LayerKV: Optimizing Large Language Model Serving with Layer-wise KV Cache Management

The expanding context windows in large language models (LLMs) have greatly enhanced their capabilities in various applications, but they also introduce significant challenges in maintaining low latency, particularly in Time to First Token (TTFT). This paper identifies that the sharp rise in TTFT as context length increases is predominantly driven by queuing delays, which are caused by the growing demands for GPU Key-Value (KV) cache allocation clashing with the limited availability of KV cache blocks. To address this issue, we propose LayerKV, a simple yet effective plug-in method that effectively reduces TTFT without requiring additional hardware or compromising output performance, while seamlessly integrating with existing parallelism strategies and scheduling techniques. Specifically, LayerKV introduces layer-wise KV block allocation, management, and offloading for fine-grained control over system memory, coupled with an SLO-aware scheduler to optimize overall Service Level Objectives (SLOs). Comprehensive evaluations on representative models, ranging from 7B to 70B parameters, across various GPU configurations, demonstrate that LayerKV improves TTFT latency up to 11x and reduces SLO violation rates by 28.7\%, significantly enhancing the user experience

vTensor: Flexible Virtual Tensor Management for Efficient LLM Serving

Large Language Models (LLMs) are widely used across various domains, processing millions of daily requests. This surge in demand poses significant challenges in optimizing throughput and latency while keeping costs manageable. The Key-Value (KV) cache, a standard method for retaining previous computations, makes LLM inference highly bounded by memory. While batching strategies can enhance performance, they frequently lead to significant memory fragmentation. Even though cutting-edge systems like vLLM mitigate KV cache fragmentation using paged Attention mechanisms, they still suffer from inefficient memory and computational operations due to the tightly coupled page management and computation kernels. This study introduces the vTensor, an innovative tensor structure for LLM inference based on GPU virtual memory management (VMM). vTensor addresses existing limitations by decoupling computation from memory defragmentation and offering dynamic extensibility. Our framework employs a CPU-GPU heterogeneous approach, ensuring efficient, fragmentation-free memory management while accommodating various computation kernels across different LLM architectures. Experimental results indicate that vTensor achieves an average speedup of 1.86x across different models, with up to 2.42x in multi-turn chat scenarios. Additionally, vTensor provides average speedups of 2.12x and 3.15x in kernel evaluation, reaching up to 3.92x and 3.27x compared to SGLang Triton prefix-prefilling kernels and vLLM paged Attention kernel, respectively. Furthermore, it frees approximately 71.25% (57GB) of memory on the NVIDIA A100 GPU compared to vLLM, enabling more memory-intensive workloads.

Reinforcement Learning-Guided Semi-Supervised Learning

In recent years, semi-supervised learning (SSL) has gained significant attention due to its ability to leverage both labeled and unlabeled data to improve model performance, especially when labeled data is scarce. However, most current SSL methods rely on heuristics or predefined rules for generating pseudo-labels and leveraging unlabeled data. They are limited to exploiting loss functions and regularization methods within the standard norm. In this paper, we propose a novel Reinforcement Learning (RL) Guided SSL method, RLGSSL, that formulates SSL as a one-armed bandit problem and deploys an innovative RL loss based on weighted reward to adaptively guide the learning process of the prediction model. RLGSSL incorporates a carefully designed reward function that balances the use of labeled and unlabeled data to enhance generalization performance. A semi-supervised teacher-student framework is further deployed to increase the learning stability. We demonstrate the effectiveness of RLGSSL through extensive experiments on several benchmark datasets and show that our approach achieves consistent superior performance compared to state-of-the-art SSL methods.

Cross-model Mutual Learning for Exemplar-based Medical Image Segmentation

Medical image segmentation typically demands extensive dense annotations for model training, which is both time-consuming and skill-intensive. To mitigate this burden, exemplar-based medical image segmentation methods have been introduced to achieve effective training with only one annotated image. In this paper, we introduce a novel Cross-model Mutual learning framework for Exemplar-based Medical image Segmentation (CMEMS), which leverages two models to mutually excavate implicit information from unlabeled data at multiple granularities. CMEMS can eliminate confirmation bias and enable collaborative training to learn complementary information by enforcing consistency at different granularities across models. Concretely, cross-model image perturbation based mutual learning is devised by using weakly perturbed images to generate high-confidence pseudo-labels, supervising predictions of strongly perturbed images across models. This approach enables joint pursuit of prediction consistency at the image granularity. Moreover, cross-model multi-level feature perturbation based mutual learning is designed by letting pseudo-labels supervise predictions from perturbed multi-level features with different resolutions, which can broaden the perturbation space and enhance the robustness of our framework. CMEMS is jointly trained using exemplar data, synthetic data, and unlabeled data in an end-to-end manner. Experimental results on two medical image datasets indicate that the proposed CMEMS outperforms the state-of-the-art segmentation methods with extremely limited supervision.

Prompt-Driven Feature Diffusion for Open-World Semi-Supervised Learning

In this paper, we present a novel approach termed Prompt-Driven Feature Diffusion (PDFD) within a semi-supervised learning framework for Open World Semi-Supervised Learning (OW-SSL). At its core, PDFD deploys an efficient feature-level diffusion model with the guidance of class-specific prompts to support discriminative feature representation learning and feature generation, tackling the challenge of the non-availability of labeled data for unseen classes in OW-SSL. In particular, PDFD utilizes class prototypes as prompts in the diffusion model, leveraging their class-discriminative and semantic generalization ability to condition and guide the diffusion process across all the seen and unseen classes. Furthermore, PDFD incorporates a class-conditional adversarial loss for diffusion model training, ensuring that the features generated via the diffusion process can be discriminatively aligned with the class-conditional features of the real data. Additionally, the class prototypes of the unseen classes are computed using only unlabeled instances with confident predictions within a semi-supervised learning framework. We conduct extensive experiments to evaluate the proposed PDFD. The empirical results show PDFD exhibits remarkable performance enhancements over many state-of-the-art existing methods.

Lightweight Unsupervised Federated Learning with Pretrained Vision Language Model

Federated learning aims to tackle the ``isolated data island" problem, where it trains a collective model from physically isolated clients while safeguarding the privacy of users' data. However, supervised federated learning necessitates that each client labels their data for training, which can be both time-consuming and resource-intensive, and may even be impractical for edge devices. Moreover, the training and transmission of deep models present challenges to the computation and communication capabilities of the clients. To address these two inherent challenges in supervised federated learning, we propose a novel lightweight unsupervised federated learning approach that leverages unlabeled data on each client to perform lightweight model training and communication by harnessing pretrained vision-language models, such as CLIP. By capitalizing on the zero-shot prediction capability and the well-trained image encoder of the pre-trained CLIP model, we have carefully crafted an efficient and resilient self-training approach. This method refines the initial zero-shot predicted pseudo-labels of unlabeled instances through the sole training of a linear classifier on top of the fixed image encoder. Additionally, to address data heterogeneity within each client, we propose a class-balanced text feature sampling strategy for generating synthetic instances in the feature space to support local training. Experiments are conducted on multiple benchmark datasets. The experimental results demonstrate that our proposed method greatly enhances model performance in comparison to CLIP's zero-shot predictions and even outperforms supervised federated learning benchmark methods given limited computational and communication overhead.

AKGNet: Attribute Knowledge-Guided Unsupervised Lung-Infected Area Segmentation

Lung-infected area segmentation is crucial for assessing the severity of lung diseases. However, existing image-text multi-modal methods typically rely on labour-intensive annotations for model training, posing challenges regarding time and expertise. To address this issue, we propose a novel attribute knowledge-guided framework for unsupervised lung-infected area segmentation (AKGNet), which achieves segmentation solely based on image-text data without any mask annotation. AKGNet facilitates text attribute knowledge learning, attribute-image cross-attention fusion, and high-confidence-based pseudo-label exploration simultaneously. It can learn statistical information and capture spatial correlations between image and text attributes in the embedding space, iteratively refining the mask to enhance segmentation. Specifically, we introduce a text attribute knowledge learning module by extracting attribute knowledge and incorporating it into feature representations, enabling the model to learn statistical information and adapt to different attributes. Moreover, we devise an attribute-image cross-attention module by calculating the correlation between attributes and images in the embedding space to capture spatial dependency information, thus selectively focusing on relevant regions while filtering irrelevant areas. Finally, a self-training mask improvement process is employed by generating pseudo-labels using high-confidence predictions to iteratively enhance the mask and segmentation. Experimental results on a benchmark medical image dataset demonstrate the superior performance of our method compared to state-of-the-art segmentation techniques in unsupervised scenarios.

Adaptive Weighted Co-Learning for Cross-Domain Few-Shot Learning

Due to the availability of only a few labeled instances for the novel target prediction task and the significant domain shift between the well annotated source domain and the target domain, cross-domain few-shot learning (CDFSL) induces a very challenging adaptation problem. In this paper, we propose a simple Adaptive Weighted Co-Learning (AWCoL) method to address the CDFSL challenge by adapting two independently trained source prototypical classification models to the target task in a weighted co-learning manner. The proposed method deploys a weighted moving average prediction strategy to generate probabilistic predictions from each model, and then conducts adaptive co-learning by jointly fine-tuning the two models in an alternating manner based on the pseudo-labels and instance weights produced from the predictions. Moreover, a negative pseudo-labeling regularizer is further deployed to improve the fine-tuning process by penalizing false predictions. Comprehensive experiments are conducted on multiple benchmark datasets and the empirical results demonstrate that the proposed method produces state-of-the-art CDFSL performance.

Efficient Low-Rank GNN Defense Against Structural Attacks

Graph Neural Networks (GNNs) have been shown to possess strong representation abilities over graph data. However, GNNs are vulnerable to adversarial attacks, and even minor perturbations to the graph structure can significantly degrade their performance. Existing methods either are ineffective against sophisticated attacks or require the optimization of dense adjacency matrices, which is time-consuming and prone to local minima. To remedy this problem, we propose an Efficient Low-Rank Graph Neural Network (ELR-GNN) defense method, which aims to learn low-rank and sparse graph structures for defending against adversarial attacks, ensuring effective defense with greater efficiency. Specifically, ELR-GNN consists of two modules: a Coarse Low-Rank Estimation Module and a Fine-Grained Estimation Module. The first module adopts the truncated Singular Value Decomposition (SVD) to initialize the low-rank adjacency matrix estimation, which serves as a starting point for optimizing the low-rank matrix. In the second module, the initial estimate is refined by jointly learning a low-rank sparse graph structure with the GNN model. Sparsity is incorporated into the learned low-rank adjacency matrix by pruning weak connections, which can reduce redundant data while maintaining valuable information. As a result, instead of using the dense adjacency matrix directly, ELR-GNN can learn a low-rank and sparse estimate of it in a simple, efficient and easy to optimize manner. The experimental results demonstrate that ELR-GNN outperforms the state-of-the-art GNN defense methods in the literature, in addition to being very efficient and easy to train.

GDM: Dual Mixup for Graph Classification with Limited Supervision

Graph Neural Networks (GNNs) require a large number of labeled graph samples to obtain good performance on the graph classification task. The performance of GNNs degrades significantly as the number of labeled graph samples decreases. To reduce the annotation cost, it is therefore important to develop graph augmentation methods that can generate new graph instances to increase the size and diversity of the limited set of available labeled graph samples. In this work, we propose a novel mixup-based graph augmentation method, Graph Dual Mixup (GDM), that leverages both functional and structural information of the graph instances to generate new labeled graph samples. GDM employs a graph structural auto-encoder to learn structural embeddings of the graph samples, and then applies mixup to the structural information of the graphs in the learned structural embedding space and generates new graph structures from the mixup structural embeddings. As for the functional information, GDM applies mixup directly to the input node features of the graph samples to generate functional node feature information for new mixup graph instances. Jointly, the generated input node features and graph structures yield new graph samples which can supplement the set of original labeled graphs. Furthermore, we propose two novel Balanced Graph Sampling methods to enhance the balanced difficulty and diversity for the generated graph samples. Experimental results on the benchmark datasets demonstrate that our proposed method substantially outperforms the state-of-the-art graph augmentation methods when the labeled graphs are scarce.