Minyi Guo

Quan Liu, Hongzi Zhu, Yunsong Zhou et al. · pku

Registration of distant outdoor LiDAR point clouds is crucial to extending the 3D vision of collaborative autonomous vehicles, and yet is challenging due to small overlapping area and a huge disparity between observed point densities. In this paper, we propose Group-wise Contrastive Learning (GCL) scheme to extract density-invariant geometric features to register distant outdoor LiDAR point clouds. We mark through theoretical analysis and experiments that, contrastive positives should be independent and identically distributed (i.i.d.), in order to train densityinvariant feature extractors. We propose upon the conclusion a simple yet effective training scheme to force the feature of multiple point clouds in the same spatial location (referred to as positive groups) to be similar, which naturally avoids the sampling bias introduced by a pair of point clouds to conform with the i.i.d. principle. The resulting fully-convolutional feature extractor is more powerful and density-invariant than state-of-the-art methods, improving the registration recall of distant scenarios on KITTI and nuScenes benchmarks by 40.9% and 26.9%, respectively. Code is available at https://github.com/liuQuan98/GCL.

Cong Guo, Chen Zhang, Jingwen Leng et al. · microsoft-research

Quantization is a technique to reduce the computation and memory cost of DNN models, which are getting increasingly large. Existing quantization solutions use fixed-point integer or floating-point types, which have limited benefits, as both require more bits to maintain the accuracy of original models. On the other hand, variable-length quantization uses low-bit quantization for normal values and high-precision for a fraction of outlier values. Even though this line of work brings algorithmic benefits, it also introduces significant hardware overheads due to variable-length encoding and decoding. In this work, we propose a fixed-length adaptive numerical data type called ANT to achieve low-bit quantization with tiny hardware overheads. Our data type ANT leverages two key innovations to exploit the intra-tensor and inter-tensor adaptive opportunities in DNN models. First, we propose a particular data type, flint, that combines the advantages of float and int for adapting to the importance of different values within a tensor. Second, we propose an adaptive framework that selects the best type for each tensor according to its distribution characteristics. We design a unified processing element architecture for ANT and show its ease of integration with existing DNN accelerators. Our design results in 2.8$\times$ speedup and 2.5$\times$ energy efficiency improvement over the state-of-the-art quantization accelerators.

Yue Guan, Zhengyi Li, Jingwen Leng et al. · meta-ai, mila

Transformer architecture has become the de-facto model for many machine learning tasks from natural language processing and computer vision. As such, improving its computational efficiency becomes paramount. One of the major computational inefficiency of Transformer-based models is that they spend the identical amount of computation throughout all layers. Prior works have proposed to augment the Transformer model with the capability of skimming tokens to improve its computational efficiency. However, they suffer from not having effectual and end-to-end optimization of the discrete skimming predictor. To address the above limitations, we propose the Transkimmer architecture, which learns to identify hidden state tokens that are not required by each layer. The skimmed tokens are then forwarded directly to the final output, thus reducing the computation of the successive layers. The key idea in Transkimmer is to add a parameterized predictor before each layer that learns to make the skimming decision. We also propose to adopt reparameterization trick and add skim loss for the end-to-end training of Transkimmer. Transkimmer achieves 10.97x average speedup on GLUE benchmark compared with vanilla BERT-base baseline with less than 1% accuracy degradation.

Cong Guo, Yuxian Qiu, Jingwen Leng et al. · microsoft-research

An activation function is an element-wise mathematical function and plays a crucial role in deep neural networks (DNN). Many novel and sophisticated activation functions have been proposed to improve the DNN accuracy but also consume massive memory in the training process with back-propagation. In this study, we propose the nested forward automatic differentiation (Forward-AD), specifically for the element-wise activation function for memory-efficient DNN training. We deploy nested Forward-AD in two widely-used deep learning frameworks, TensorFlow and PyTorch, which support the static and dynamic computation graph, respectively. Our evaluation shows that nested Forward-AD reduces the memory footprint by up to 1.97x than the baseline model and outperforms the recomputation by 20% under the same memory reduction ratio.

Guan Shen, Jieru Zhao, Quan Chen et al.

The attention mechanisms of transformers effectively extract pertinent information from the input sequence. However, the quadratic complexity of self-attention w.r.t the sequence length incurs heavy computational and memory burdens, especially for tasks with long sequences. Existing accelerators face performance degradation in these tasks. To this end, we propose SALO to enable hybrid sparse attention mechanisms for long sequences. SALO contains a data scheduler to map hybrid sparse attention patterns onto hardware and a spatial accelerator to perform the efficient attention computation. We show that SALO achieves 17.66x and 89.33x speedup on average compared to GPU and CPU implementations, respectively, on typical workloads, i.e., Longformer and ViL.

Mingzhe Gao, Jieru Zhao, Zhe Lin et al.

Recently, the use of large language models (LLMs) for software code generation, e.g., C/C++ and Python, has proven a great success. However, LLMs still suffer from low syntactic and functional correctness when it comes to the generation of register-transfer level (RTL) code, such as Verilog. To address this issue, in this paper, we develop AutoVCoder, a systematic open-source framework that significantly improves the LLMs' correctness of generating Verilog code and enhances the quality of its output at the same time. Our framework integrates three novel techniques, including a high-quality hardware dataset generation approach, a two-round LLM fine-tuning method and a domain-specific retrieval-augmented generation (RAG) mechanism. Experimental results demonstrate that AutoVCoder outperforms both industrial and academic LLMs in Verilog code generation. Specifically, AutoVCoder shows a 0.5% and 2.2% improvement in functional correctness on the EvalMachine and EvalHuman benchmarks compared with BetterV, and also achieves a 3.4% increase in syntax correctness and a 3.4% increase in functional correctness on the RTLLM benchmark compared with RTLCoder.

Yunsong Zhou, Hongzi Zhu, Quan Liu et al.

Mobile monocular 3D object detection (Mono3D) (e.g., on a vehicle, a drone, or a robot) is an important yet challenging task. Existing transformer-based offline Mono3D models adopt grid-based vision tokens, which is suboptimal when using coarse tokens due to the limited available computational power. In this paper, we propose an online Mono3D framework, called MonoATT, which leverages a novel vision transformer with heterogeneous tokens of varying shapes and sizes to facilitate mobile Mono3D. The core idea of MonoATT is to adaptively assign finer tokens to areas of more significance before utilizing a transformer to enhance Mono3D. To this end, we first use prior knowledge to design a scoring network for selecting the most important areas of the image, and then propose a token clustering and merging network with an attention mechanism to gradually merge tokens around the selected areas in multiple stages. Finally, a pixel-level feature map is reconstructed from heterogeneous tokens before employing a SOTA Mono3D detector as the underlying detection core. Experiment results on the real-world KITTI dataset demonstrate that MonoATT can effectively improve the Mono3D accuracy for both near and far objects and guarantee low latency. MonoATT yields the best performance compared with the state-of-the-art methods by a large margin and is ranked number one on the KITTI 3D benchmark.

Xinwei Qiang, Hongmin Chen, Shixuan Sun et al.

Determinism is indispensable for reproducibility in large language model (LLM) training, yet it often exacts a steep performance cost. In widely used attention implementations such as FlashAttention-3, the deterministic backward pass can incur up to a 37.9% throughput reduction relative to its non-deterministic counterpart, primarily because gradient accumulation operations must be serialized to guarantee numerical consistency. This performance loss stems from suboptimal scheduling of compute and gradient-reduction phases, leading to significant hardware underutilization. To address this challenge, we formulate the backward pass of deterministic attention as a scheduling problem on a Directed Acyclic Graph (DAG) and derive schedules that minimize the critical path length. Building on this formulation, we present DASH (Deterministic Attention Scheduling for High-Throughput), which encapsulates two complementary scheduling strategies: (i) Descending Q-Tile Iteration, a reversed query-block traversal that shrinks pipeline stalls in causal attention, and (ii) Shift Scheduling, a theoretically optimal schedule within our DAG model that reduces pipeline stalls for both full and causal masks. Our empirical evaluations on NVIDIA H800 GPUs demonstrate that DASH narrows the performance gap of deterministic attention. The proposed strategies improve the throughput of the attention backward pass by up to 1.28$\times$ compared to the baseline, significantly advancing the efficiency of reproducible LLM training. Our code is open-sourced at https://github.com/SJTU-Liquid/deterministic-FA3.

Jiale Xu, Rui Zhang, Yi Xiong et al.

Large Language Models are increasingly being deployed in datacenters. Serving these models requires careful memory management, as their memory usage includes static weights, dynamic activations, and key-value caches. While static weights are constant and predictable, dynamic components such as activations and KV caches change frequently during runtime, presenting significant challenges for efficient memory management. Modern LLM serving systems typically handle runtime memory and KV caches at distinct abstraction levels: runtime memory management relies on static tensor abstractions, whereas KV caches utilize a page table-based virtualization layer built on top of the tensor abstraction. This virtualization dynamically manages KV caches to mitigate memory fragmentation. However, this dual-level approach fundamentally isolates runtime memory and KV cache management, resulting in suboptimal memory utilization under dynamic workloads, which can lead to a nearly 20% drop in throughput. To address these limitations, we propose eLLM, an elastic memory management framework inspired by the classical memory ballooning mechanism in operating systems. The core components of eLLM include: (1) Virtual Tensor Abstraction, which decouples the virtual address space of tensors from the physical GPU memory, creating a unified and flexible memory pool; (2) an Elastic Memory Mechanism that dynamically adjusts memory allocation through runtime memory inflation and deflation, leveraging CPU memory as an extensible buffer; and (3) a Lightweight Scheduling Strategy employing SLO-aware policies to optimize memory utilization and effectively balance performance trade-offs under stringent SLO constraints. Comprehensive evaluations demonstrate that eLLM significantly outperforms state-of-the-art systems, 2.32x higher decoding throughput, and supporting 3x larger batch sizes for 128K-token inputs.

Zhuoshan Zhou, Chen Zhang, Shuyi Zhang et al.

The Mixture-of-Experts (MoE) architecture is crucial for scaling large language models, but its scalability is severely limited by inter-GPU communication bottlenecks in multi-GPU systems. Although overlapping communication with computation is a widely recognized optimization, its effective deployment still remains challenging, both in terms of performance and programmability. In this work, we identify the root cause as a fundamental abstraction mismatch between MoE's dynamic, irregular token-to-expert mapping and the static, address-centric communication model of modern GPUs, which necessitates a complex software mediation phase to resolve addresses before data transfers, limiting performance and software flexibility. To resolve this, we propose MoE-Hub, a hardware-software co-design that introduces a destination-agnostic communication paradigm. MoE-Hub decouples data transmission from address management, allowing producers to send data immediately after routing using only a logical destination, while address allocation and data-flow orchestration are handled transparently by lightweight hardware in the GPU hub. By hardware-accelerating the entire communication control plane, MoE-Hub enables seamless and transparent overlap. Our evaluation shows that MoE-Hub achieves 1.40x-3.08x per-layer and 1.21x-1.98x end-to-end speedup over state-of-the-art systems.

Yunsong Zhou, Quan Liu, Hongzi Zhu et al.

Monocular 3D object detection (Mono3D) in mobile settings (e.g., on a vehicle, a drone, or a robot) is an important yet challenging task. Due to the near-far disparity phenomenon of monocular vision and the ever-changing camera pose, it is hard to acquire high detection accuracy, especially for far objects. Inspired by the insight that the depth of an object can be well determined according to the depth of the ground where it stands, in this paper, we propose a novel Mono3D framework, called MoGDE, which constantly estimates the corresponding ground depth of an image and then utilizes the estimated ground depth information to guide Mono3D. To this end, we utilize a pose detection network to estimate the pose of the camera and then construct a feature map portraying pixel-level ground depth according to the 3D-to-2D perspective geometry. Moreover, to improve Mono3D with the estimated ground depth, we design an RGB-D feature fusion network based on the transformer structure, where the long-range self-attention mechanism is utilized to effectively identify ground-contacting points and pin the corresponding ground depth to the image feature map. We conduct extensive experiments on the real-world KITTI dataset. The results demonstrate that MoGDE can effectively improve the Mono3D accuracy and robustness for both near and far objects. MoGDE yields the best performance compared with the state-of-the-art methods by a large margin and is ranked number one on the KITTI 3D benchmark.

Kangkang Sun, Jun Wu, Minyi Guo et al.

Federated Learning (FL) enables collaborative model training without data sharing, yet participants face a fundamental challenge, e.g., simultaneously ensuring fairness across demographic groups while protecting sensitive client data. We introduce a differentially private fair FL algorithm (\textit{FedPF}) that transforms this multi-objective optimization into a zero-sum game where fairness and privacy constraints compete against model utility. Our theoretical analysis reveals a surprising inverse relationship, i.e., stricter privacy protection fundamentally limits the system's ability to detect and correct demographic biases, creating an inherent tension between privacy and fairness. Counterintuitively, we prove that moderate fairness constraints initially improve model generalization before causing performance degradation, where a non-monotonic relationship that challenges conventional wisdom about fairness-utility tradeoffs. Experimental validation demonstrates up to 42.9 % discrimination reduction across three datasets while maintaining competitive accuracy, but more importantly, reveals that the privacy-fairness tension is unavoidable, i.e., achieving both objectives simultaneously requires carefully balanced compromises rather than optimization of either in isolation. The source code for our proposed algorithm is publicly accessible at https://github.com/szpsunkk/FedPF.

Jiale Xu, Rui Zhang, Cong Guo et al.

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.

Guan Shen, Jieru Zhao, Zeke Wang et al.

Along with the fast evolution of deep neural networks, the hardware system is also developing rapidly. As a promising solution achieving high scalability and low manufacturing cost, multi-accelerator systems widely exist in data centers, cloud platforms, and SoCs. Thus, a challenging problem arises in multi-accelerator systems: selecting a proper combination of accelerators from available designs and searching for efficient DNN mapping strategies. To this end, we propose MARS, a novel mapping framework that can perform computation-aware accelerator selection, and apply communication-aware sharding strategies to maximize parallelism. Experimental results show that MARS can achieve 32.2% latency reduction on average for typical DNN workloads compared to the baseline, and 59.4% latency reduction on heterogeneous models compared to the corresponding state-of-the-art method.

Jiawen Wang, Quan Chen, Deze Zeng et al.

Many emerging user-facing services adopt Graph Neural Networks (GNNs) to improve serving accuracy. When the graph used by a GNN model changes, representations (embedding) of nodes in the graph should be updated accordingly. However, the node representation update is too slow, resulting in either long response latency of user queries (the inference is performed after the update completes) or high staleness problem (the inference is performed based on stale data). Our in-depth analysis shows that the slow update is mainly due to neighbor explosion problem in graphs and duplicated computation. Based on such findings, we propose STAG, a GNN serving framework that enables low latency and low staleness of GNN-based services. It comprises a collaborative serving mechanism and an additivity-based incremental propagation strategy. With the collaborative serving mechanism, only part of node representations are updated during the update phase, and the final representations are calculated in the inference phase. It alleviates the neighbor explosion problem. The additivity-based incremental propagation strategy reuses intermediate data during the update phase, eliminating duplicated computation problem. Experimental results show that STAG accelerates the update phase by 1.3x~90.1x, and greatly reduces staleness time with a slight increase in response latency.

Zhengyi Li, Cong Guo, Zhanda Zhu et al.

Post-training quantization attracts increasing attention due to its convenience in deploying quantized neural networks. Although rounding-to-nearest remains the prevailing method for DNN quantization, prior research has demonstrated its suboptimal nature when applied to weight quantization. They propose optimizing weight rounding schemes by leveraging output error rather than the traditional weight quantization error. Our study reveals that similar rounding challenges also extend to activation quantization. Despite the easy generalization, the challenges lie in the dynamic nature of activation. Adaptive rounding is expected for varying activations and the method is subjected to runtime overhead. To tackle this, we propose the AQuant quantization framework with a novel perspective to reduce output error by adjusting rounding schemes of activations. Instead of using the constant rounding border 0.5 of the rounding-to-nearest operation, we make the border become a function w.r.t. the activation value to change the activation rounding by the adaptive border. To deal with the runtime overhead, we use a coarse-grained version of the border function. Finally, we introduce our framework to optimize the border function. Extensive experiments show that AQuant achieves notable improvements compared to state-of-the-art works and pushes the accuracy of ResNet-18 up to 60.31% under the 2-bit weight and activation quantization.

Weichuang Zhang, Yiquan Wang, Xinzhou Zhang et al.

FPGAs are well-suited for dataflow architectures that process data in a streaming or pipelined manner, thus satisfying the high computational and communication demands of emerging applications. However, manually implementing an efficient dataflow architecture for large-scale applications is still challenging, even for specialists who use high-level synthesis (HLS) to simplify FPGA programming. To address this, we introduce CODO, an automated compiler that generates feasible and efficient dataflow accelerators on FPGAs. CODO features a systematic method for detecting and eliminating both coarse-grained and fine-grained dataflow violations. Building on this, CODO performs both on- and off-chip data movement optimizations to maximize transfer efficiency. To guarantee a higher design quality, CODO performs automatic scheduling to generate high-performance dataflow accelerators, ensuring a balanced performance-resource trade-off. Synthesis results show that CODO delivers $1.45\times$ to $4.52\times$ latency speedups on typical computation kernels and $3.7\times$ to $33.8\times$ speedups on DNN models compared to SOTA frameworks. In on-board evaluations, CODO achieves $7.3\times$ average speedup on CNN models and $2.07\times$ average speedup on the GPT-2 model over SOTA frameworks. The compiler is open-sourced at https://github.com/sjtu-zhao-lab/codo-artifact.

Shuwen Lu, Zhihui Zhang, Cong Guo et al.

Graph neural networks (GNNs) have shown significant accuracy improvements in a variety of graph learning domains, sparking considerable research interest. To translate these accuracy improvements into practical applications, it is essential to develop high-performance and efficient hardware acceleration for GNN models. However, designing GNN accelerators faces two fundamental challenges: the high bandwidth requirement of GNN models and the diversity of GNN models. Previous works have addressed the first challenge by using more expensive memory interfaces to achieve higher bandwidth. For the second challenge, existing works either support specific GNN models or have generic designs with poor hardware utilization. In this work, we tackle both challenges simultaneously. First, we identify a new type of partition-level operator fusion, which we utilize to internally reduce the high bandwidth requirement of GNNs. Next, we introduce partition-level multi-threading to schedule the concurrent processing of graph partitions, utilizing different hardware resources. To further reduce the extra on-chip memory required by multi-threading, we propose fine-grained graph partitioning to generate denser graph partitions. Importantly, these three methods make no assumptions about the targeted GNN models, addressing the challenge of model variety. We implement these methods in a framework called SwitchBlade, consisting of a compiler, a graph partitioner, and a hardware accelerator. Our evaluation demonstrates that SwitchBlade achieves an average speedup of $1.85\times$ and energy savings of $19.03\times$ compared to the NVIDIA V100 GPU. Additionally, SwitchBlade delivers performance comparable to state-of-the-art specialized accelerators.

Yue Guan, Changming Yu, Shihan Fang et al.

Speculative decoding improves LLM inference by generating and verifying multiple tokens in parallel, but existing systems suffer from suboptimal performance due to a mismatch between dynamic speculation and static runtime assumptions. We present Yggdrasil, a co-designed system that enables latency-optimal speculative decoding through context-aware tree drafting and compiler-friendly execution. Yggdrasil introduces an equal-growth tree structure for static graph compatibility, a latency-aware optimization objective for draft selection, and stage-based scheduling to reduce overhead. Yggdrasil supports unmodified LLMs and achieves up to $3.98\times$ speedup over state-of-the-art baselines across multiple hardware setups.

Zhenyu Ning, Jieru Zhao, Qihao Jin et al.

Multimodal Large Language Models (MLLMs) are distinguished by their multimodal comprehensive ability and widely used in many real-world applications including GPT-4o, autonomous driving and robotics. Despite their impressive performance, the multimodal inputs always incur long context. The inference under long context requires caching massive Key and Value states (KV cache) of previous tokens, which introduces high latency and excessive memory consumption. Due to this reason, it is challenging to deploy streaming inference of MLLMs on edge devices, which largely constrains the power and usage of MLLMs in real-world applications. In this paper, we introduce Inf-MLLM, an efficient inference framework for MLLMs, which enable streaming inference of MLLM on a single GPU with infinite context. Inf-MLLM is based on our key observation of the attention pattern in both LLMs and MLLMs called "attention saddles". Thanks to the newly discovered attention pattern, Inf-MLLM maintains a size-constrained KV cache by dynamically caching recent tokens and relevant tokens. Furthermore, Inf-MLLM proposes attention bias, a novel approach to enable MLLMs to capture long-term dependency. We show that Inf-MLLM enables multiple LLMs and MLLMs to achieve stable performance over 4M-token long texts and multi-round conversations with 1-hour-long videos on a single GPU. In addition, Inf-MLLM exhibits superior streaming reasoning quality than existing methods such as StreamingLLM and 2x speedup than H2O.

Kangkang Sun, Jianhua Li, Xiuzhen Chen et al.

Federated Learning (FL) has emerged as a prominent paradigm for privacy-preserving distributed machine learning, yet two fundamental challenges hinder its large-scale adoption. First, gradient inversion attacks can reconstruct sensitive training data from uploaded model updates, so privacy risk persists even when raw data remain local. Second, without adequate monetary compensation, rational clients have little incentive to contribute high-quality gradients, limiting participation at scale. To address these challenges, a privacy trading market is developed in which clients sell their differential privacy budgets as a commodity and receive explicit economic compensation for privacy sacrifice. This market is formalized as a Privacy Auction Game (PAG), and the existence of a Bayesian Nash Equilibrium is established under dominant-strategy incentive compatibility (DSIC), individual rationality (IR), and budget feasibility. To overcome the NP-hard, high-dimensional Nash Equilibrium computation at scale, \textit{MFG-RegretNet} is introduced as a deep-learning-based auction mechanism that combines mean-field game (MFG) approximation with differentiable mechanism design. The MFG reduction lowers per-round computational complexity from $\mathcal{O}(N^2 \log N)$ to $\mathcal{O}(N)$ while incurring only an $\mathcal{O}(N^{-1/2})$ equilibrium approximation gap. Extensive experiments on MNIST and CIFAR-10 demonstrate that MFG-RegretNet outperforms state-of-the-art baselines in incentive compatibility, auction revenue, and social welfare, while maintaining competitive downstream FL model accuracy.

Chen Zhang, Qijun Zhang, Zhuoshan Zhou et al.

Tensor parallelism (TP) in large-scale LLM inference and training introduces frequent collective operations that dominate inter-GPU communication. While in-switch computing, exemplified by NVLink SHARP (NVLS), accelerates collective operations by reducing redundant data transfer, its communication-centric design philosophy introduces the mismatch between its communication mode and the memory semantic requirement of LLM's computation kernel. Such a mismatch isolates the compute and communication phases, resulting in underutilized resources and limited overlap in multi-GPU systems. To address the limitation, we propose CAIS, the first Compute-Aware In-Switch computing framework that aligns communication modes with computation's memory semantics requirement. CAIS consists of three integral techniques: (1) compute-aware ISA and microarchitecture extension to enable compute-aware in-switch computing. (2) merging-aware TB (Thread Block) coordination to improve the temporal alignment for efficient request merging. (3) graph-level dataflow optimizer to achieve a tight cross-kernel overlap. Evaluations on LLM workloads show that CAIS achieves 1.38$\times$ average end-to-end training speedup over the SOTA NVLS-enabled solution, and 1.61$\times$ over T3, the SOTA compute-communicate overlap solutions but do not leverage NVLS, demonstrating its effectiveness in accelerating TP on multi-GPU systems.

Guangda Liu, Chengwei Li, Jieru Zhao et al.

Large Language Models (LLMs) have been widely deployed in a variety of applications, and the context length is rapidly increasing to handle tasks such as long-document QA and complex logical reasoning. However, long context poses significant challenges for inference efficiency, including high memory costs of key-value (KV) cache and increased latency due to extensive memory accesses. Recent works have proposed compressing KV cache to approximate computation, but these methods either evict tokens permanently, never recalling them for later inference, or recall previous tokens at the granularity of pages divided by textual positions. Both approaches degrade the model accuracy and output quality. To achieve efficient and accurate recallable KV cache compression, we introduce ClusterKV, which recalls tokens at the granularity of semantic clusters. We design and implement efficient algorithms and systems for clustering, selection, indexing and caching. Experiment results show that ClusterKV attains negligible accuracy loss across various tasks with 32k context lengths, using only a 1k to 2k KV cache budget, and achieves up to a 2$\times$ speedup in latency and a 2.5$\times$ improvement in decoding throughput. Compared to SoTA recallable KV compression methods, ClusterKV demonstrates higher model accuracy and output quality, while maintaining or exceeding inference efficiency. Our code is available at https://github.com/sjtu-zhao-lab/ClusterKV.

Jiacheng Liu, Peng Tang, Wenfeng Wang et al.

The emergence of large-scale Mixture of Experts (MoE) models represents a significant advancement in artificial intelligence, offering enhanced model capacity and computational efficiency through conditional computation. However, deploying and running inference on these models presents significant challenges in computational resources, latency, and energy efficiency. This comprehensive survey analyzes optimization techniques for MoE models across the entire system stack. We first establish a taxonomical framework that categorizes optimization approaches into model-level, system-level, and hardware-level optimizations. At the model level, we examine architectural innovations including efficient expert design, attention mechanisms, various compression techniques such as pruning, quantization, and knowledge distillation, as well as algorithm improvement including dynamic routing strategies and expert merging methods. At the system level, we investigate distributed computing approaches, load balancing mechanisms, and efficient scheduling algorithms that enable scalable deployment. Furthermore, we delve into hardware-specific optimizations and co-design strategies that maximize throughput and energy efficiency. This survey provides both a structured overview of existing solutions and identifies key challenges and promising research directions in MoE inference optimization. To facilitate ongoing updates and the sharing of cutting-edge advances in MoE inference optimization research, we have established a repository accessible at https://github.com/MoE-Inf/awesome-moe-inference/.

Weiye Wang, Chen Chen, Junxue Zhang et al.

Distributed prefix caching has become a core technique for efficient LLM serving. However, for long-context requests with high cache hit ratios, retrieving reusable KVCache blocks from remote servers has emerged as a new performance bottleneck. Such network-intensive LLM inference is expected to become increasingly common as agentic AI workloads continue to grow. However, existing LLM inference engines remain largely compute-centric: they treat KVCache loading as a subordinate phase to GPU execution and often fail to account for its delay explicitly during scheduling. We present CALVO, an LLM serving engine that treats KVCache loading as a first-class concern. CALVO decouples KVCache loading and GPU computation into independently managed, asynchronously progressing stages, enabling better utilization of network, PCIe, and computation resources. In addition, CALVO incorporates KVCache loading delay as an explicit component of per-request service cost, leading to more accurate scheduling decisions. Experiments on a real testbed with diverse long-context workloads show that CALVO substantially improves the efficiency of network-intensive LLM inference, achieving up to 61.67% higher SLO attainment than the baseline.

Kangkang Sun, Jun Wu, Jianhua Li et al.

Uncertainty estimation in multi-LLM systems remains largely single-model-centric: existing methods quantify uncertainty within each model but do not adequately capture semantic disagreement across models. To address this gap, we propose Collaborative Entropy (CoE), a unified information-theoretic metric for semantic uncertainty in multi-LLM collaboration. CoE is defined on a shared semantic cluster space and combines two components: intra-model semantic entropy and inter-model divergence to the ensemble mean. CoE is not a weighted ensemble predictor; it is a system-level uncertainty measure that characterizes collaborative confidence and disagreement. We analyze several core properties of CoE, including non-negativity, zero-value certainty under perfect semantic consensus, and the behavior of CoE when individual models collapse to delta distributions. These results clarify when reducing per-model uncertainty is sufficient and when residual inter-model disagreement remains. We also present a simple CoE-guided, training-free post-hoc coordination heuristic as a practical application of the metric. Experiments on \textit{TriviaQA} and \textit{SQuAD} with LLaMA-3.1-8B-Instruct, Qwen-2.5-7B-Instruct, and Mistral-7B-Instruct show that CoE provides stronger uncertainty estimation than standard entropy- and divergence-based baselines, with gains becoming larger as additional heterogeneous models are introduced. Overall, CoE offers a useful uncertainty-aware perspective on multi-LLM collaboration.

Ao Xu, Han Zhao, Weihao Cui et al.

Large language models (LLMs) are increasingly deployed under the Model-as-a-Service (MaaS) paradigm. To meet stringent quality-of-service (QoS) requirements, existing LLM serving systems disaggregate the prefill and decode phases of inference. However, decode instances often experience low GPU utilization due to their memory-bound nature and insufficient batching in dynamic workloads, leaving compute resources underutilized. We introduce Harli, a serving system that improves GPU utilization by co-locating parameter-efficient finetuning (PEFT) tasks with LLM decode instances. PEFT tasks are compute-bound and memory-efficient, making them ideal candidates for safe co-location. Specifically, Harli addresses key challenges--limited memory and unpredictable interference--using three components: a unified memory allocator for runtime memory reuse, a two-stage latency predictor for decode latency modeling, and a QoS-guaranteed throughput-maximizing scheduler for throughput maximization. Experimental results show that Harli improves the finetune throughput by 46.2% on average (up to 92.0%) over state-of-the-art serving systems, while maintaining strict QoS guarantees for inference decode.

Wenfeng Wang, Xiaofeng Hou, Peng Tang et al.

Retrieval-Augmented Generation (RAG) systems enhance the performance of large language models (LLMs) by incorporating supplementary retrieved documents, enabling more accurate and context-aware responses. However, integrating these external documents often results in very long input sequences, which significantly increases computation costs during the prefill stage, where key-value (KV) representations for all input tokens are generated. This latency bottleneck becomes especially pronounced under high-throughput serving scenarios. KV-cache reuse offers a promising solution by storing previously computed KV states for shared input prefixes, thereby avoiding redundant computation across requests that contain overlapping context. Yet, the effectiveness of cache reuse is often limited by three practical challenges: low cache hit rates due to naive eviction policies, high CPU-GPU data transfer overhead, and slow SSD I/O when caches spill to storage. To address these issues, we propose PCR, a system designed to maximize KV-cache reuse efficiency through intelligent prefetching and pipelined data movement. Specifically, PCR introduces three key techniques: (1) a prefix-tree caching structure with a look-ahead LRU replacement policy that uses pending requests in the scheduler queue to improve cache hit ratios; (2) layer-wise overlapping that pipelines KV-cache loading and GPU computation across CUDA streams to hide communication latency; and (3) queue-based prefetching that proactively loads relevant KV caches from SSD into DRAM before they are needed. Extensive experiments show that PCR outperforms existing KV-cache reuse methods, achieving up to a 2.47x speedup in terms of average TTFT.

Boxiao Du, Boning Huangfu, Yizhou Luo et al.

Large Language Models (LLMs) play a critical role in emerging agentic applications, where the timely completion of each entire inference is critical. Meanwhile, agentic LLM inferences are increasingly served on heterogeneous GPUs in operator's resource pools. Therefore, it is crucial to route incoming inference requests to appropriate GPUs so that their end-to-end latency requirements are satisfied whenever possible, thereby achieving high goodput. In this paper, we propose GoodServe, a goodput-optimized serving system for agentic inferences over heterogeneous resources. GoodServe performs inference routing in a predict-and-rectify manner. It estimates the request output lengths as well as the GPU serving status in an accurate and also practical manner. Based on information from both the demand and resource sides, it then makes high-quality routing decisions using a just-enough instance selection heuristic. It also periodically monitors SLO-violation risks of active requests and triggers runtime request migrations to address unexpected dynamics. Our evaluations show that GoodServe improves goodput by up to 27.4% over existing routing methods.

Kangkang Sun, Junyi He, Juntong Liu et al.

Autonomous platoons traversing infrastructure gaps increasingly depend on LEO satellite backhaul for safety-critical updates, yet no existing framework jointly addresses compound Doppler from simultaneous satellite and vehicle motion, sub-slot handover outages that exceed collision-alert deadlines, and heterogeneous freshness requirements across three vehicular priority classes. The core challenge is a \emph{timescale mismatch}: coarse control slots hide sub-slot outages, which makes both AoI spike analysis and safety verification ill-posed. Ping-pong handover oscillations further compound AoI cost in a way that purely reactive schedulers cannot mitigate. We address these challenges through a unified framework that couples a two-timescale AoI model with tiered time-average safety constraints enforced by virtual queues. A closed-form ping-pong AoI envelope reveals that cumulative penalty grows quadratically in oscillation length, analytically justifying oscillation suppression as the highest-leverage safety mechanism. The resulting drift-plus-penalty template is instantiated as SafeScale-MATD3 with proactive handover timing and multi-task dual-critic MARL. A key finding is that suppressing brief but repeated ping-pong oscillations yields larger safety returns than shortening any single outage, and that tick-level AoI accounting is a necessary condition for verifiable collision-alert guarantees under LEO handovers. Simulations show that SafeScale-MATD3 is the only method satisfying the strict 1 % collision-alert violation budget, reducing violation rate by 4 to 5.5 times versus baselines, while achieving 35 % lower collision-alert AoI and strict Pareto dominance on the energy and freshness tradeoff.

Yunzhe Li, Jiajun Yan, Yuzhou Wei et al.

Failing to be aware of speeding vehicles approaching from behind poses a huge threat to the road safety of pedestrians and cyclists. In this paper, we propose BlinkBud, which utilizes a single earbud and a paired phone to online detect hazardous objects approaching from behind of a user. The core idea is to accurately track visually identified objects utilizing a small number of sampled camera images taken from the earbud. To minimize the power consumption of the earbud and the phone while guaranteeing the best tracking accuracy, a novel 3D object tracking algorithm is devised, integrating both a Kalman filter based trajectory estimation scheme and an optimal image sampling strategy based on reinforcement learning. Moreover, the impact of constant user head movements on the tracking accuracy is significantly eliminated by leveraging the estimated pitch and yaw angles to correct the object depth estimation and align the camera coordinate system to the user's body coordinate system, respectively. We implement a prototype BlinkBud system and conduct extensive real-world experiments. Results show that BlinkBud is lightweight with ultra-low mean power consumptions of 29.8 mW and 702.6 mW on the earbud and smartphone, respectively, and can accurately detect hazards with a low average false positive ratio (FPR) and false negative ratio (FNR) of 4.90% and 1.47%, respectively.

Xinhao Luo, Zihan Liu, Yangjie Zhou et al.

Large language model (LLM) decoding suffers from high latency due to fragmented execution across operators and heavy reliance on off-chip memory for data exchange and reduction. This execution model limits opportunities for fusion and incurs significant memory traffic and kernel launch overhead. While modern architectures such as NVIDIA Hopper provide distributed shared memory and low-latency intra-cluster interconnects, they expose only low-level data movement instructions, lacking structured abstractions for collective on-chip communication. To bridge this software-hardware gap, we introduce two cluster-level communication primitives, ClusterReduce and ClusterGather, which abstract common communication patterns and enable structured, high-speed data exchange and reduction between thread blocks within a cluster, allowing intermediate results to be on-chip without involving off-chip memory. Building on these abstractions, we design ClusterFusion, an execution framework that schedules communication and computation jointly to expand operator fusion scope by composing decoding stages such as QKV Projection, Attention, and Output Projection into a single fused kernels. Evaluations on H100 GPUs show that ClusterFusion outperforms state-of-the-art inference frameworks by 1.61x on average in end-to-end latency across different models and configurations. The source code is available at https://github.com/xinhao-luo/ClusterFusion.

Weitian Chen, Shixuan Sun, Cheng Chen et al.

Subgraph matching is a core operation in graph analytics, supporting a broad spectrum of applications from social network analysis to bioinformatics. Recent GPU-based approaches accelerate subgraph matching by leveraging parallelism but rely on a coarse-grained execution model that suffers from scalability and efficiency issues due to high memory overhead and thread underutilization. In this paper, we propose gMatch, a hardware-efficient subgraph matching approach on GPUs. gMatch introduces a fine-grained execution model that reduces memory consumption and enables flexible task scheduling among threads. We further design warp-level batch exploration and lightweight load balancing to improve execution efficiency and scalability. Experiments on diverse workloads and real-world datasets show that gMatch outperforms state-of-the-art subgraph matching methods, including STMatch, T-DFS, and EGSM, in both performance and scalability. We also compare against state-of-the-art systems for mining small patterns, such as BEEP and G$^2$Miner. While these systems achieve better performance on small datasets, gMatch scales to substantially larger queries and datasets, where existing approaches degrade or fail to complete.

Di Liu, Yifei Liu, Chen Chen et al.

With the increasing volumes of Large Language Models (LLMs) and the expanding context lengths, attention computation has become a key performance bottleneck in LLM serving. For fast attention computation, recent practices often parallelize the attention heads on multiple GPUs, and also widely adopt attention sparsification to reduce the computation amount -- which selectively computes a subset of attention pairs under a preset sparsity budget. In this paper, we notice that attention heads of an LLM model often exhibit heterogeneous-yet-stable sparsity elasticities, which motivates us to enforce head-adaptive sparsity budgets to attain better efficiency while preserving high inference quality. Yet, from the system aspect, with heterogeneous sparsity levels, attention computation time on different heads would be inconsistent, yielding cross-GPU resource bubbles under head-parallel deployment. To further minimize such bubbles, we propose a novel attention deployment strategy called Sparsity-aware Head-Parallel Load Balance (S-HPLB). Experiments on long-context benchmark show that, S-HPLB can achieve a $2.88\times$ improvement in average attention computation latency without quality degradation.

Yunzhe Li, Jianan Wang, Hongzi Zhu et al.

Large Language Models (LLMs) have become foundational components in a wide range of applications, including natural language understanding and generation, embodied intelligence, and scientific discovery. As their computational requirements continue to grow, these models are increasingly deployed as cloud-based services, allowing users to access powerful LLMs via the Internet. However, this deployment model introduces a new class of threat: denial-of-service (DoS) attacks via unbounded reasoning, where adversaries craft specially designed inputs that cause the model to enter excessively long or infinite generation loops. These attacks can exhaust backend compute resources, degrading or denying service to legitimate users. To mitigate such risks, many LLM providers adopt a closed-source, black-box setting to obscure model internals. In this paper, we propose ThinkTrap, a novel input-space optimization framework for DoS attacks against LLM services even in black-box environments. The core idea of ThinkTrap is to first map discrete tokens into a continuous embedding space, then undertake efficient black-box optimization in a low-dimensional subspace exploiting input sparsity. The goal of this optimization is to identify adversarial prompts that induce extended or non-terminating generation across several state-of-the-art LLMs, achieving DoS with minimal token overhead. We evaluate the proposed attack across multiple commercial, closed-source LLM services. Our results demonstrate that, even far under the restrictive request frequency limits commonly enforced by these platforms, typically capped at ten requests per minute (10 RPM), the attack can degrade service throughput to as low as 1% of its original capacity, and in some cases, induce complete service failure.

Di Liu, Ruitian Wang, Chen Chen et al.

As large language models scale to longer contexts, loading the growing KV cache during attention computation becomes a critical bottleneck. Previous work has shown that attention computation is dominated by a small subset of tokens. This motivates block sparse attention methods that partition the KV cache into fixed-size blocks and selectively compute attention over those blocks exhibiting high importance. However, these methods assign a uniform block size across all attention heads, implicitly assuming homogeneous behavior throughout the model. Our analysis reveals that this assumption is flawed: attention heads exhibit widely varying sensitivity to block granularity, and uniformity leads to suboptimal accuracy. We present AB-Sparse, a training-free algorithm-system co-designed framework that improves accuracy while preserving throughput. AB-Sparse introduces lightweight adaptive block size allocation across attention heads to improve accuracy. To compensate for the additional memory overhead, it further employs lossless block centroid quantization. In addition, custom GPU kernels are developed to support efficient execution with variable block sizes. Evaluation results demonstrate that AB-Sparse achieves an accuracy improvement of up to 5.43% over existing block sparse attention baselines without throughput overhead.

Quan Liu, Yunsong Zhou, Hongzi Zhu et al.

For many driving safety applications, it is of great importance to accurately register LiDAR point clouds generated on distant moving vehicles. However, such point clouds have extremely different point density and sensor perspective on the same object, making registration on such point clouds very hard. In this paper, we propose a novel feature extraction framework, called APR, for online distant point cloud registration. Specifically, APR leverages an autoencoder design, where the autoencoder reconstructs a denser aggregated point cloud with several frames instead of the original single input point cloud. Our design forces the encoder to extract features with rich local geometry information based on one single input point cloud. Such features are then used for online distant point cloud registration. We conduct extensive experiments against state-of-the-art (SOTA) feature extractors on KITTI and nuScenes datasets. Results show that APR outperforms all other extractors by a large margin, increasing average registration recall of SOTA extractors by 7.1% on LoKITTI and 4.6% on LoNuScenes. Code is available at https://github.com/liuQuan98/APR.

Cong Guo, Yuxian Qiu, Jingwen Leng et al.

Quantization of deep neural networks (DNN) has been proven effective for compressing and accelerating DNN models. Data-free quantization (DFQ) is a promising approach without the original datasets under privacy-sensitive and confidential scenarios. However, current DFQ solutions degrade accuracy, need synthetic data to calibrate networks, and are time-consuming and costly. This paper proposes an on-the-fly DFQ framework with sub-second quantization time, called SQuant, which can quantize networks on inference-only devices with low computation and memory requirements. With the theoretical analysis of the second-order information of DNN task loss, we decompose and approximate the Hessian-based optimization objective into three diagonal sub-items, which have different areas corresponding to three dimensions of weight tensor: element-wise, kernel-wise, and output channel-wise. Then, we progressively compose sub-items and propose a novel data-free optimization objective in the discrete domain, minimizing Constrained Absolute Sum of Error (or CASE in short), which surprisingly does not need any dataset and is even not aware of network architecture. We also design an efficient algorithm without back-propagation to further reduce the computation complexity of the objective solver. Finally, without fine-tuning and synthetic datasets, SQuant accelerates the data-free quantization process to a sub-second level with >30% accuracy improvement over the existing data-free post-training quantization works, with the evaluated models under 4-bit quantization. We have open-sourced the SQuant framework at https://github.com/clevercool/SQuant.

Xing Ma, Yangjie Zhou, Wu Sun et al.

Efficient CUDA implementations of attention mechanisms are critical to modern deep learning systems, yet supporting diverse and evolving attention variants remains challenging. Existing frameworks and compilers trade performance for flexibility, while expert-written kernels achieve high efficiency but are difficult to adapt. Recent work explores large language models (LLMs) for GPU kernel generation, but prior studies report unstable correctness and significant performance gaps for complex operators such as attention. We present CuBridge, an LLM-based framework that adapts expert-written attention kernels through a structured lift-transfer-lower workflow. CuBridge starts from expert-written CUDA attention kernels and lifts them into an executable intermediate representation that makes execution orchestration explicit while abstracting low-level CUDA syntax. Given a user-provided PyTorch specification, CuBridge generates and verifies a target IR program, then reconstructs optimized CUDA code via reference-guided lowering. Across diverse attention variants and GPU platforms, CuBridge consistently produces correct kernels and substantially outperforms general frameworks, compiler-based approaches, and prior LLM-based methods.

Qijun Zhang, Chen Zhang, Zhuoshan Zhou et al.

Mixture-of-Experts (MoE) has been adopted by many leading large models to reduce computational requirements. However, frequent inter-GPU communication in MoE expert parallelism (EP) becomes a performance challenge. We observe substantial redundant inter-GPU data transfers in MoE that can be potentially addressed by in-switch computing. Unfortunately, the existing solution, NVLink SHARP (NVLS), can only support static collectives with regular patterns, incapable of dynamic communication with irregular patterns in MoE. To bridge the functionality gap, we propose DySHARP, an integral dynamic in-switch computing solution to accelerate MoE, encompassing both communication primitives and communication-aware scheduling: 1) Dynamic multimem addressing co-designs ISA, architecture, and runtime, as a dynamic extension to NVLS, reducing redundant traffic. However, the resulting traffic reduction is inherently asymmetric between two directions, preventing it from directly translating into speedup. 2) Token-centric kernel fusion deeply fuses the dispatch-computation-combine pipeline, resolving this asymmetry to translate traffic reduction into actual speedup. Compared with the state-of-the-art solution, DySHARP achieves up to 1.79$\times$ speedup.

Yunsong Zhou, Linyan Huang, Qingwen Bu et al.

Embodied scene understanding serves as the cornerstone for autonomous agents to perceive, interpret, and respond to open driving scenarios. Such understanding is typically founded upon Vision-Language Models (VLMs). Nevertheless, existing VLMs are restricted to the 2D domain, devoid of spatial awareness and long-horizon extrapolation proficiencies. We revisit the key aspects of autonomous driving and formulate appropriate rubrics. Hereby, we introduce the Embodied Language Model (ELM), a comprehensive framework tailored for agents' understanding of driving scenes with large spatial and temporal spans. ELM incorporates space-aware pre-training to endow the agent with robust spatial localization capabilities. Besides, the model employs time-aware token selection to accurately inquire about temporal cues. We instantiate ELM on the reformulated multi-faced benchmark, and it surpasses previous state-of-the-art approaches in all aspects. All code, data, and models will be publicly shared.

Zhengyi Li, Yakai Wang, Kang Yang et al.

For Transformer models, cryptographically secure inference ensures that the client learns only the final output, while the server learns nothing about the client's input. However, securely computing nonlinear layers remains a major efficiency bottleneck due to the substantial communication rounds and data transmission required. To address this issue, prior works reveal intermediate activations to the client, allowing nonlinear operations to be computed in plaintext. Although this approach significantly improves efficiency, exposing activations enables adversaries to extract model weights. To mitigate this risk, existing works employ a shuffling defense that reveals only randomly permuted activations to the client. In this work, we show that the shuffling defense is not as robust as previously claimed. We propose an attack that aligns differently shuffled activations to a common permutation and subsequently exploits them to extract model weights. Experiments on Pythia-70m and GPT-2 demonstrate that the proposed attack can align shuffled activations with mean squared errors ranging from $10^{-9}$ to $10^{-6}$. With a query cost of approximately \$1, the adversary can recover model weights with L1-norm differences ranging from $10^{-4}$ to $10^{-2}$ compared to the oracle weights.

Kangkang Sun, Jianhua Li, Xiuzhen Chen et al.

Coordinating mixed fleets of $10^4$ to $10^5$ vehicles, passenger cars, freight trucks, and autonomous vehicles, under stringent delay constraints is a central scalability bottleneck in next-generation V2X networks. Heterogeneous mean field games (HMFG) offer a principled coordination framework, yet a fundamental design question lacks theoretical guidance: how many agent types $K$ should be used for a fleet of size $N$? The core challenge is a two-sided trade-off that existing theory does not resolve: increasing $K$ reduces type-discretization error but simultaneously starves each class of the samples needed for reliable mean-field approximation. We resolve this trade-off by deriving an explicit $\varepsilon$-Nash error decomposition driven by a Wasserstein-based heterogeneity measure, and prove that the unique error-minimizing type count satisfies $K^*(N)=Θ(N^{1/3})$ in the canonical one-dimensional queue setting. We further establish a heterogeneity-aware convergence condition for G-prox PDHG and extend the framework to temporal-graph LEO satellite backhaul dynamics with provable robustness guarantees. A perhaps surprising consequence is that even for $N=10^5$ vehicles, only about 28 type classes suffice, cube-root compression rather than per-vehicle modeling, so type-granularity selection is largely a set-once design decision. Experiments validate the scaling law, achieve $2.3\times$ faster PDHG convergence at $K=5$, and deliver up to $29.5\%$ lower delay and $60\%$ higher throughput compared with homogeneous baselines.

Peng Tang, Jiacheng Liu, Xiaofeng Hou et al.

The Mixture-of-Experts (MoE) architecture has demonstrated significant advantages in the era of Large Language Models (LLMs), offering enhanced capabilities with reduced inference costs. However, deploying MoE-based LLMs on memoryconstrained edge devices remains challenging due to their substantial memory requirements. While existing expertoffloading methods alleviate the memory requirements, they often incur significant expert-loading costs or compromise model accuracy. We present HOBBIT, a mixed precision expert offloading system to enable flexible and efficient MoE inference. Our key insight is that dynamically replacing less critical cache-miss experts with low precision versions can substantially reduce expert-loading latency while preserving model accuracy. HOBBIT introduces three innovative techniques that map the natural hierarchy of MoE computation: (1) a token-level dynamic expert loading mechanism, (2) a layer-level adaptive expert prefetching technique, and (3) a sequence-level multidimensional expert caching policy. These innovations fully leverage the benefits of mixedprecision expert inference. By implementing HOBBIT on top of the renowned LLM inference framework Llama.cpp, we evaluate its performance across different edge devices with representative MoE models. The results demonstrate that HOBBIT achieves up to a 9.93x speedup in decoding compared to state-of-the-art MoE offloading systems.

Quan Liu, Hongzi Zhu, Zhenxi Wang et al.

Registration of point clouds collected from a pair of distant vehicles provides a comprehensive and accurate 3D view of the driving scenario, which is vital for driving safety related applications, yet existing literature suffers from the expensive pose label acquisition and the deficiency to generalize to new data distributions. In this paper, we propose EYOC, an unsupervised distant point cloud registration method that adapts to new point cloud distributions on the fly, requiring no global pose labels. The core idea of EYOC is to train a feature extractor in a progressive fashion, where in each round, the feature extractor, trained with near point cloud pairs, can label slightly farther point cloud pairs, enabling self-supervision on such far point cloud pairs. This process continues until the derived extractor can be used to register distant point clouds. Particularly, to enable high-fidelity correspondence label generation, we devise an effective spatial filtering scheme to select the most representative correspondences to register a point cloud pair, and then utilize the aligned point clouds to discover more correct correspondences. Experiments show that EYOC can achieve comparable performance with state-of-the-art supervised methods at a lower training cost. Moreover, it outwits supervised methods regarding generalization performance on new data distributions.

Zhengyi Li, Kang Yang, Jin Tan et al.

Transformer models have gained significant attention due to their power in machine learning tasks. Their extensive deployment has raised concerns about the potential leakage of sensitive information during inference. However, when being applied to Transformers, existing approaches based on secure two-party computation (2PC) bring about efficiency limitations in two folds: (1) resource-intensive matrix multiplications in linear layers, and (2) complex non-linear activation functions like $\mathsf{GELU}$ and $\mathsf{Softmax}$. This work presents a new two-party inference framework $\mathsf{Nimbus}$ for Transformer models. For the linear layer, we propose a new 2PC paradigm along with an encoding approach to securely compute matrix multiplications based on an outer-product insight, which achieves $2.9\times \sim 12.5\times$ performance improvements compared to the state-of-the-art (SOTA) protocol. For the non-linear layer, through a new observation of utilizing the input distribution, we propose an approach of low-degree polynomial approximation for $\mathsf{GELU}$ and $\mathsf{Softmax}$, which improves the performance of the SOTA polynomial approximation by $2.9\times \sim 4.0\times$, where the average accuracy loss of our approach is 0.08\% compared to the non-2PC inference without privacy. Compared with the SOTA two-party inference, $\mathsf{Nimbus}$ improves the end-to-end performance of \bert{} inference by $2.7\times \sim 4.7\times$ across different network settings.

Mingzhe Gao, Jieru Zhao, Zhe Lin et al.

High-level synthesis (HLS) notably speeds up the hardware design process by avoiding RTL programming. However, the turnaround time of HLS increases significantly when post-route quality of results (QoR) are considered during optimization. To tackle this issue, we propose a hierarchical post-route QoR prediction approach for FPGA HLS, which features: (1) a modeling flow that directly estimates latency and post-route resource usage from C/C++ programs; (2) a graph construction method that effectively represents the control and data flow graph of source code and effects of HLS pragmas; and (3) a hierarchical GNN training and prediction method capable of capturing the impact of loop hierarchies. Experimental results show that our method presents a prediction error of less than 10% for different types of QoR metrics, which gains tremendous improvement compared with the state-of-the-art GNN methods. By adopting our proposed methodology, the runtime for design space exploration in HLS is shortened to tens of minutes and the achieved ADRS is reduced to 6.91% on average.

Zhenyu Ning, Guangda Liu, Qihao Jin et al.

Recent developments in Video Large Language Models (Video LLMs) have enabled models to process long video sequences and demonstrate remarkable performance. Nonetheless, studies predominantly focus on offline video question answering, neglecting memory usage and response speed that are essential in various real-world applications, such as Deepseek services, autonomous driving, and robotics. To mitigate these challenges, we propose $\textbf{LiveVLM}$, a training-free framework specifically designed for streaming, online video understanding and real-time interaction. Unlike existing works that process videos only after one question is posed, LiveVLM constructs an innovative streaming-oriented KV cache to process video streams in real-time, retain long-term video details and eliminate redundant KVs, ensuring prompt responses to user queries. For continuous video streams, LiveVLM generates and compresses video key-value tensors (video KVs) to reserve visual information while improving memory efficiency. Furthermore, when a new question is proposed, LiveVLM incorporates an online question-answering process that efficiently fetches both short-term and long-term visual information, while minimizing interference from redundant context. Extensive experiments demonstrate that LiveVLM enables the foundation LLaVA-OneVision model to process 44$\times$ number of frames on the same device, and achieves up to 5$\times$ speedup in response speed compared with SoTA online methods at an input of 256 frames, while maintaining the same or better model performance.

Huanqi Hu, Bowen Xiao, Shixuan Sun et al.

Quantization is a critical technique for accelerating LLM inference by reducing memory footprint and improving computational efficiency. Among various schemes, 4-bit weight and 8-bit activation quantization (W4A8) offers a strong balance between accuracy and performance. However, existing W4A8 GEMM kernels fall short in practice due to inefficient dequantization on CUDA Cores, which cannot keep pace with the high throughput of Tensor Cores. In this paper, we present LiquidGEMM, a hardware-efficient W4A8 GEMM kernel for efficient LLM serving. LiquidGEMM designs two key techniques: LiquidQuant, a hardware-efficient quantization method that enables fast, overflow-safe dequantization using just two arithmetic instructions per four elements; and an implicit fine-grained pipeline that fully overlaps weight loading, dequantization, and MMA across warp groups without software synchronization or redundant memory traffic. Experimental results show that LiquidGEMM achieves up to 2.90x speedup over state-of-the-art W4A8 kernels and up to 4.94x end-to-end system-level speedup. Compared to various quantized GEMM kernels in NVIDIA TensorRT-LLM, LiquidGEMM delivers 1.12-1.63x performance gains, and achieves up to 1.63x system-level speedup.

Chunyu Xue, Weihao Cui, Han Zhao et al.

Joint consideration of scheduling and adaptive parallelism offers great opportunities for improving the training efficiency of large models on heterogeneous GPU clusters. However, integrating adaptive parallelism into a cluster scheduler expands the cluster scheduling space. The new space is the product of the original scheduling space and the parallelism exploration space of adaptive parallelism (also a product of pipeline, data, and tensor parallelism). The exponentially enlarged scheduling space and ever-changing optimal parallelism plan from adaptive parallelism together result in the contradiction between low-overhead and accurate performance data acquisition for efficient cluster scheduling. This paper presents Crius, a training system for efficiently scheduling multiple large models with adaptive parallelism in a heterogeneous cluster. Crius proposes a novel scheduling granularity called Cell. It represents a job with deterministic resources and pipeline stages. The exploration space of Cell is shrunk to the product of only data and tensor parallelism, thus exposing the potential for accurate and low-overhead performance estimation. Crius then accurately estimates Cells and efficiently schedules training jobs. When a Cell is selected as a scheduling choice, its represented job runs with the optimal parallelism plan explored. Experimental results show that Crius reduces job completion time by up to 48.9% and schedules large models with up to 1.49x cluster throughput improvement.