Ming-Chang Yang

Bohao Peng, Jian Wang, Yuechen Zhang et al.

Diffusion models have demonstrated remarkable and robust abilities in both image and video generation. To achieve greater control over generated results, researchers introduce additional architectures, such as ControlNet, Adapters and ReferenceNet, to integrate conditioning controls. However, current controllable generation methods often require substantial additional computational resources, especially for video generation, and face challenges in training or exhibit weak control. In this paper, we propose ControlNeXt: a powerful and efficient method for controllable image and video generation. We first design a more straightforward and efficient architecture, replacing heavy additional branches with minimal additional cost compared to the base model. Such a concise structure also allows our method to seamlessly integrate with other LoRA weights, enabling style alteration without the need for additional training. As for training, we reduce up to 90% of learnable parameters compared to the alternatives. Furthermore, we propose another method called Cross Normalization (CN) as a replacement for Zero-Convolution' to achieve fast and stable training convergence. We have conducted various experiments with different base models across images and videos, demonstrating the robustness of our method.

Yifan Hou, Jian Zhang, James Cheng et al. · eth-zurich

Graph neural networks (GNNs) have been widely used for representation learning on graph data. However, there is limited understanding on how much performance GNNs actually gain from graph data. This paper introduces a context-surrounding GNN framework and proposes two smoothness metrics to measure the quantity and quality of information obtained from graph data. A new GNN model, called CS-GNN, is then designed to improve the use of graph information based on the smoothness values of a graph. CS-GNN is shown to achieve better performance than existing methods in different types of real graphs.

Zihan Zhao, Baotong Lu, Shengjie Lin et al.

Long-context LLM serving is bottlenecked by the cost of attending over ever-growing KV caches. Dynamic sparse attention promises relief by accessing only a small, query-dependent subset of the KV state per decoding step and extending the KV storage to CPU memory. In practice, however, these algorithmic savings rarely translate into end-to-end system-level gains because sparse methods typically operate at different granularities and thus rely on ad hoc, per-algorithm implementations. At the same time, hierarchical KV storage introduces a new systems bottleneck: retrieving fine-grained, irregular KV subsets across the GPU-CPU boundary can easily erase the benefits of sparsity. We present SPIN, a sparse-attention-aware inference framework that co-designs the execution pipeline with hierarchical KV storage through three techniques: (1) a unified partition abstraction that maps different sparsity granularities onto a shared page-based KV substrate; (2) a locality-aware KV cache manager that dynamically sizes per-request HBM budgets and uses a GPU-friendly bucketed LRU policy to cut PCIe round-trips; and (3) a two-level hierarchical metadata layout sized to the active working set rather than the worst-case address space. Built on vLLM with three representative sparse attention algorithms, SPIN delivers 1.66-5.66x higher end-to-end throughput and 7-9x lower TTFT than vLLM, and reduces TPOT by up to 58% over the original sparse-attention implementations.

Bohao Peng, Zhuotao Tian, Shu Liu et al.

Continual learning has gained increasing importance as it facilitates the acquisition and refinement of scalable knowledge and skills in language models. However, existing methods typically encounter strict limitations and challenges in real-world scenarios, such as reliance on experience replay, optimization constraints, and inference task-ID. In this study, we introduce the Scalable Language Model (SLM) to overcome these limitations within a more challenging and generalized setting, representing a significant advancement toward practical applications for continual learning. Specifically, we propose the Joint Adaptive Re-Parameterization (JARe), integrated with Dynamic Task-related Knowledge Retrieval (DTKR), to enable adaptive adjustment of language models based on specific downstream tasks. This approach leverages the task distribution within the vector space, aiming to achieve a smooth and effortless continual learning process. Our method demonstrates state-of-the-art performance on diverse backbones and benchmarks, achieving effective continual learning in both full-set and few-shot scenarios with minimal forgetting. Moreover, while prior research primarily focused on a single task type such as classification, our study goes beyond, with the large language model, i.e., LLaMA-2, to explore the effects across diverse domains and task types, such that a single language model can be decently scaled to broader applications.

Pengguang Chen, Yixin Chen, Shu Liu et al.

We study the vision transformer structure in the mobile level in this paper, and find a dramatic performance drop. We analyze the reason behind this phenomenon, and propose a novel irregular patch embedding module and adaptive patch fusion module to improve the performance. We conjecture that the vision transformer blocks (which consist of multi-head attention and feed-forward network) are more suitable to handle high-level information than low-level features. The irregular patch embedding module extracts patches that contain rich high-level information with different receptive fields. The transformer blocks can obtain the most useful information from these irregular patches. Then the processed patches pass the adaptive patch merging module to get the final features for the classifier. With our proposed improvements, the traditional uniform vision transformer structure can achieve state-of-the-art results in mobile level. We improve the DeiT baseline by more than 9\% under the mobile-level settings and surpass other transformer architectures like Swin and CoaT by a large margin.

Jie Liu, Xiao Yan, Xinyan Dai et al.

The inner-product navigable small world graph (ip-NSW) represents the state-of-the-art method for approximate maximum inner product search (MIPS) and it can achieve an order of magnitude speedup over the fastest baseline. However, to date it is still unclear where its exceptional performance comes from. In this paper, we show that there is a strong norm bias in the MIPS problem, which means that the large norm items are very likely to become the result of MIPS. Then we explain the good performance of ip-NSW as matching the norm bias of the MIPS problem - large norm items have big in-degrees in the ip-NSW proximity graph and a walk on the graph spends the majority of computation on these items, thus effectively avoids unnecessary computation on small norm items. Furthermore, we propose the ip-NSW+ algorithm, which improves ip-NSW by introducing an additional angular proximity graph. Search is first conducted on the angular graph to find the angular neighbors of a query and then the MIPS neighbors of these angular neighbors are used to initialize the candidate pool for search on the inner-product proximity graph. Experiment results show that ip-NSW+ consistently and significantly outperforms ip-NSW and provides more robust performance under different data distributions.