Xinjian Wu

Xinjian Wu, Fanhu Zeng, Xiudong Wang et al.

Vision Transformers (ViTs) have emerged as powerful models in the field of computer vision, delivering superior performance across various vision tasks. However, the high computational complexity poses a significant barrier to their practical applications in real-world scenarios. Motivated by the fact that not all tokens contribute equally to the final predictions and fewer tokens bring less computational cost, reducing redundant tokens has become a prevailing paradigm for accelerating vision transformers. However, we argue that it is not optimal to either only reduce inattentive redundancy by token pruning, or only reduce duplicative redundancy by token merging. To this end, in this paper we propose a novel acceleration framework, namely token Pruning & Pooling Transformers (PPT), to adaptively tackle these two types of redundancy in different layers. By heuristically integrating both token pruning and token pooling techniques in ViTs without additional trainable parameters, PPT effectively reduces the model complexity while maintaining its predictive accuracy. For example, PPT reduces over 37% FLOPs and improves the throughput by over 45% for DeiT-S without any accuracy drop on the ImageNet dataset. The code is available at https://github.com/xjwu1024/PPT and https://github.com/mindspore-lab/models/

Wenzhuo Liu, Xinjian Wu, Fei Zhu et al.

Deep Neural Network (DNN) has achieved great success on datasets of closed class set. However, new classes, like new categories of social media topics, are continuously added to the real world, making it necessary to incrementally learn. This is hard for DNN because it tends to focus on fitting to new classes while ignoring old classes, a phenomenon known as catastrophic forgetting. State-of-the-art methods rely on knowledge distillation and data replay techniques but still have limitations. In this work, we analyze the causes of catastrophic forgetting in class incremental learning, which owes to three factors: representation drift, representation confusion, and classifier distortion. Based on this view, we propose a two-stage learning framework with a fixed encoder and an incrementally updated prototype classifier. The encoder is trained with self-supervised learning to generate a feature space with high intrinsic dimensionality, thus improving its transferability and generality. The classifier incrementally learns new prototypes while retaining the prototypes of previously learned data, which is crucial in preserving the decision boundary.Our method does not rely on preserved samples of old classes, is thus a non-exemplar based CIL method. Experiments on public datasets show that our method can significantly outperform state-of-the-art exemplar-based methods when they reserved 5 examplers per class, under the incremental setting of 10 phases, by 18.24% on CIFAR-100 and 9.37% on ImageNet100.

Xinjian Wu, Jingzhi Gong, Gunel Jahangirova et al.

Retrieval-augmented generation (RAG) pipelines for code completion rely on chunking to segment source files into retrievable units, yet chunking strategies are typically adopted without empirical justification, and practitioner recommendations are notably inconsistent. We present a controlled empirical study isolating the effect of chunking on code completion quality by crossing four representative strategies (Function, Declaration, Sliding Window, and cAST) with four retrievers, five generators, and nine parameter configurations on two benchmarks (RepoEval and CrossCodeEval), totaling 864 experimental settings. Our results reveal that chunking strategy has a statistically significant effect on RAG-based code completion. Contrary to intuition, chunking based on functions underperforms all other strategies by 3.57--5.64 percentage points on RepoEval (Cliff's delta = -1.0), while the remaining chunking strategies perform comparably. Our further analysis demonstrates that this observation holds across all retriever--generator combinations. We also find that cross-file context length is the dominant parameter: doubling from 2,048 to 8,192 tokens yields up to 4.2 percentage points of improvement, whereas chunk size has a weaker, non-monotonic effect. On the cost--quality Pareto front, Sliding Window and cAST dominate both benchmarks; Function chunking is never Pareto-optimal.

Xinjian Wu, Ruisong Zhang, Jie Qin et al.

Segmenting and recognizing diverse object parts is crucial in computer vision and robotics. Despite significant progress in object segmentation, part-level segmentation remains underexplored due to complex boundaries and scarce annotated data. To address this, we propose a novel Weakly-supervised Part Segmentation (WPS) setting and an approach called WPS-SAM, built on the large-scale pre-trained vision foundation model, Segment Anything Model (SAM). WPS-SAM is an end-to-end framework designed to extract prompt tokens directly from images and perform pixel-level segmentation of part regions. During its training phase, it only uses weakly supervised labels in the form of bounding boxes or points. Extensive experiments demonstrate that, through exploiting the rich knowledge embedded in pre-trained foundation models, WPS-SAM outperforms other segmentation models trained with pixel-level strong annotations. Specifically, WPS-SAM achieves 68.93% mIOU and 79.53% mACC on the PartImageNet dataset, surpassing state-of-the-art fully supervised methods by approximately 4% in terms of mIOU.

Cheng Deng, Luoyang Sun, Jiwen Jiang et al.

While scaling laws have been continuously validated in large language models (LLMs) with increasing model parameters, the inherent tension between the inference demands of LLMs and the limited resources of edge devices poses a critical challenge to the development of edge intelligence. Recently, numerous small language models have emerged, aiming to distill the capabilities of LLMs into smaller footprints. However, these models often retain the fundamental architectural principles of their larger counterparts, still imposing considerable strain on the storage and bandwidth capacities of edge devices. In this paper, we introduce the PLM, a Peripheral Language Model, developed through a co-design process that jointly optimizes model architecture and edge system constraints. The PLM utilizes a Multi-head Latent Attention mechanism and employs the squared ReLU activation function to encourage sparsity, thereby reducing peak memory footprint during inference. During training, we collect and reorganize open-source datasets, implement a multi-phase training strategy, and empirically investigate the Warmup-Stable-Decay-Constant (WSDC) learning rate scheduler. Additionally, we incorporate Reinforcement Learning from Human Feedback (RLHF) by adopting the ARIES preference learning approach. Following a two-phase SFT process, this method yields performance gains of 2% in general tasks, 9% in the GSM8K task, and 11% in coding tasks. In addition to its novel architecture, evaluation results demonstrate that PLM outperforms existing small language models trained on publicly available data while maintaining the lowest number of activated parameters. Furthermore, deployment across various edge devices, including consumer-grade GPUs, mobile phones, and Raspberry Pis, validates PLM's suitability for peripheral applications. The PLM series models are publicly available at https://github.com/plm-team/PLM.

Luoyang Sun, Cheng Deng, Jiwen Jiang et al.

Attention mechanisms underpin the success of large language models (LLMs), yet their substantial computational and memory overhead poses challenges for optimizing efficiency and performance. A critical bottleneck arises as KV cache and attention computations scale rapidly with text length, challenging deployment on hardware with limited computational and memory resources. We observe that attention mechanisms exhibit substantial redundancy, since the KV cache can be significantly compressed and attention maps across heads display high similarity, revealing that much of the computation and storage is unnecessary. Leveraging these insights, we propose \textbf{G}rouped-Head Laten\textbf{T} \textbf{A}ttention (GTA), a novel attention mechanism that reduces memory usage and computational complexity while maintaining performance. GTA comprises two components: (1) a shared attention map mechanism that reuses attention scores across multiple heads, decreasing the key cache size; and (2) a nonlinear value decoder with learned projections that compresses the value cache into a latent space, further cutting memory needs. GTA cuts attention computation FLOPs by up to \emph{62.5\%} versus Grouped-Query Attention and shrink the KV cache by up to \emph{70\%}, all while avoiding the extra overhead of Multi-Head Latent Attention to improve LLM deployment efficiency. Consequently, GTA models achieve a \emph{2x} increase in end-to-end inference speed, with prefill benefiting from reduced computational cost and decoding benefiting from the smaller cache footprint.

Xinjian Wu, Hongmei Wang, Yuan Zhou et al.

The scalability of high-fidelity video diffusion models (VDMs) is constrained by two key sources of redundancy: the quadratic complexity of global spatio-temporal attention and the computational overhead of long iterative denoising trajectories. Existing accelerators -- such as sparse attention and step-distilled samplers -- typically target a single dimension in isolation and quickly encounter diminishing returns, as the remaining bottlenecks become dominant. In this work, we introduce USV (Unified Sparsification for Video diffusion models), an end-to-end trainable framework that overcomes this limitation by jointly orchestrating sparsification across both the model's internal computation and its sampling process. USV learns a dynamic, data- and timestep-dependent sparsification policy that prunes redundant attention connections, adaptively merges semantically similar tokens, and reduces denoising steps, treating them not as independent tricks but as coordinated actions within a single optimization objective. This multi-dimensional co-design enables strong mutual reinforcement among previously disjoint acceleration strategies. Extensive experiments on large-scale video generation benchmarks demonstrate that USV achieves up to 83.3% speedup in the denoising process and 22.7% end-to-end acceleration, while maintaining high visual fidelity. Our results highlight unified, dynamic sparsification as a practical path toward efficient, high-quality video generation.

Zhongwei Wan, Xinjian Wu, Yu Zhang et al.

Generative inference in Large Language Models (LLMs) is impeded by the growing memory demands of Key-Value (KV) cache, especially for longer sequences. Traditional KV cache eviction strategies, which discard less critical KV pairs based on attention scores, often degrade generation quality, leading to issues such as context loss or hallucinations. In this work, we introduce Dynamic Discriminative Operations (D2O), a KV cache compression method that optimizes KV cache size dynamically and discriminatively at two levels without fine-tuning, while preserving essential context. At layer level, D2O leverages the varying densities of attention weights between shallow and deep layers to dynamically determine which layers should avoid excessive eviction via a novel dynamic allocation strategy to minimize information loss. At token level, D2O incorporates a compensation mechanism that maintains a similarity threshold to re-discriminate the importance of currently discarded tokens, determining whether they should be recalled and merged with similar tokens. We conduct experiments on various benchmarks and LLM architectures. Our results show that D2O not only achieves significant memory savings and enhances inference throughput by more than 3$\times$ but also maintains high-quality long-text generation.