Efficient LLM inference on resource-constrained devices presents significant challenges in compute and memory utilization. Due to limited GPU memory, existing systems offload model weights to CPU memory, incurring substantial I/O overhead between the CPU and GPU. This leads to two major inefficiencies: (1) GPU cores are underutilized, often remaining idle while waiting for data to be loaded; and (2) GPU memory has low impact on performance, as reducing its capacity has minimal effect on overall throughput.In this paper, we propose SpecOffload, a high-throughput inference engine that embeds speculative decoding into offloading. Our key idea is to unlock latent GPU resources for storing and executing a draft model used for speculative decoding, thus accelerating inference at near-zero additional cost. To support this, we carefully orchestrate the interleaved execution of target and draft models in speculative decoding within the offloading pipeline, and propose a planner to manage tensor placement and select optimal parameters. Compared to the best baseline, SpecOffload improves GPU core utilization by 4.49x and boosts inference throughput by 2.54x. Our code is available at https://github.com/MobiSense/SpecOffload-public .

The parameter-efficient adaptation of the image-text pre-training model CLIP for video-text retrieval is a prominent area of research. While CLIP is focused on image-level vision-language matching, video-text retrieval demands comprehensive understanding at the video level. Three key discrepancies emerge in the transfer from image-level to video-level: vision, language, and alignment. However, existing methods mainly focus on vision while neglecting language and alignment. In this paper, we propose Discrepancy Reduction in Vision, Language, and Alignment (DiscoVLA), which simultaneously mitigates all three discrepancies. Specifically, we introduce Image-Video Features Fusion to integrate image-level and video-level features, effectively tackling both vision and language discrepancies. Additionally, we generate pseudo image captions to learn fine-grained image-level alignment. To mitigate alignment discrepancies, we propose Image-To-Video Alignment Distillation, which leverages image-level alignment knowledge to enhance video-level alignment. Extensive experiments demonstrate the superiority of our DiscoVLA. In particular, on MSRVTT with CLIP (ViT-B/16), DiscoVLA outperforms previous methods by 2.2% R@1 and 7.5% R@sum. The code is available at https://github.com/LunarShen/DsicoVLA.

Most text-video retrieval methods utilize the text-image pre-trained models like CLIP as a backbone. These methods process each sampled frame independently by the image encoder, resulting in high computational overhead and limiting practical deployment. Addressing this, we focus on efficient text-video retrieval by tackling two key challenges: 1. From the perspective of trainable parameters, current parameter-efficient fine-tuning methods incur high inference costs; 2. From the perspective of model complexity, current token compression methods are mainly designed for images to reduce spatial redundancy but overlook temporal redundancy in consecutive frames of a video. To tackle these challenges, we propose Temporal Token Merging (TempMe), a parameter-efficient and training-inference efficient text-video retrieval architecture that minimizes trainable parameters and model complexity. Specifically, we introduce a progressive multi-granularity framework. By gradually combining neighboring clips, we reduce spatio-temporal redundancy and enhance temporal modeling across different frames, leading to improved efficiency and performance. Extensive experiments validate the superiority of our TempMe. Compared to previous parameter-efficient text-video retrieval methods, TempMe achieves superior performance with just 0.50M trainable parameters. It significantly reduces output tokens by 95% and GFLOPs by 51%, while achieving a 1.8X speedup and a 4.4% R-Sum improvement. With full fine-tuning, TempMe achieves a significant 7.9% R-Sum improvement, trains 1.57X faster, and utilizes 75.2% GPU memory usage. The code is available at https://github.com/LunarShen/TempMe.

In order to boost the performance of a convolutional neural network (CNN), several approaches have shown the benefit of enhancing the spatial encoding of feature maps. However, few works paid attention to the positional properties of convolutional kernels. In this paper, we demonstrate that different kernel positions are of different importance, which depends on the task, dataset and architecture, and adaptively emphasizing the informative parts in convolutional kernels can lead to considerable improvement. Therefore, we propose a novel structural re-parameterization Position Boosting Convolution (PBConv) to exploit and enhance the position information in the convolutional kernel. PBConv consists of several concurrent small convolutional kernels, which can be equivalently converted to the original kernel and bring no extra inference cost. Different from existing structural re-parameterization methods, PBconv searches for the optimal re-parameterized structure by a fast heuristic algorithm based on the dispersion of kernel weights. Such heuristic search is efficient yet effective, well adapting the varying kernel weight distribution. As a result, PBConv can significantly improve the representational power of a model, especially its ability to extract fine-grained low-level features. Importantly, PBConv is orthogonal to procedural re-parameterization methods and can further boost performance based on them.

With the continuous advancement of deep learning, object detection has made remarkable progress in accurately identifying a wide range of object categories, even within increasingly complex scenes. However, as the number of categories grows, visual concepts naturally organize into a label hierarchy. We contend that existing hierarchical classification and detection methods predominantly prioritize fine-grained prediction, potentially leading to inconsistencies with realistic human perception. From this perspective, we investigate the Hierarchical Object Detection (HOD) problem to better align with real-world perception. To address the lack of benchmarks in the field, we build a large-scale HOD benchmark termed RHOD with open-source datasets, comprising 740 categories. To better align the hierarchical object detectors towards realistic perception, we propose a new evaluation metric named Hierarchical Average Precision (HAP). Furthermore, we present a novel hierarchical object detection method that includes two components, Tree Soft Labeling (TSL) and Hierarchical Extension and Suppression (HES). Our method mitigates the issue of overconfidence in fine-grained predictions, which has been prevalent in previous approaches. We evaluate a range of existing methods on the RHOD benchmark, including plain, hierarchical, and open-vocabulary models. Additionally, we perform comprehensive experiments to assess the performance of our proposed method. The experimental results show that our method achieves state-of-the-art performance on the RHOD benchmark.

Few-shot multi-label recognition (FS-MLR) presents a significant challenge due to the need to assign multiple labels to images with limited examples. Existing methods often struggle to balance the learning of novel classes and the retention of knowledge from base classes. To address this issue, we propose a novel Cross-Modality Prompts (CMP) approach. Unlike conventional methods that rely on additional semantic information to mitigate the impact of limited samples, our approach leverages multimodal prompts to adaptively tune the feature extraction network. A new FS-MLR benchmark is also proposed, which includes single-label training and multi-label testing, accompanied by benchmark datasets constructed from MS-COCO and NUS-WIDE. Extensive experiments on these datasets demonstrate the superior performance of our CMP approach, highlighting its effectiveness and adaptability. Our results show that CMP outperforms CoOp on the MS-COCO dataset with a maximal improvement of 19.47% and 23.94% in mAPharmonicfor 5-way 1-shot and 5-way 5-shot settings, respectively.

Temporal modeling plays an important role in the effective adaption of the powerful pretrained text-image foundation model into text-video retrieval. However, existing methods often rely on additional heavy trainable modules, such as transformer or BiLSTM, which are inefficient. In contrast, we avoid introducing such heavy components by leveraging frozen foundation models. To this end, we propose temporal modeling with frozen vision-language foundation models (TFVL) to model the temporal dynamics with fixed encoders. Specifically, text encoder temporal modeling (TextTemp) and image encoder temporal modeling (ImageTemp) apply frozen text and image encoders within the video head and video backbone, respectively. TextTemp uses a frozen text encoder to interpret frame representations as "visual words" within a temporal "sentence," capturing temporal dependencies. On the other hand, ImageTemp uses a frozen image encoder to treat all frame tokens as a unified visual entity, learning spatiotemporal information. The total trainable parameters of our method, comprising a lightweight projection and several prompt tokens, are significantly fewer than those in other existing methods. We evaluate the effectiveness of our method on MSR-VTT, DiDeMo, ActivityNet, and LSMDC. Compared with full fine-tuning on MSR-VTT, our TFVL achieves an average 3.25% gain in R@1 with merely 0.35% of the parameters. Extensive experiments demonstrate that the proposed TFVL outperforms state-of-the-art methods with significantly fewer parameters.

The existence of redundancy in convolutional neural networks (CNNs) enables us to remove some filters/channels with acceptable performance drops. However, the training objective of CNNs usually tends to minimize an accuracy-related loss function without any attention paid to the redundancy, making the redundancy distribute randomly on all the filters, such that removing any of them may trigger information loss and accuracy drop, necessitating a fine-tuning step for recovery. In this article, we propose to manipulate the redundancy during training to facilitate network pruning. To this end, we propose a novel centripetal SGD (C-SGD) to make some filters identical, resulting in ideal redundancy patterns, as such filters become purely redundant due to their duplicates, hence removing them does not harm the network. As shown on CIFAR and ImageNet, C-SGD delivers better performance because the redundancy is better organized, compared to the existing methods. The efficiency also characterizes C-SGD because it is as fast as regular SGD, requires no fine-tuning, and can be conducted simultaneously on all the layers even in very deep CNNs. Besides, C-SGD can improve the accuracy of CNNs by first training a model with the same architecture but wider layers and then squeezing it into the original width.

We propose ResRep, a novel method for lossless channel pruning (a.k.a. filter pruning), which slims down a CNN by reducing the width (number of output channels) of convolutional layers. Inspired by the neurobiology research about the independence of remembering and forgetting, we propose to re-parameterize a CNN into the remembering parts and forgetting parts, where the former learn to maintain the performance and the latter learn to prune. Via training with regular SGD on the former but a novel update rule with penalty gradients on the latter, we realize structured sparsity. Then we equivalently merge the remembering and forgetting parts into the original architecture with narrower layers. In this sense, ResRep can be viewed as a successful application of Structural Re-parameterization. Such a methodology distinguishes ResRep from the traditional learning-based pruning paradigm that applies a penalty on parameters to produce sparsity, which may suppress the parameters essential for the remembering. ResRep slims down a standard ResNet-50 with 76.15% accuracy on ImageNet to a narrower one with only 45% FLOPs and no accuracy drop, which is the first to achieve lossless pruning with such a high compression ratio. The code and models are at https://github.com/DingXiaoH/ResRep.