergence time from several days to a few hours. Extensive experiments for 3 model series across 13 benchmarks demonstrate that MoDES far outperforms previous approaches. For instance, when skipping 88% experts for Qwen3-VL-MoE-30B-A3B-Instruct, the performance boost is up to 10.67% (97.33% vs. 86.66%). Furthermore, MoDES significantly enhances inference speed, improving the prefilling time by 2.16$\times$ and the decoding time by 1.26$\times$.

Shaoyuan Chen, Zhixuan Chen, Dawei Yang et al.

Large Language Models (LLMs) fine-tuning techniques not only improve the adaptability to diverse downstream tasks, but also mitigate adverse effects of model quantization. Despite this, conventional quantization suffers from its structural limitation that hinders flexibility during the fine-tuning and deployment stages. Practical on-device tasks demand different quantization precisions (i.e. different bit-widths), e.g., understanding tasks tend to exhibit higher tolerance to reduced precision compared to generation tasks. Conventional quantization, typically relying on scaling factors that are incompatible across bit-widths, fails to support the on-device switching of precisions when confronted with complex real-world scenarios. To overcome the dilemma, we propose OTARo, a novel method that enables on-device LLMs to flexibly switch quantization precisions while maintaining performance robustness through once fine-tuning. OTARo introduces Shared Exponent Floating Point (SEFP), a distinct quantization mechanism, to produce different bit-widths through simple mantissa truncations of a single model. Moreover, to achieve bit-width robustness in downstream applications, OTARo performs a learning process toward losses induced by different bit-widths. The method involves two critical strategies: (1) Exploitation-Exploration Bit-Width Path Search (BPS), which iteratively updates the search path via a designed scoring mechanism; (2) Low-Precision Asynchronous Accumulation (LAA), which performs asynchronous gradient accumulations and delayed updates under low bit-widths. Experiments on popular LLMs, e.g., LLaMA3.2-1B, LLaMA3-8B, demonstrate that OTARo achieves consistently strong and robust performance for all precisions.

Mengzhao Chen, Meng Wu, Hui Jin et al.

Modern AI hardware, such as Nvidia's Blackwell architecture, is increasingly embracing low-precision floating-point (FP) formats to handle the pervasive activation outliers in Large Language Models (LLMs). Despite this industry trend, a unified comparison of FP and integer (INT) quantization across varying granularities has been missing, leaving algorithm and hardware co-design without clear guidance. This paper fills that gap by systematically investigating the trade-offs between FP and INT formats. We reveal a critical performance crossover: while FP excels in coarse-grained quantization, the comparison at fine-grained (block-wise) levels is more nuanced. Our comprehensive comparison demonstrates that for popular 8-bit fine-grained formats (e.g., MX with block size 32), MXINT8 is superior to its FP counterpart in both algorithmic accuracy and hardware efficiency. However, for 4-bit formats, FP (e.g., MXFP4, NVFP4) often holds an accuracy advantage , though we show that NVINT4 can surpass NVFP4 when outlier-mitigation techniques like Hadamard rotation are applied. We also introduce a symmetric clipping method that resolves gradient bias in fine-grained low-bit INT training, enabling nearly lossless performance for MXINT8 training. These findings challenge the current hardware trajectory, demonstrating that a one-size-fits-all FP approach is suboptimal and advocating that fine-grained INT formats, particularly MXINT8, offer a better balance of accuracy, power, and efficiency for future AI accelerators.

Ang Li, Yifei Wang, Zhihang Yuan et al.

Reinforcement learning in large language models (LLMs) often relies on scalar rewards, a practice that discards valuable textual rationale buried in the rollouts, forcing the model to explore \textit{de novo} with each attempt and hindering sample efficiency. While LLMs can uniquely learn from language feedback provided in-context, naively integrating on-line experiences into RL training presents a paradox: feedback from the same problem risks information leakage and memorization, while feedback from different problems often leads to behavior collapse due to irrelevant context. To resolve this tension, we propose \textbf{Language-And-Numerical Policy Optimization (LANPO)}, a framework that cleanly separates the roles of feedback: language guides exploration, while numerical rewards drive optimization. LANPO builds a dynamic experience pool from past trials and introduces two principles to ensure feedback is effective: \emph{Reward-Agnostic Reflection} for safe intra-sample self-correction and \emph{Relevant Abstraction} to distill generalizable lessons from inter-sample experiences. Across mathematical reasoning benchmarks, LANPO enables 7B and 14B models to significantly outperform strong baselines trained with GRPO in test accuracy. Our work provides a robust method for integrating historical experiences into the LLM RL loop, creating more effective and data-efficient learning agents.

Hanling Zhang, Yayu Zhou, Tongcheng Fang et al.

Small Language Models (SLMs) provide computational advantages in resource-constrained environments, yet memory limitations remain a critical bottleneck for edge device deployment. A substantial portion of SLMs' memory footprint stems from vocabulary-related components, particularly embeddings and language modeling (LM) heads, due to large vocabulary sizes. Existing static vocabulary pruning, while reducing memory usage, suffers from rigid, one-size-fits-all designs that cause information loss from the prefill stage and a lack of flexibility. In this work, we identify two key principles underlying the vocabulary reduction challenge: the lexical locality principle, the observation that only a small subset of tokens is required during any single inference, and the asymmetry in computational characteristics between vocabulary-related components of SLM. Based on these insights, we introduce VocabTailor, a novel decoupled dynamic vocabulary selection framework that addresses memory constraints through offloading embedding and implements a hybrid static-dynamic vocabulary selection strategy for LM Head, enabling on-demand loading of vocabulary components. Comprehensive experiments across diverse downstream tasks demonstrate that VocabTailor achieves a reduction of up to 99% in the memory usage of vocabulary-related components with minimal or no degradation in task performance, substantially outperforming existing static vocabulary pruning.

Zhihang Yuan, Rui Xie, Yuzhang Shang et al.

Diffusion Transformer(DiT)-based generation models have achieved remarkable success in video generation. However, their inherent computational demands pose significant efficiency challenges. In this paper, we exploit the inherent temporal non-uniformity of real-world videos and observe that videos exhibit dynamic information density, with high-motion segments demanding greater detail preservation than static scenes. Inspired by this temporal non-uniformity, we propose VGDFR, a training-free approach for Diffusion-based Video Generation with Dynamic Latent Frame Rate. VGDFR adaptively adjusts the number of elements in latent space based on the motion frequency of the latent space content, using fewer tokens for low-frequency segments while preserving detail in high-frequency segments. Specifically, our key contributions are: (1) A dynamic frame rate scheduler for DiT video generation that adaptively assigns frame rates for video segments. (2) A novel latent-space frame merging method to align latent representations with their denoised counterparts before merging those redundant in low-resolution space. (3) A preference analysis of Rotary Positional Embeddings (RoPE) across DiT layers, informing a tailored RoPE strategy optimized for semantic and local information capture. Experiments show that VGDFR can achieve a speedup up to 3x for video generation with minimal quality degradation.

Zhihang Yuan, Hanling Zhang, Pu Lu et al.

Diffusion Transformers (DiT) excel at image and video generation but face computational challenges due to the quadratic complexity of self-attention operators. We propose DiTFastAttn, a post-training compression method to alleviate the computational bottleneck of DiT. We identify three key redundancies in the attention computation during DiT inference: (1) spatial redundancy, where many attention heads focus on local information; (2) temporal redundancy, with high similarity between the attention outputs of neighboring steps; (3) conditional redundancy, where conditional and unconditional inferences exhibit significant similarity. We propose three techniques to reduce these redundancies: (1) Window Attention with Residual Sharing to reduce spatial redundancy; (2) Attention Sharing across Timesteps to exploit the similarity between steps; (3) Attention Sharing across CFG to skip redundant computations during conditional generation. We apply DiTFastAttn to DiT, PixArt-Sigma for image generation tasks, and OpenSora for video generation tasks. Our results show that for image generation, our method reduces up to 76% of the attention FLOPs and achieves up to 1.8x end-to-end speedup at high-resolution (2k x 2k) generation.

Zhihang Yuan, Yuzhang Shang, Yue Song et al.

In this paper, we introduce a new post-training compression paradigm for Large Language Models (LLMs) to facilitate their wider adoption. We delve into LLM weight low-rank decomposition, and find that the challenges of this task stem from (1) the distribution variance in the LLM activations and (2) the sensitivity difference among various kinds of layers. To address these issues, we propose a training-free approach called Activation-aware Singular Value Decomposition (ASVD). Specifically, ASVD manages activation outliers by transforming the weight matrix based on the activation distribution. This transformation allows the outliers in the activation matrix to be absorbed into the transformed weight matrix, thereby enhancing decomposition accuracy. Additionally, we propose an efficient iterative calibration process to optimize layer-specific decomposition by addressing the varying sensitivity of different LLM layers. In this way, ASVD can compress a network by 10%-30%. Based on the success of the low-rank decomposition of projection matrices in the self-attention module, we further introduce ASVD to compress the KV cache. By reducing the channel dimension of KV activations, memory requirements for KV cache can be largely reduced. ASVD can further achieve 50% KV cache reductions without performance drop in a training-free manner.