Large language models (LLMs) demonstrate exceptional performance but incur high serving costs due to substantial memory demands, with the key-value (KV) cache being a primary bottleneck. Existing KV cache compression methods, including quantization and pruning, struggle with limitations such as uniform treatment of keys and values and static memory allocation across attention heads. To address these challenges, we introduce LeanKV, a unified KV cache compression framework that enhances LLM serving efficiency without compromising accuracy through three innovations: (1) Hetero-KV quantization, which stores keys at a higher precision than values to reflect their greater impact on attention computations; (2) per-head dynamic sparsity, which allocates memory based on token importance per head and per request; and (3) unified KV compression, integrating mixed-precision quantization and selective pruning to enable a smooth tradeoff between model accuracy and memory efficiency. To efficiently support these techniques, LeanKV introduces systems optimizations including unified paging and on-GPU parallel memory management. Implemented on vLLM, LeanKV compresses the KV cache by $3.0\times$ to $5.0\times$ without accuracy loss and up to $11.0\times$ with under 5% accuracy loss, enhancing throughput by $1.9\times$ to $2.5\times$, and up to $6.9\times$.

Recent research has explored the use of Large Language Models (LLMs) for tackling complex graph reasoning tasks. However, due to the intricacies of graph structures and the inherent limitations of LLMs in handling long text, current approaches often fail to deliver satisfactory accuracy, even on small-scale graphs and simple tasks. To address these challenges, we introduce GraphAgent-Reasoner, a fine-tuning-free framework that utilizes a multi-agent collaboration strategy for explicit and precise graph reasoning. Inspired by distributed graph computation theory, our framework decomposes graph problems into smaller, node-centric tasks that are distributed among multiple agents. The agents collaborate to solve the overall problem, significantly reducing the amount of information and complexity handled by a single LLM, thus enhancing the accuracy of graph reasoning. By simply increasing the number of agents, GraphAgent-Reasoner can efficiently scale to accommodate larger graphs with over 1,000 nodes. Evaluated on the GraphInstruct dataset, our framework demonstrates near-perfect accuracy on polynomial-time graph reasoning tasks, significantly outperforming the best available models, both closed-source and fine-tuned open-source variants. Our framework also demonstrates the capability to handle real-world graph reasoning applications such as webpage importance analysis.

Low-Rank Adaptation (LoRA) offers an efficient way to fine-tune large language models (LLMs). Its modular and plug-and-play nature allows the integration of various domain-specific LoRAs, enhancing LLM capabilities. Open-source platforms like Huggingface and Modelscope have introduced a new computational paradigm, Uploadable Machine Learning (UML). In UML, contributors use decentralized data to train specialized adapters, which are then uploaded to a central platform to improve LLMs. This platform uses these domain-specific adapters to handle mixed-task requests requiring personalized service. Previous research on LoRA composition either focuses on specific tasks or fixes the LoRA selection during training. However, in UML, the pool of LoRAs is dynamically updated with new uploads, requiring a generalizable selection mechanism for unseen LoRAs. Additionally, the mixed-task nature of downstream requests necessitates personalized services. To address these challenges, we propose Retrieval-Augmented Mixture of LoRA Experts (RAMoLE), a framework that adaptively retrieves and composes multiple LoRAs based on input prompts. RAMoLE has three main components: LoraRetriever for identifying and retrieving relevant LoRAs, an on-the-fly MoLE mechanism for coordinating the retrieved LoRAs, and efficient batch inference for handling heterogeneous requests. Experimental results show that RAMoLE consistently outperforms baselines, highlighting its effectiveness and scalability.

Graph Neural Networks (GNNs) excel across various applications but remain vulnerable to adversarial attacks, particularly Graph Injection Attacks (GIAs), which inject malicious nodes into the original graph and pose realistic threats. Text-attributed graphs (TAGs), where nodes are associated with textual features, are crucial due to their prevalence in real-world applications and are commonly used to evaluate these vulnerabilities. However, existing research only focuses on embedding-level GIAs, which inject node embeddings rather than actual textual content, limiting their applicability and simplifying detection. In this paper, we pioneer the exploration of GIAs at the text level, presenting three novel attack designs that inject textual content into the graph. Through theoretical and empirical analysis, we demonstrate that text interpretability, a factor previously overlooked at the embedding level, plays a crucial role in attack strength. Among the designs we investigate, the Word-frequency-based Text-level GIA (WTGIA) is particularly notable for its balance between performance and interpretability. Despite the success of WTGIA, we discover that defenders can easily enhance their defenses with customized text embedding methods or large language model (LLM)--based predictors. These insights underscore the necessity for further research into the potential and practical significance of text-level GIAs.

Reinforcement learning has been applied to train the dialog systems in many works. Previous approaches divide the dialog system into multiple modules including DST (dialog state tracking) and DP (dialog policy), and train these modules simultaneously. However, different modules influence each other during training. The errors from DST might misguide the dialog policy, and the system action brings extra difficulties for the DST module. To alleviate this problem, we propose Asynchronous Updating Reinforcement Learning framework (AURL) that updates the DST module and the DP module asynchronously under a cooperative setting. Furthermore, curriculum learning is implemented to address the problem of unbalanced data distribution during reinforcement learning sampling, and multiple user models are introduced to increase the dialog diversity. Results on the public SSD-PHONE dataset show that our method achieves a compelling result with a 31.37% improvement on the dialog success rate. The code is publicly available via https://github.com/shunjiu/AURL.

A slot value might be provided segment by segment over multiple-turn interactions in a dialog, especially for some important information such as phone numbers and names. It is a common phenomenon in daily life, but little attention has been paid to it in previous work. To fill the gap, this paper defines a new task named Sub-Slot based Task-Oriented Dialog (SSTOD) and builds a Chinese dialog dataset SSD for boosting research on SSTOD. The dataset includes a total of 40K dialogs and 500K utterances from four different domains: Chinese names, phone numbers, ID numbers and license plate numbers. The data is well annotated with sub-slot values, slot values, dialog states and actions. We find some new linguistic phenomena and interactive manners in SSTOD which raise critical challenges of building dialog agents for the task. We test three state-of-the-art dialog models on SSTOD and find they cannot handle the task well on any of the four domains. We also investigate an improved model by involving slot knowledge in a plug-in manner. More work should be done to meet the new challenges raised from SSTOD which widely exists in real-life applications. The dataset and code are publicly available via https://github.com/shunjiu/SSTOD.

Quantization can improve the execution latency and energy efficiency of neural networks on both commodity GPUs and specialized accelerators. The majority of existing literature focuses on training quantized DNNs, while this work examines the less-studied topic of quantizing a floating-point model without (re)training. DNN weights and activations follow a bell-shaped distribution post-training, while practical hardware uses a linear quantization grid. This leads to challenges in dealing with outliers in the distribution. Prior work has addressed this by clipping the outliers or using specialized hardware. In this work, we propose outlier channel splitting (OCS), which duplicates channels containing outliers, then halves the channel values. The network remains functionally identical, but affected outliers are moved toward the center of the distribution. OCS requires no additional training and works on commodity hardware. Experimental evaluation on ImageNet classification and language modeling shows that OCS can outperform state-of-the-art clipping techniques with only minor overhead.

We propose unitary group convolutions (UGConvs), a building block for CNNs which compose a group convolution with unitary transforms in feature space to learn a richer set of representations than group convolution alone. UGConvs generalize two disparate ideas in CNN architecture, channel shuffling (i.e. ShuffleNet) and block-circulant networks (i.e. CirCNN), and provide unifying insights that lead to a deeper understanding of each technique. We experimentally demonstrate that dense unitary transforms can outperform channel shuffling in DNN accuracy. On the other hand, different dense transforms exhibit comparable accuracy performance. Based on these observations we propose HadaNet, a UGConv network using Hadamard transforms. HadaNets achieve similar accuracy to circulant networks with lower computation complexity, and better accuracy than ShuffleNets with the same number of parameters and floating-point multiplies.

Scalable frameworks, such as TensorFlow, MXNet, Caffe, and PyTorch drive the current popularity and utility of deep learning. However, these frameworks are optimized for a narrow range of server-class GPUs and deploying workloads to other platforms such as mobile phones, embedded devices, and specialized accelerators (e.g., FPGAs, ASICs) requires laborious manual effort. We propose TVM, an end-to-end optimization stack that exposes graph-level and operator-level optimizations to provide performance portability to deep learning workloads across diverse hardware back-ends. We discuss the optimization challenges specific to deep learning that TVM solves: high-level operator fusion, low-level memory reuse across threads, mapping to arbitrary hardware primitives, and memory latency hiding. Experimental results demonstrate that TVM delivers performance across hardware back-ends that are competitive with state-of-the-art libraries for low-power CPU and server-class GPUs. We also demonstrate TVM's ability to target new hardware accelerator back-ends by targeting an FPGA-based generic deep learning accelerator. The compiler infrastructure is open sourced.