Yao Lu

Shang Yang, Junxian Guo, Haotian Tang et al. · mit

Large language models (LLMs) have shown remarkable potential in processing long sequences and complex reasoning tasks, yet efficiently serving these models remains challenging due to the quadratic computational complexity of attention in the prefilling stage and the large memory footprint of the KV cache in the decoding stage. To address these issues, we introduce LServe, an efficient system that accelerates long-sequence LLM serving via hybrid sparse attention. This method unifies different hardware-friendly, structured sparsity patterns for both prefilling and decoding attention into a single framework, where computations on less important tokens are skipped block-wise. LServe demonstrates the compatibility of static and dynamic sparsity in long-context LLM attention. This design enables multiplicative speedups by combining these optimizations. Specifically, we convert half of the attention heads to nearly free streaming heads in both the prefilling and decoding stages. Additionally, we find that only a constant number of KV pages is required to preserve long-context and reasoning capabilities, irrespective of context length. We then design a hierarchical KV page selection policy that dynamically prunes KV pages based on query-centric similarity. On average, LServe accelerates LLM prefilling by up to 2.9x and decoding by 1.3-2.1x over vLLM, maintaining long-context accuracy. Code is released at https://github.com/mit-han-lab/omniserve.

Yao Lu, Peixin Zhang, Jingyi Wang et al.

Deep learning has revolutionized computing in many real-world applications, arguably due to its remarkable performance and extreme convenience as an end-to-end solution. However, deep learning models can be costly to train and to use, especially for those large-scale models, making it necessary to optimize the original overly complicated models into smaller ones in scenarios with limited resources such as mobile applications or simply for resource saving. The key question in such model optimization is, how can we effectively identify and measure the redundancy in a deep learning model structure. While several common metrics exist in the popular model optimization techniques to measure the performance of models after optimization, they are not able to quantitatively inform the degree of remaining redundancy. To address the problem, we present a novel testing approach, i.e., RedTest, which proposes a novel testing metric called Model Structural Redundancy Score (MSRS) to quantitatively measure the degree of redundancy in a deep learning model structure. We first show that MSRS is effective in both revealing and assessing the redundancy issues in many state-of-the-art models, which urgently calls for model optimization. Then, we utilize MSRS to assist deep learning model developers in two practical application scenarios: 1) in Neural Architecture Search, we design a novel redundancy-aware algorithm to guide the search for the optimal model structure and demonstrate its effectiveness by comparing it to existing standard NAS practice; 2) in the pruning of large-scale pre-trained models, we prune the redundant layers of pre-trained models with the guidance of layer similarity to derive less redundant ones of much smaller size. Extensive experimental results demonstrate that removing such redundancy has a negligible effect on the model utility.

Yao Lu, Hiram Rayo Torres Rodriguez, Sebastian Vogel et al.

Neural Architecture Search (NAS) has become the de-facto approach for designing accurate and efficient networks for edge devices. Since models are typically quantized for edge deployment, recent work has investigated quantization-aware NAS (QA-NAS) to search for highly accurate and efficient quantized models. However, existing QA-NAS approaches, particularly few-bit mixed-precision (FB-MP) methods, do not scale to larger tasks. Consequently, QA-NAS has mostly been limited to low-scale tasks and tiny networks. In this work, we present an approach to enable QA-NAS (INT8 and FB-MP) on large-scale tasks by leveraging the block-wise formulation introduced by block-wise NAS. We demonstrate strong results for the semantic segmentation task on the Cityscapes dataset, finding FB-MP models 33% smaller and INT8 models 17.6% faster than DeepLabV3 (INT8) without compromising task performance.

Yao Lu, Chunfeng Sun, Dongwei Xu et al.

Deep learning-based Automatic Modulation Recognition (AMR) model has made significant progress with the support of large-scale labeled data. However, when developing new models or performing hyperparameter tuning, the time and energy consumption associated with repeated training using massive amounts of data are often unbearable. To address the above challenges, we propose \emph{FoQuS}, which approximates the effect of full training by selecting a coreset from the original dataset, thereby significantly reducing training overhead. Specifically, \emph{FoQuS} records the prediction trajectory of each sample during full-dataset training and constructs three importance metrics based on training dynamics. Experiments show that \emph{FoQuS} can maintain high recognition accuracy and good cross-architecture generalization on multiple AMR datasets using only 1\%-30\% of the original data.

Kangfu Mei, Yao Lu, Qiaosi Yi et al.

Pursuing realistic results according to human visual perception is the central concern in the image transformation tasks. Perceptual learning approaches like perceptual loss are empirically powerful for such tasks but they usually rely on the pre-trained classification network to provide features, which are not necessarily optimal in terms of visual perception of image transformation. In this paper, we argue that, among the features representation from the pre-trained classification network, only limited dimensions are related to human visual perception, while others are irrelevant, although both will affect the final image transformation results. Under such an assumption, we try to disentangle the perception-relevant dimensions from the representation through our proposed online contrastive learning. The resulted network includes the pre-training part and a feature selection layer, followed by the contrastive learning module, which utilizes the transformed results, target images, and task-oriented distorted images as the positive, negative, and anchor samples, respectively. The contrastive learning aims at activating the perception-relevant dimensions and suppressing the irrelevant ones by using the triplet loss, so that the original representation can be disentangled for better perceptual quality. Experiments on various image transformation tasks demonstrate the superiority of our framework, in terms of human visual perception, to the existing approaches using pre-trained networks and empirically designed losses.

Yao Lu, Sören Pirk, Jan Dlabal et al.

Many computer vision tasks address the problem of scene understanding and are naturally interrelated e.g. object classification, detection, scene segmentation, depth estimation, etc. We show that we can leverage the inherent relationships among collections of tasks, as they are trained jointly, supervising each other through their known relationships via consistency losses. Furthermore, explicitly utilizing the relationships between tasks allows improving their performance while dramatically reducing the need for labeled data, and allows training with additional unsupervised or simulated data. We demonstrate a distributed joint training algorithm with task-level parallelism, which affords a high degree of asynchronicity and robustness. This allows learning across multiple tasks, or with large amounts of input data, at scale. We demonstrate our framework on subsets of the following collection of tasks: depth and normal prediction, semantic segmentation, 3D motion and ego-motion estimation, and object tracking and 3D detection in point clouds. We observe improved performance across these tasks, especially in the low-label regime.

Yao Lu, Jukka Corander, Zhirong Yang

Stochastic Neighbor Embedding (SNE) methods minimize the divergence between the similarity matrix of a high-dimensional data set and its counterpart from a low-dimensional embedding, leading to widely applied tools for data visualization. Despite their popularity, the current SNE methods experience a crowding problem when the data include highly imbalanced similarities. This implies that the data points with higher total similarity tend to get crowded around the display center. To solve this problem, we introduce a fast normalization method and normalize the similarity matrix to be doubly stochastic such that all the data points have equal total similarities. Furthermore, we show empirically and theoretically that the doubly stochasticity constraint often leads to embeddings which are approximately spherical. This suggests replacing a flat space with spheres as the embedding space. The spherical embedding eliminates the discrepancy between the center and the periphery in visualization, which efficiently resolves the crowding problem. We compared the proposed method (DOSNES) with the state-of-the-art SNE method on three real-world datasets and the results clearly indicate that our method is more favorable in terms of visualization quality.