Large Vision-Language Models (VLMs) exhibit impressive multi-modal capabilities but suffer from prohibitive computational and memory demands, due to their long visual token sequences and massive parameter sizes. To address these issues, recent works have proposed training-free compression methods. However, existing efforts often suffer from three major limitations: (1) Current approaches do not decompose techniques into comparable modules, hindering fair evaluation across spatial and temporal redundancy. (2) Evaluation confined to simple single-turn tasks, failing to reflect performance in realistic scenarios. (3) Isolated use of individual compression techniques, without exploring their joint potential. To overcome these gaps, we introduce LLMC+, a comprehensive VLM compression benchmark with a versatile, plug-and-play toolkit. LLMC+ supports over 20 algorithms across five representative VLM families and enables systematic study of token-level and model-level compression. Our benchmark reveals that: (1) Spatial and temporal redundancies demand distinct technical strategies. (2) Token reduction methods degrade significantly in multi-turn dialogue and detail-sensitive tasks. (3) Combining token and model compression achieves extreme compression with minimal performance loss. We believe LLMC+ will facilitate fair evaluation and inspire future research in efficient VLM.

Lingkun Long, Yushi Huang, Shihao Bai, Ruihao Gong, Jun Zhang, Ao Zhou, Jianlei Yang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

LLM inference demands substantial GPU resources, but its highly variable workload characteristics create efficiency challenges at both inter-GPU and intra-GPU levels. To meet Service Level Objectives (SLOs), existing systems typically overprovision resources in two ways: allocating excess GPUs to handle peak loads and reserving excessive memory per request to prevent out-of-memory during token generation. We introduce PiLLM (Predictable inference for LLMs), a system that addresses these inefficiencies through accurate workload prediction and dynamic resource allocation.

Tiny machine learning (TinyML) has attracted heightened attention for its ability to provide low-cost and instantaneous performance on edge devices. Particularly, the commonly used microcontroller unit (MCU) imposes extreme constraints on peak memory (SRAM) and storage (Flash). Existing TinyML methods often rely on a customized and hard-to-obtain inference libraries, as well as necessitate a time-consuming search for a deployable architecture using advanced Neural Architecture Search (NAS) algorithms. To solve these problems, we fully exploit the resources on MCU and deduce hardware-oriented guidelines for designing models under extreme MCU constraints. In detail, we delve into thorough information about the atom operators by collecting the runtime data of Flash, SRAM, and latency to build a dataset named AtomDB. Based on AtomDB, several critical operator guidelines are established to fully utilize limited Flash and SRAM, while minimizing latency. By transferring the guidelines to analyze blocks, we propose a hybrid pattern that organizes appropriate blocks at different network stages to form the AtomNet, a more hardware-oriented architecture, to handle the former SRAM bottleneck and the latter Flash bottleneck. Extensive experiments demonstrate the effectiveness of the exploitation of the hardware characteristics. Remarkably, AtomNet pioneeringly achieve 3.5% accuracy enhancement and more than 15% latency reduction on 320KB MCU using readily available official inference libraries for ImageNet tasks, surpassing the current state-of-the-art method.

Junyi Chen, Shihao Bai, Zaijun Wang, Siyu Wu, Chuheng Du, Hailong Yang, Ruihao Gong, Shengzhong Liu, Fan Wu, Guihai Chen. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

The exploration and application of Large Language Models (LLMs) is thriving. To reduce deployment costs, continuous batching has become an essential feature in current service frameworks. The effectiveness of continuous batching relies on an accurate estimate of the memory requirements of requests. However, due to the diversity in request output lengths, existing frameworks tend to adopt aggressive or conservative schedulers, which often result in significant overestimation or underestimation of memory consumption. Consequently, they suffer from harmful request evictions or prolonged queuing times, failing to achieve satisfactory throughput under strict Service Level Agreement (SLA) guarantees (a.k.a. goodput), across various LLM application scenarios with differing input-output length distributions. To address this issue, we propose a novel Past-Future scheduler that precisely estimates the peak memory resources required by the running batch via considering the historical distribution of request output lengths and calculating memory occupancy at each future time point. It adapts to applications with all types of input-output length distributions, balancing the trade-off between request queuing and harmful evictions, thereby consistently achieving better goodput. Furthermore, to validate the effectiveness of the proposed scheduler, we developed a high-performance LLM serving framework, LightLLM, that implements the Past-Future scheduler. Compared to existing aggressive or conservative schedulers, LightLLM demonstrates superior goodput, achieving up to 2-3× higher goodput than other schedulers under heavy loads. LightLLM is open source to boost the research in such direction (https://github.com/ModelTC/lightllm).

The rapid advancement of large language models (LLMs) has paved the way for their use in solving real-world problems, which in turn has significantly driven the development of tool-assisted LLMs. This progress necessitates thorough evaluation methods. However, existing benchmarks typically only provide end-to-end scores but lack in-depth analysis and often suffer from issues such as instability. To address this gap, we have meticulously designed the Tool Playgrounds framework, a comprehensive, analyzable, and extensible benchmark. This framework evaluates boundary dimensions such as parameter missing interaction, parameter correction, tool failover, and leveraging internal knowledge. Our findings indicate that even the most advanced commercial models frequently overlook these essential aspects and face challenges in managing complex tool usage. To foster further research and development, we have made our code, dataset, and leaderboard publicly available on https://github.com/zhiwei-dong/ToolPlaygrounds.

Although knowledge distillation (KD) is an effective approach to improve the performance of a smaller LLM (i.e., the student model) by transferring knowledge from a large LLM (i.e., the teacher model), it still suffers from high training cost. Existing LLM distillation methods ignore the difficulty difference among different samples, making the distillation of easy samples unnecessary. This leads to high distillation cost. In this paper, we propose difficulty-aware knowledge distillation (DA-KD) framework for efficient knowledge distillation, in which we dynamically adjust the distillation dataset based on the difficulty of samples. We further observe existing KD loss cannot perform well when most of samples are difficult in the distillation dataset because of unstable optimization and the neglect of hard samples. Therefore, we also propose a new KD loss called bidirectional discrepancy loss (BDL) for effective KD. Extensive experiments demonstrate that our DA-KD framework is effective and efficient. Without bells and whistles, DA-KD can outperform existing state-of-the-art KD methods by 2% with half training cost and even surpass the teacher model with 4.7$\times$ compression.

Diffusion Transformers (DiTs) excel in generative tasks but face practical deployment challenges due to high inference costs. Feature caching, which stores and retrieves redundant computations, offers the potential for acceleration. Existing learning-based caching, though adaptive, overlooks the impact of the prior timestep. It also suffers from misaligned objectives-*aligned predicted noise vs. high-quality images*-between training and inference. These two discrepancies compromise both performance and efficiency. To this end, we *harmonize* training and inference with a novel learning-based *caching* framework dubbed **HarmoniCa**. It first incorporates *Step-Wise Denoising Training* (SDT) to ensure the continuity of the denoising process, where prior steps can be leveraged. In addition, an *Image Error Proxy-Guided Objective* (IEPO) is applied to balance image quality against cache utilization through an efficient proxy to approximate the image error. Extensive experiments across $8$ models, $4$ samplers, and resolutions from $256\times256$ to $2K$ demonstrate superior performance and speedup of our framework. For instance, it achieves over $40\\%$ latency reduction (*i.e.*, $2.07\times$ theoretical speedup) and improved performance on PixArt-$\alpha$. Remarkably, our *image-free* approach reduces training time by $25\\%$ compared with the previous method. Our code is available at https://github.com/ModelTC/HarmoniCa.

Vision-language instruction-tuning models have recently achieved significant performance improvements. In this work, we discover that large-scale 3D parallel training on those models leads to an imbalanced computation load across different devices. The vision and language parts are inherently heterogeneous: their data distribution and model architecture differ significantly, which affects distributed training efficiency. To address this issue, we rebalance the computational load from data, model, and memory perspectives, achieving more balanced computation across devices. Specifically, for the data, instances are grouped into new balanced mini-batches within and across devices. A search-based method is employed for the model to achieve a more balanced partitioning. For memory optimization, we adaptively adjust the re-computation strategy for each partition to utilize the available memory fully. These three perspectives are not independent but are closely connected, forming an omniverse balanced training framework. Extensive experiments are conducted to validate the effectiveness of our method. Compared with the open-source training code of InternVL-Chat, training time is reduced greatly, achieving about 1.8$\times$ speed-up. Our method’s efficacy and generalizability are further validated across various models and datasets. Codes will be released at https://github.com/ModelTC/OmniBal.

Training Long-Context Large Language Models (LLMs) is challenging, as hybrid training with long-context and short-context data often leads to workload imbalances. Existing works mainly use data packing to alleviate this issue, but fail to consider imbalanced attention computation and wasted communication overhead. This paper proposes Hierarchical Balance Packing (HBP), which designs a novel batch-construction method and training recipe to address those inefficiencies. In particular, the HBP constructs multi-level data packing groups, each optimized with a distinct packing length. It assigns training samples to their optimal groups and configures each group with the most effective settings, including sequential parallelism degree and gradient checkpointing configuration. To effectively utilize multi-level groups of data, we design a dynamic training pipeline specifically tailored to HBP, including curriculum learning, adaptive sequential parallelism, and stable loss. Our extensive experiments demonstrate that our method significantly reduces training time over multiple datasets and open-source models while maintaining strong performance. For the largest DeepSeek-V2 (236B) MoE model, our method speeds up the training by 2.4$\times$ with competitive performance. Codes will be released at https://github.com/ModelTC/HBP.

Large language models (LLMs) have gained considerable attention for their remarkable performance across a wide range of tasks. However, efficient LLM inference remains challenging because of the autoregressive decoding process, which generates only one token at a time. Speculative decoding has been introduced to address the limitation by using small speculative models (SSMs) to speed up LLM inference. However, the low acceptance rate of SSMs and the high verification cost of LLM prohibit further performance improvement. In this paper, we present Smurfs, an LLM inference system designed to accelerate LLM inference through collective and adaptive speculative decoding. Smurfs adopts a majority-voted mechanism that harnesses multiple SSMs to collaboratively predict LLM outputs in multi-task scenarios, while avoiding high verification cost. It also decouples SSM speculation from LLM verification and uses a pipelined execution to hide the latency of SSM speculation. Additionally, Smurfs proposes a mechanism to dynamically determine the optimal speculation length of SSM at runtime, balancing the performance impact of accepted tokens and verification cost. The experimental results demonstrate the superiority of Smurfs in terms of inference throughput and latency compared to the state-of-the-art LLM inference systems.

Recently, post-training quantization (PTQ) has become the de facto way to produce efficient low-precision neural networks without long-time retraining. Despite its low cost, current PTQ works fail to succeed under the extremely low-bit setting. In this work, we delve into extremely low-bit quantization and construct a unified theoretical analysis, which provides an in-depth understanding of the reason for the failure of low-bit quantization. According to the theoretical study, we argue that the existing methods fail in low-bit schemes due to significant perturbation on weights and lack of consideration of activation quantization. To this end, we propose Brecq and QDrop to respectively solve these two challenges, based on which a Q-Limit framework is constructed. Then the Q-Limit framework is further extended to support a mixed precision quantization scheme. To the best of our knowledge, this is the first work that can push the limit of PTQ down to INT2. Extensive experiments on various handcrafted and searched neural architectures are conducted for both visual recognition/detection tasks and language processing tasks. Without bells and whistles, our PTQ framework can attain low-bit ResNet and MobileNetV2 comparable with quantization-aware training (QAT), establishing a new state-of-the-art for PTQ.

Diffusion models, widely used for image generation, face significant challenges related to their broad applicability due to prolonged inference times and high memory demands. Efficient Post-Training Quantization (PTQ) is crucial to address these issues. However, unlike traditional models, diffusion models critically rely on the time-step for the multi-round denoising. Typically, each time-step is encoded into a hypersensitive temporal feature by several modules. Despite this, existing PTQ methods do not optimize these modules individually. Instead, they employ unsuitable reconstruction objectives and complex calibration methods, leading to significant disturbances in the temporal feature and denoising trajectory, as well as reduced compression efficiency. To address these challenges, we introduce a novel quantization framework that includes three strategies: 1) TIB-based Maintenance: Based on our innovative Temporal Information Block (TIB) definition, Temporal Information-aware Reconstruction (TIAR) and Finite Set Calibration (FSC) are developed to efficiently align original temporal features. 2) Cache-based Maintenance: Instead of indirect and complex optimization for the related modules, pre-computing and caching quantized counterparts of temporal features are developed to minimize errors. 3) Disturbance-aware Selection: Employ temporal feature errors to guide a fine-grained selection between the two maintenance strategies for further disturbance reduction. This framework preserves most of the temporal information and ensures high-quality end-to-end generation. Extensive testing on various datasets, diffusion models and hardware confirms our superior performance and acceleration.

The resurgence of H5Nx reassortment has caused multiple epidemics resulting in severe disease even death in wild birds and poultry. Assessing H5Nx reassortment risk is crucial for designing targeted interventions and enhancing preparedness efforts to manage H5Nx outbreaks effectively. However, the complexity in H5Nx reassortment, driven by the diversity of influenza A viruses (IAVs) and wide range of hosts, has hindered the effective quantification of reassortment risk. In this study, we utilized a network approach to explore the reassortment history using a large-scale dataset. By inferring genomic homogeneity among IAVs, we constructed an IAVs homologous network with reassortment history embedded within it. We estimated the communities within the IAVs homologous network to represent the reassortment risk of various viruses, revealing diverse reassortment risks across different H5Nx viruses. Our analysis also identified the primary hosts contributing to reassortment: domestic poultry in China, and wild birds in North America and Europe. These primary hosts are critical targets for future H5Nx reassortment interventions. Our study provides a framework for quantifying and ranking H5Nx reassortment risk, contributing to enhanced preparedness and prevention efforts.

Lane detection (LD) plays a crucial role in enhancing the L2+ capabilities of autonomous driving, capturing widespread attention. The Post-Processing Quantization (PTQ) could facilitate the practical application of LD models, enabling fast speeds and limited memories without labeled data. However, prior PTQ methods do not consider the complex LD outputs that contain physical semantics, such as offsets, locations, etc., and thus cannot be directly applied to LD models. In this paper, we pioneeringly investigate semantic sensitivity to post-processing for lane detection with a novel Lane Distortion Score. Moreover, we identify two main factors impacting the LD performance after quantization, namely intra-head sensitivity and inter-head sensitivity, where a small quantization error in specific semantics can cause significant lane distortion. Thus, we propose a Selective Focus framework deployed with Semantic Guided Focus and Sensitivity Aware Selection modules, to incorporate post-processing information into PTQ reconstruction. Based on the observed intra-head sensitivity, Semantic Guided Focus is introduced to prioritize foreground-related semantics using a practical proxy. For inter-head sensitivity, we present Sensitivity Aware Selection, efficiently recognizing influential prediction heads and refining the optimization objectives at runtime. Extensive experiments have been done on a wide variety of models including keypoint-, anchor-, curve-, and segmentation-based ones. Our method produces quantized models in minutes on a single GPU and can achieve 6.4\% F1 Score improvement on the CULane dataset. Code and supplementary statement can be found at https://github.com/PannenetsF/SelectiveFocus.

Neural network sparsity has attracted many research interests due to its similarity to biological schemes and high energy efficiency. However, existing methods depend on long-time training or fine-tuning, which prevents large-scale applications. Recently, some works focusing on post-training sparsity (PTS) have emerged. They get rid of the high training cost but usually suffer from distinct accuracy degradation due to neglect of the reasonable sparsity rate at each layer. Previous methods for finding sparsity rates mainly focus on the training-aware scenario, which usually fails to converge stably under the PTS setting with limited data and much less training cost. In this paper, we propose a fast and controllable post-training sparsity (FCPTS) framework. By incorporating a differentiable bridge function and a controllable optimization objective, our method allows for rapid and accurate sparsity allocation learning in minutes, with the added assurance of convergence to a predetermined global sparsity rate. Equipped with these techniques, we can surpass the state-of-the-art methods by a large margin, e.g., over 30\% improvement for ResNet-50 on ImageNet under the sparsity rate of 80\%. Our plug-and-play code and supplementary materials are open-sourced at https://github.com/ModelTC/FCPTS.

The Diffusion model, a prevalent framework for image generation, encounters significant challenges in terms of broad applicability due to its extended inference times and substantial memory requirements. Efficient Post-training Quantization (PTQ) is pivotal for addressing these issues in traditional models. Different from traditional models, diffusion models heavily depend on the time-step t to achieve satisfactory multi-round denoising. Usually, t from the finite set {1, …, T} is encoded to a temporal feature by a few modules totally irrespective of the sampling data. However, existing PTQ methods do not optimize these modules separately. They adopt inappropriate reconstruction targets and complex calibration methods, resulting in a severe disturbance of the temporal feature and denoising trajectory, as well as a low compression efficiency. To solve these, we propose a Temporal Feature Maintenance Quantization (TFMQ) framework building upon a Temporal Information Block which is just related to the time-step t and unrelated to the sampling data. Powered by the pioneering block design, we devise temporal information aware reconstruction (TIAR) and finite set calibration (FSC) to align the full-precision temporal features in a limited time. Equipped with the framework, we can maintain the most temporal information and ensure the end-to-end generation quality. Extensive experiments on various datasets and diffusion models prove our state-of-the-art results. Remarkably, our quantization approach, for the first time, achieves model performance nearly on par with the full-precision model under 4-bit weight quantization. Additionally, our method incurs almost no extra computational cost and accelerates quan-tization time by 2.0× on LSUN-Bedrooms 256 × 256 compared to previous works. Our code is publicly available at https://github.com/ModelTC/TFMQ-DM.

Recent advancements in generative large language models (LLMs) have significantly boosted the performance in natural language processing tasks. However, their efficiency is constrained by the inherent limitations in autoregressive token generation. While parallel decoding with token tree verification, e.g., Medusa, has been proposed to improve decoding parallelism and efficiency, it often struggles with maintaining contextual relationships due to its independent token prediction approach and incurs significant verification overhead, especially with large tree sizes and batch processing. In this paper, we propose ProPD, an efficient LLM parallel decoding framework based on dynamic token tree pruning and generation. ProPD features an advanced early pruning mechanism to efficiently eliminate unpromising token sequences to improve verification efficiency. Additionally, it introduces a dynamic token tree generation algorithm to balance the computation and parallelism of the verification phase in real-time and maximize the overall efficiency across different batch sizes, sequence lengths, and tasks, etc. We verify ProPD across a diverse set of datasets, LLMs, and batch sizes and demonstrate ProPD consistently outperforms existing decoding algorithms by 1.1--3.2 ×.

Large Language Models (LLMs) have demonstrated unparalleled efficacy in natural language processing. However, their high computational demands and memory overheads hinder their broad deployment. To address this, two quantization strategies emerge, including Quantization-Aware Training (QAT) and Post-Training Quantization (PTQ). For LLMs, the billions of parameters make the QAT impractical due to the prohibitive training cost and thus PTQ becomes more prevalent. In existing studies, activation outliers in particular channels are identified as the biggest challenge to PTQ accuracy. They propose to transform the magnitudes from activations to weights, which however offers limited alleviation or suffers from unstable gradients, resulting in a severe performance drop at low-bitwidth. In this paper, we propose QLLM, an accurate and efficient low-bitwidth PTQ method designed for LLMs. QLLM introduces an adaptive channel reassembly technique that reallocates the magnitude of outliers to other channels, thereby mitigating their impact on the quantization range. This is achieved by channel disassembly and channel assembly, which first breaks down the outlier channels into several sub-channels to ensure a more balanced distribution of activation magnitudes. Then similar channels are merged to maintain the original channel number for efficiency. Additionally, an adaptive strategy is designed to autonomously determine the optimal number of sub-channels for channel disassembly. To further compensate for the performance loss caused by quantization, we propose an efficient tuning method that only learns a small number of low-rank weights while freezing the pre-trained quantized model. After training, these low-rank parameters can be fused into the frozen weights without affecting inference. Extensive experiments on LLaMA-1 and LLaMA-2 show that QLLM is able to obtain accurate quantized models efficiently. For example, QLLM quantizes the 4-bit LLaMA-2-70B within 10 hours on a single A100-80G GPU, outperforming the previous state-of-the-art method by 7.89% on the average accuracy across five zero-shot tasks. Code is available at [ZIP Lab](https://github.com/ziplab/QLLM) and [ModelTC](https://github.com/ModelTC/QLLM).

In this paper, we introduce a novel model compression technique named Joint Sparsification and Quantization (JSQ), explicitly tailored for large language models (LLMs). Traditional methods employ either sparsification or quantization individually to compress LLMs, leading to performance degradation at high compression ratios. In contrast, our JSQ approach integrates sparsification and quantization cohesively. As sparsification tend to preserve outliers that is harmful to quantization, we introduce a novel sparsity metric to serves as a bridge between the sparsification and quantization. Moreover, it is proven outliers in LLMs have significant impact but harmful to compression. Current solutions are highly coupled with quantization process, which is not helpful to sparsification. To this end, we also introduce a search-based activation editor to automatically eliminate relatively useless outliers. Comprehensive experiments across various datasets and architectures affirm the efficacy of our JSQ framework. Notably, our JSQ achieves 7.96$\times$ computation reduction without crashing for the representative model LLaMA. This accomplishment stands in stark contrast to the limitations of most state-of-the-art LLM compression methods, which typically fail under such extreme compression ratios. Our code is released at https://github.com/uanu2002/JSQ.

The increasing diversity of deep neural network (DNN) models and hardware platforms necessitates effective model profiling for high-performance inference deployment. Current DNN profiling tools suffer from either limited optimization insights due to the missing correlation between high-level DNN layer design and low-level hardware performance metrics, or prohibitive profiling overhead due to the large amount of performance measurement through hardware performance counters. Meanwhile, the roofline model has been widely used in the high-performance computing (HPC) domain for identifying performance bottlenecks and guiding optimizations. However, it lacks hierarchical (e.g., kernel/operator/layer), fine-grained, multi-platform support for profiling DNN models.

With the increased attention to model efficiency, post-training sparsity (PTS) has become more and more prevalent because of its effectiveness and efficiency. However, there remain questions on better practice of PTS algorithms and the sparsification ability of models, which hinders the further development of this area.Therefore, a benchmark to comprehensively investigate the issues above is urgently needed. In this paper, we propose the first comprehensive post-training sparsity benchmark called PTSBench towards algorithms and models. We benchmark 10+ PTS general-pluggable fine-grained techniques on 3 typical tasks using over 40 off-the-shelf model architectures. Through extensive experiments and analyses, we obtain valuable conclusions and provide several insights from both algorithms and model aspects. Our PTSBench can provide (1) new observations for a better understanding of the PTS algorithms, (2) in-depth and comprehensive evaluations for the sparsification ability of models, and (3) a well-structured and easy-integrate open-source framework. We hope this work will provide illuminating conclusions and advice for future studies of post-training sparsity methods and sparsification-friendly model design. The code for our PTSBench is released at https://github.com/ModelTC/msbench.

Adversarial training has been demonstrated to be one of the most effective remedies for defending adversarial examples, yet it often suffers from the huge robustness generalization gap on unseen testing adversaries, deemed as the adversarially robust generalization problem. Despite the preliminary understandings devoted to adversarially robust generalization, little is known from the architectural perspective. To bridge the gap, this paper for the first time systematically investigated the relationship between adversarially robust generalization and architectural design. In particular, we comprehensively evaluated 20 most representative adversarially trained architectures on ImageNette and CIFAR-10 datasets towards multiple$\ell_{p}$-norm adversarial attacks. Based on the extensive experiments, we found that, under aligned settings, Vision Transformers (e.g., PVT, CoAtNet) often yield better adversarially robust generalization while CNNs tend to overfit on specific attacks and fail to generalize on multiple adversaries. To better understand the nature behind it, we conduct theoretical analysis via the lens of Rademacher complexity. We revealed the fact that the higher weight sparsity contributes significantly towards the better adversarially robust generalization of Transformers, which can be often achieved by the specially-designed attention blocks. We hope our paper could help to better understand the mechanism for designing robust DNNs. Our model weights can be found at http://robust.art.

Open world object detection (OWOD) has attracted extensive attention due to its practicability in the real world. Previous OWOD works manually designed unknown-discover strategies to select unknown proposals from the background, suffering from uncertainties without appropriate priors. In this paper, we claim the learning of object detection could be seen as an object-level feature-entanglement process, where unknown traits are propagated to the known proposals through convolutional operations and could be distilled to benefit unknown recognition without manual selection. Therefore, we propose a simple yet effective Annealing-based Label-Transfer framework, which sufficiently explores the known proposals to alleviate the uncertainties. Specifically, a Label-Transfer Learning paradigm is introduced to decouple the known and unknown features, while a Sawtooth Annealing Scheduling strategy is further employed to rebuild the decision boundaries of the known and unknown classes, thus promoting both known and unknown recognition. Moreover, previous OWOD works neglected the trade-off of known and unknown performance, and we thus introduce a metric called Equilibrium Index to comprehensively evaluate the effectiveness of the OWOD models. To the best of our knowledge, this is the first OWOD work without manual unknown selection. Extensive experiments conducted on the common-used benchmark validate that our model achieves superior detection performance (200% unknown mAP improvement with the even higher known detection performance) compared to other state-of-the-art methods. Our code is available at https://github.com/DIG-Beihang/ALLOW.git.

Post-training quantization (PTQ) of transformer language models faces significant challenges due to the existence of detrimental outliers in activations.We observe that these outliers are concentrated in specific channels and are asymmetric across channels.To address this issue, we propose the Outlier Suppression+ (OS+) framework, which contains the channel-wise shifting for asymmetry and channel-wise scaling for concentration.We show that these operations can be seamlessly migrated into subsequent modules while maintaining equivalence.Second, we propose a fast and stable scheme to calculate effective shifting and scaling values.The channel-wise shifting aligns the center of each channel for removal of outlier asymmetry.The channel-wise scaling quantitatively evaluates changes brought by migration and quantization for better quantization burden balance.We validate our OS+ under both standard and fine-grained quantization settings with models including BERT, OPT, BLOOM, BLOOMZ, and LLaMA.Comprehensive results across various tasks demonstrate the superiority of our approach.Especially, with standard quantization, OS+ can achieve near-floating-point performance on both small models and large language models on 8-bit and 6-bit.Besides, we establish a new state-of-the-art for 4-bit BERT with 15.5% improvement.Our code is available at https://github.com/ModelTC/ Outlier_Suppression_Plus.

Deep neural networks have delivered remarkable performance and have been widely used in various visual tasks. However, their huge sizes cause significant inconvenience for transmission and storage. Many previous studies have explored model size compression. However, these studies often approach various lossy and lossless compression methods in isolation, leading to challenges in achieving high compression ratios efficiently. This work proposes a post-training model size compression method that combines lossy and lossless compression in a unified way. We first propose a unified parametric weight transformation, which ensures different lossy compression methods can be performed jointly in a post-training manner. Then, a dedicated differentiable counter is introduced to guide the optimization of lossy compression to arrive at a more suitable point for later lossless compression. Additionally, our method can easily control a desired global compression ratio and allocate adaptive ratios for different layers. Finally, our method can achieve a stable 10× compression ratio without sacrificing accuracy and a 20× compression ratio with minor accuracy loss in a short time. Our code is available at https://github.com/ModelTC/L2_Compression.

The requirement for deploying deep learning (DL) models efficiently has boosted the research of DL compilers. Especially, the difficulty of generating optimized tensor programs has driven DL compilers to commonly adopt the auto-tuning approaches. Consequently, there are increasing demands to improve the effectiveness of auto-tuning in terms of both search efficiency and search quality. However, existing auto-tuning approaches commonly treat subgraphs individually and overlook the similarities among them, and thus fail to generate better tensor programs under limited time budget. To address the above drawbacks, we propose FamilySeer, an auto-tuning framework that can generate better tensor programs by exploiting the subgraph similarities. Specifically, FamilySeer organizes similar subgraphs into subgraph families, where the cost models are built at family basis with improved accuracy for estimating high potential program candidates. To further leverage the similarity, FamilySeer uses the accurate cost model per family to reduce the number of program candidates for costly hardware measurements without degrading search quality. The experiment results on various DL models demonstrate that FamilySeer can achieve better search efficiency/quality on both CPU and GPU platforms compared to the state-of-the-art auto-tuning framework.

Deep neural networks (DNNs) are widely used in various applications. The accurate and latency feedback is essential for model design and deployment. In this work, we attempt to alleviate the cost of model latency acquisition from two aspects: latency query and latency prediction. To ease the difficulty of acquiring model latency on multi-platform, our latency query system can automatically convert DNN model into the corresponding executable format, and measure latency on the target hardware. Powered by this, latency queries can be fulfilled with a simple interface calling. For the efficient utilization of previous latency knowledge, we employ a MySQL database to store numerous models and the corresponding latencies. In our system, the efficiency of latency query can be boosted by 1.8 ×. For latency prediction, we first represent neural networks with the unified GNN-based graph embedding. With the help of the evolving database, our model-based latency predictor achieves better performance, which realizes 12.31% accuracy improvement compared with existing methods. Our codes are open-sourced at https://github.com/ModelTC/NNLQP.

Vision transformers (ViTs) are prevailing among several visual recognition tasks, therefore drawing intensive interest in generating adversarial examples against them. Different from CNNs, ViTs enjoy unique architectures, e.g., self-attention and image-embedding, which are commonly-shared features among various types of transformer-based models. However, existing adversarial methods suffer from weak transferable attacking ability due to the overlook of these architectural features. To address the problem, we propose an Architecture-oriented Transferable Attacking (ATA) framework to generate transferable adversarial examples by activating the uncertain attention and perturbing the sensitive embedding.Specifically, we first locate the patch-wise attentional regions that mostly affect model perception, therefore intensively activating the uncertainty of the attention mechanism and confusing the model decisions in turn.Furthermore, we search the pixel-wise attacking positions that are more likely to derange the embedded tokens using sensitive embedding perturbation, which could serve as a strong transferable attacking pattern.By jointly confusing the unique yet widely-used architectural features among transformer-based models, we can activate strong attacking transferability among diverse ViTs. Extensive experiments on large-scale dataset ImageNet using various popular transformers demonstrate that our ATA outperforms other baselines by large margins (at least +15% Attack Success Rate). Our code is available at https://github.com/nlsde-safety-team/ATA

Transformer architecture has become the fundamental element of the widespread natural language processing~(NLP) models. With the trends of large NLP models, the increasing memory and computation costs hinder their efficient deployment on resource-limited devices. Therefore, transformer quantization attracts wide research interest. Recent work recognizes that structured outliers are the critical bottleneck for quantization performance. However, their proposed methods increase the computation overhead and still leave the outliers there. To fundamentally address this problem, this paper delves into the inherent inducement and importance of the outliers. We discover that $\boldsymbol \gamma$ in LayerNorm (LN) acts as a sinful amplifier for the outliers, and the importance of outliers varies greatly where some outliers provided by a few tokens cover a large area but can be clipped sharply without negative impacts. Motivated by these findings, we propose an outlier suppression framework including two components: Gamma Migration and Token-Wise Clipping. The Gamma Migration migrates the outlier amplifier to subsequent modules in an equivalent transformation, contributing to a more quantization-friendly model without any extra burden. The Token-Wise Clipping takes advantage of the large variance of token range and designs a token-wise coarse-to-fine pipeline, obtaining a clipping range with minimal final quantization loss in an efficient way. This framework effectively suppresses the outliers and can be used in a plug-and-play mode. Extensive experiments prove that our framework surpasses the existing works and, for the first time, pushes the 6-bit post-training BERT quantization to the full-precision (FP) level. Our code is available at https://github.com/wimh966/outlier_suppression.

Quantization has emerged as one of the most prevalent approaches to compress and accelerate neural networks. Recently, data-free quantization has been widely studied as a practical and promising solution. It synthesizes data for calibrating the quantized model according to the batch normalization (BN) statistics of FP32 ones and significantly relieves the heavy dependency on real training data in traditional quantization methods. Unfortunately, we find that in practice, the synthetic data identically constrained by BN statistics suffers serious homogenization at both distribution level and sample level and further causes a significant performance drop of the quantized model. We propose Diverse Sample Generation (DSG) scheme to mitigate the adverse effects caused by homogenization. Specifically, we slack the alignment of feature statistics in the BN layer to relax the constraint at the distribution level and design a layerwise enhancement to reinforce specific layers for different data samples. Our DSG scheme is versatile and even able to be applied to the state-of-the-art post-training quantization method like AdaRound. We evaluate the DSG scheme on the large-scale image classification task and consistently obtain significant improvements over various network architectures and quantization methods, especially when quantized to lower bits (e.g., up to 22% improvement on W4A4). Moreover, benefiting from the enhanced diversity, models calibrated with synthetic data perform close to those calibrated with real data and even outperform them on W4A4.

User data confidentiality protection is becoming a rising challenge in the present deep learning research. Without access to data, conventional data-driven model compression faces a higher risk of performance degradation. Recently, some works propose to generate images from a specific pretrained model to serve as training data. However, the inversion process only utilizes biased feature statistics stored in one model and is from low-dimension to high-dimension. As a consequence, it inevitably encounters the difficulties of generalizability and inexact inversion, which leads to unsatisfactory performance. To address these problems, we propose MixMix based on two simple yet effective techniques: (1) Feature Mixing: utilizes various models to construct a universal feature space for generalized inversion; (2) Data Mixing: mixes the synthesized images and labels to generate exact label information. We prove the effectiveness of MixMix from both theoretical and empirical perspectives. Extensive experiments show that MixMix outperforms existing methods on the mainstream compression tasks, including quantization, knowledge distillation and pruning. Specifically, MixMix achieves up to 4% and 20% accuracy uplift on quantization and pruning, respectively, compared to existing data-free compression work.

Quantization Neural Networks (QNN) have attracted a lot of attention due to their high efficiency. To enhance the quantization accuracy, prior works mainly focus on designing advanced quantization algorithms but still fail to achieve satisfactory results under the extremely low-bit case. In this work, we take an architecture perspective to investigate the potential of high-performance QNN. Therefore, we propose to combine Network Architecture Search methods with quantization to enjoy the merits of the two sides. However, a naive combination inevitably faces unacceptable time consumption or unstable training problem. To alleviate these problems, we first propose the joint training of architecture and quantization with a shared step size to acquire a large number of quantized models. Then a bit-inheritance scheme is introduced to transfer the quantized models to the lower bit, which further reduces the time cost and meanwhile improves the quantization accuracy. Equipped with this overall framework, dubbed as Once Quantization-Aware Training (OQAT), our searched model family, OQATNets, achieves a new state-of-the-art compared with various architectures under different bit-widths. In particular, OQAT-2bit-M achieves 61.6% ImageNet Top-1 accuracy, outperforming 2-bit counterpart MobileNetV3 by a large margin of 9% with 10% less computation cost. A series of quantization-friendly architectures are identified easily and extensive analysis can be made to summarize the interaction between quantization and neural architectures. Codes and models are released at https://github.com/LaVieEnRoseSMZ/OQA

Spiking Neural Network (SNN) has been recognized as one of the next generation of neural networks. Conventionally, SNN can be converted from a pre-trained ANN by only replacing the ReLU activation to spike activation while keeping the parameters intact. Perhaps surprisingly, in this work we show that a proper way to calibrate the parameters during the conversion of ANN to SNN can bring significant improvements. We introduce SNN Calibration, a cheap but extraordinarily effective method by leveraging the knowledge within a pre-trained Artificial Neural Network (ANN). Starting by analyzing the conversion error and its propagation through layers theoretically, we propose the calibration algorithm that can correct the error layer-by-layer. The calibration only takes a handful number of training data and several minutes to finish. Moreover, our calibration algorithm can produce SNN with state-of-the-art architecture on the large-scale ImageNet dataset, including MobileNet and RegNet. Extensive experiments demonstrate the effectiveness and efficiency of our algorithm. For example, our advanced pipeline can increase up to 69% top-1 accuracy when converting MobileNet on ImageNet compared to baselines. Codes are released at https://github.com/yhhhli/SNN_Calibration.

Weight and activation binarization is an effective approach to deep neural network compression and can accelerate the inference by leveraging bitwise operations. Although many binarization methods have improved the accuracy of the model by minimizing the quantization error in forward propagation, there remains a noticeable performance gap between the binarized model and the full-precision one. Our empirical study indicates that the quantization brings information loss in both forward and backward propagation, which is the bottleneck of training accurate binary neural networks. To address these issues, we propose an Information Retention Network (IR-Net) to retain the information that consists in the forward activations and backward gradients. IR-Net mainly relies on two technical contributions: (1) Libra Parameter Binarization (Libra-PB): simultaneously minimizing both quantization error and information loss of parameters by balanced and standardized weights in forward propagation; (2) Error Decay Estimator (EDE): minimizing the information loss of gradients by gradually approximating the sign function in backward propagation, jointly considering the updating ability and accurate gradients. We are the first to investigate both forward and backward processes of binary networks from the unified information perspective, which provides new insight into the mechanism of network binarization. Comprehensive experiments with various network structures on CIFAR-10 and ImageNet datasets manifest that the proposed IR-Net can consistently outperform state-of-the-art quantization methods.

In this paper, we consider the low-bit quantization problem of face recognition (FR) under the open-set protocol. Different from well explored low-bit quantization on closed-set image classification task, the open-set task is more sensitive to quantization errors (QEs). We redefine the QEs in angular space and disentangle it into class error and individual error. These two parts correspond to inter-class separability and intra-class compactness, respectively. Instead of eliminating the entire QEs, we propose the rotation consistent margin (RCM) loss to minimize the individual error, which is more essential to feature discriminative power. Extensive experiments on popular benchmark datasets such as MegaFace Challenge, Youtube Faces (YTF), Labeled Face in the Wild (LFW) and IJB-C show the superiority of proposed loss in low-bit FR quantization tasks.

Recently low-bit (e.g., 8-bit) network quantization has been extensively studied to accelerate the inference. Besides inference, low-bit training with quantized gradients can further bring more considerable acceleration, since the backward process is often computation-intensive. Unfortunately, the inappropriate quantization of backward propagation usually makes the training unstable and even crash. There lacks a successful unified low-bit training framework that can support diverse networks on various tasks. In this paper, we give an attempt to build a unified 8-bit (INT8) training framework for common convolutional neural networks from the aspects of both accuracy and speed. First, we empirically find the four distinctive characteristics of gradients, which provide us insightful clues for gradient quantization. Then, we theoretically give an in-depth analysis of the convergence bound and derive two principles for stable INT8 training. Finally, we propose two universal techniques, including Direction Sensitive Gradient Clipping that reduces the direction deviation of gradients and Deviation Counteractive Learning Rate Scaling that avoids illegal gradient update along the wrong direction. The experiments show that our unified solution promises accurate and efficient INT8 training for a variety of networks and tasks, including MobileNetV2, InceptionV3 and object detection that prior studies have never succeeded. Moreover, it enjoys a strong flexibility to run on off-the-shelf hardware, and reduces the training time by 22% on Pascal GPU without too much optimization effort. We believe that this pioneering study will help lead the community towards a fully unified INT8 training for convolutional neural networks.

Binary neural networks have attracted numerous attention in recent years. However, mainly due to the information loss stemming from the biased binarization, how to preserve the accuracy of networks still remains a critical issue. In this paper, we attempt to maintain the information propagated in the forward process and propose a Balanced Binary Neural Networks with Gated Residual (BBG for short). First, a weight balanced binarization is introduced and thus the informative binary weights can capture more information contained in the activations. Second, for binary activations, a gated residual is further appended to compensate their information loss during the forward process, with a slight overhead. Both techniques can be wrapped as a generic network module that supports various network architectures for different tasks including classification and detection. The experimental results show that BBG-Net performs remarkably well across various network architectures such as VGG, ResNet and SSD with the superior performance over state-of-the-art methods.

With the continuous demand for higher accuracy of deep neural networks, the model size has increased significantly. Quantization is one of the most widely used model compression methods, which can effectively reduce the model size without severe accuracy loss. Modern processors such as ARM CPU and NVIDIA GPU have already provided the support of low-bit arithmetic instructions. However, there lack efficient and practical optimizations for convolution computation towards extremely low-bit on ARM CPU (e.g., 2 ∼ 8-bit) and NVIDIA GPU (e.g., 4-bit and 8-bit). This paper explores the performance optimization methods of extremely low-bit convolution on diverse architectures. On ARM CPU, we propose two instruction schemes for 2 ∼ 3-bit and 4 ∼ 8-bit convolution with corresponding register allocation methods. In addition, we re-design the GEMM computation with data padding and packing optimizations. We also implement winograd algorithm for convolution with some specific bit width (e.g., 4 ∼ 6-bit) to achieve higher performance. On NVIDIA GPU, we propose a data partition mechanism and multi-level memory access optimizations, to better adapt the computation to GPU thread and memory hierarchy. We also propose quantization fusion to eliminate unnecessary data access. The experiment results demonstrate our implementations achieve better performance of extremely low-bit convolution compared to the state-of-the-art frameworks and libraries such as ncnn and cuDNN. To the best of our knowledge, this is the first work that provides efficient implementations of extremely low-bit convolutions covering 2 ∼ 8-bit on ARM CPU and 4-bit/8-bit on NVIDIA GPU.

Hardware-friendly network quantization (e.g., binary/uniform quantization) can efficiently accelerate the inference and meanwhile reduce memory consumption of the deep neural networks, which is crucial for model deployment on resource-limited devices like mobile phones. However, due to the discreteness of low-bit quantization, existing quantization methods often face the unstable training process and severe performance degradation. To address this problem, in this paper we propose Differentiable Soft Quantization (DSQ) to bridge the gap between the full-precision and low-bit networks. DSQ can automatically evolve during training to gradually approximate the standard quantization. Owing to its differentiable property, DSQ can help pursue the accurate gradients in backward propagation, and reduce the quantization loss in forward process with an appropriate clipping range. Extensive experiments over several popular network structures show that training low-bit neural networks with DSQ can consistently outperform state-of-the-art quantization methods. Besides, our first efficient implementation for deploying 2 to 4-bit DSQ on devices with ARM architecture achieves up to 1.7× speed up, compared with the open-source 8-bit high-performance inference framework NCNN [31].