arXiv subjects

Peiqi Yin

Publications and source records attributed to Peiqi Yin.

11 recordsLinked to original sources

arXiv2026

CAM: Question Answering on Entity-Centric Videos with Continuous Extraction and Adaptive Querying

Memory facilitates question answering over long videos by extracting and retrieving facts to fit within the limited context windows of multimodal LLMs (MLLMs). Existing solutions typically extract independent memory entries from fixed-length video clips and thus cannot capture high-level semantics that need to be summarized over extended time periods, such as character traits and relations. Moreover, they rely solely on similarity-based retrieval and may fail to retrieve the fine-grained details required for question answering. To tackle these problems, we propose CAM, featuring continuous extraction for high-level semantics and adaptive querying for fine-grained details. In particular, CAM stores the entities and relations extracted from video clips in a knowledge graph. To capture the high-level semantics of each entity or relation, CAM summarizes the local subgraph of the target entity or relation once the subgraph reaches a predefined size. To retrieve the fine-grained details required for question answering, CAM supports multiple search methods, including knowledge graph traversal, video re-watching, and audio listening. It utilizes a planner-executor-verifier pipeline to adaptively compose these search methods according to question intent. Evaluations on three benchmarks show that CAM outperforms SOTA baselines and improves their accuracy by up to 23 percentage points. Code is available at https://github.com/Jake-Tian/CAM.

cs.CV↗

arXiv2026

LiteTopK: Exploiting the Curse of Dimensionality for a Fused Indexer-TopK Kernel in Long-Context Sparse Attention

Indexer-TopK, the operation to compute the scores and select the top-k candidates, is widely used by sparse attention algorithms in large language models and vector retrieval in recommendation systems and vector databases. However, existing GPU-based Indexer-TopK kernels like DeepSeek Sparse Attention (DSA) remain inefficient due to excessive global memory traffic, costly synchronization, and prohibitive memory overhead. In this study, inspired by the curse of dimensionality phenomenon, we first observe that sparse attention scores exhibit a score concentration phenomenon, where scores tend to fall within a narrow range. Based on this observation, we propose LITETOPK, an efficient fused Indexer-TopK kernel. LITETOPK first samples a small subset of data to estimate query-data score ranges, then partitions candidates into bins accordingly. This organization allows the LITETOPK kernel to maintain a tight approximate threshold online, write back only promising candidates, reduce unnecessary I/O and memory overhead while preserving exact Top-k correctness. Building on LITETOPK, we further propose LITEDSA, which exploits the similarity of top-k candidate sets among neighboring tokens. LITEDSA packs neighboring tokens' candidates for joint computation and masks out extra scores for each query, thereby reducing memory traffic while preserving correctness. Experimental results in a real-world deployment environ ment with eight B200 GPUs show that LITETOPK+LITEDSA accelerates the prefill stage of GLM 5.2 by 1.35x, with no performance loss and lower memory overhead.

cs.LG↗

arXiv2026

LiveServe: Interaction-Aware Serving for Real-Time Omni-Modal LLMs

Realtime omni-modal LMs support speech-centric conversations where users stream inputs, hear generated audio, and interrupt freely. Existing Omni-LM serving systems still rely on throughput-oriented LLM scheduling and LRU KV offloading. These policies ignore audio playback and multi-turn reuse: they may generate tokens far beyond what users hear, wasting work after barge-in, and evict KV state needed in the next turn. LiveServe is an interaction-aware serving system for realtime Omni-LM interaction. It exposes playback progress, speech activity, and barge-in events to the serving pipeline. The scheduler prioritizes first-audio and near-underrun sessions while limiting generation beyond the playback frontier. The KV manager uses next-use-aware eviction and preloads likely-needed KV during user speech to hide reload latency. On vLLM-Omni, LiveServe improves realtime serving across two Omni-LMs and mixed workloads. It lowers P90 audio TTFP by $1.55\times$ on average and up to $2.21\times$, while improving completed-request throughput by $1.15\times$ on average and up to $1.56\times$, and moves most KV reload work off the next-turn critical path.

cs.DC↗

arXiv2026

ScaleGANN: Accelerate Large-Scale ANN Indexing by Cost-effective Cloud GPUs

Graph-based ANNS algorithms have gained increasing research interest and market adoption due to their efficiency and accuracy in retrieval. Existing approaches primarily rely on CPUs for graph index construction and retrieval, but this often requires significant time, especially for large-scale and high-dimensional datasets. Some studies have explored GPU-based solutions. However, GPUs are costly and their limited memory makes handling large datasets challenging. In this paper, we propose a novel end-to-end system ScaleGANN that enables users to efficiently construct graph indexes for large-scale, high-dimensional datasets by leveraging low-cost spot GPU resources in a distributed cloud system. ScaleGANN utilized the idea of divide-and-merge, with an optimized vector partitioning algorithm to further improve the indexing time and space efficiency while guaranteeing good index quality. Its novel resource allocation strategy realized multi-GPU indexing parallelism and overall cost-effectiveness for both build and query. Besides, we designed a task scheduler and cost model for better spot instance management and evaluation. We tested our system on large real-world datasets. Experiment results show that our approach can significantly accelerate the index build time to up to 9x times at even 6x lower price compared with the state-of-the-art extendable ANNS benchmark DiskANN.

cs.DB↗

arXiv2026

vLLM-Omni: Fully Disaggregated Serving for Any-to-Any Multimodal Models

Any-to-any multimodal models that jointly handle text, images, video, and audio represent a significant advance in multimodal AI. However, their complex architectures (typically combining multiple autoregressive LLMs, diffusion transformers, and other specialized components) pose substantial challenges for efficient model serving. Existing serving systems are mainly tailored to a single paradigm, such as autoregressive LLMs for text generation or diffusion transformers for visual generation. They lack support for any-to-any pipelines that involve multiple interconnected model components. As a result, developers must manually handle cross-stage interactions, leading to huge performance degradation. We present vLLM-Omni, a fully disaggregated serving system for any-to-any models. vLLM-Omni features a novel stage abstraction that enables users to decompose complex any-to-any architectures into interconnected stages represented as a graph, and a disaggregated stage execution backend that optimizes resource utilization and throughput across stages. Each stage is independently served by an LLM or diffusion engine with per-stage request batching, flexible GPU allocation, and unified inter-stage connectors for data routing. Experimental results demonstrate that vLLM-Omni reduces job completion time (JCT) by up to 91.4% compared to baseline methods. The code is public available at https://github.com/vllm-project/vllm-omni.

cs.DC↗

arXiv2025

SparseServe: Unlocking Parallelism for Dynamic Sparse Attention in Long-Context LLM Serving

Serving long-context LLMs is costly because attention computation grows linearly with context length. Dynamic sparse attention algorithms (DSAs) mitigate this by attending only to the key-value (KV) cache of critical tokens. However, with DSAs, the main performance bottleneck shifts from HBM bandwidth to HBM capacity: KV caches for unselected tokens must remain in HBM for low-latency decoding, constraining parallel batch size and stalling further throughput gains. Offloading these underutilized KV caches to DRAM could free HBM capacity, allowing larger parallel batch sizes. Yet, achieving such hierarchical HBM-DRAM storage raises new challenges, including fragmented KV cache access, HBM cache contention, and high HBM demands of hybrid batching, that remain unresolved in prior work. This paper proposes SparseServe, an LLM serving system that unlocks the parallel potential of DSAs through efficient hierarchical HBM-DRAM management. SparseServe introduces three key innovations to address the challenges mentioned above: (1) fragmentation-aware KV cache transfer, which accelerates HBM-DRAM data movement through GPU-direct loading (FlashH2D) and CPU-assisted saving (FlashD2H); (2) working-set-aware batch size control that adjusts batch sizes based on real-time working set estimation to minimize HBM cache thrashing; (3) layer-segmented prefill that bounds HBM use during prefill to a single layer, enabling efficient execution even for long prompts. Extensive experimental results demonstrate that SparseServe achieves up to 9.26x lower mean time-to-first-token (TTFT) latency and up to 3.14x higher token generation throughput compared to state-of-the-art LLM serving systems.

cs.DC↗

arXiv2025

Gorgeous: Revisiting the Data Layout for Disk-Resident High-Dimensional Vector Search

Similarity-based vector search underpins many important applications, but a key challenge is processing massive vector datasets (e.g., in TBs). To reduce costs, some systems utilize SSDs as the primary data storage. They employ a proximity graph, which connects similar vectors to form a graph and is the state-of-the-art index for vector search. However, these systems are hindered by sub-optimal data layouts that fail to effectively utilize valuable memory space to reduce disk access and suffer from poor locality for accessing disk-resident data. Through extensive profiling and analysis, we found that the structure of the proximity graph index is accessed more frequently than the vectors themselves, yet existing systems do not distinguish between the two. To address this problem, we design the Gorgeous system with the principle of prioritizing graph structure over vectors. Specifically, Gorgeous features a memory cache that keeps the adjacency lists of graph nodes to improve cache hits and a disk block format that explicitly stores neighbors' adjacency lists along with a vector to enhance data locality. Experimental results show that Gorgeous consistently outperforms two state-of-the-art disk-based systems for vector search, boosting average query throughput by over 60% and reducing query latency by over 35%.

cs.DB↗

arXiv2025

PilotANN: Memory-Bounded GPU Acceleration for Vector Search

Approximate Nearest Neighbor Search (ANNS) has become fundamental to modern deep learning applications, having gained particular prominence through its integration into recent generative models that work with increasingly complex datasets and higher vector dimensions. Existing CPU-only solutions, even the most efficient graph-based ones, struggle to meet these growing computational demands, while GPU-only solutions face memory constraints. As a solution, we propose PilotANN, a hybrid CPU-GPU system for graph-based ANNS that utilizes both CPU's abundant RAM and GPU's parallel processing capabilities. Our approach decomposes the graph traversal process of top-$k$ search into three stages: GPU-accelerated subgraph traversal using SVD-reduced vectors, CPU refinement and precise search using complete vectors. Furthermore, we introduce fast entry selection to improve search starting points while maximizing GPU utilization. Experimental results demonstrate that PilotANN achieves $3.9 - 5.4 \times$ speedup in throughput on 100-million scale datasets, and is able to handle datasets up to $12 \times$ larger than the GPU memory. We offer a complete open-source implementation at https://github.com/ytgui/PilotANN.

cs.DC↗

arXiv2025

Progressive Sparse Attention: Algorithm and System Co-design for Efficient Attention in LLM Serving

Processing long contexts has become a critical capability for modern large language models (LLMs). However, serving long-context LLMs comes with significant inference costs due to the high memory overhead of the key-value (KV) cache. Existing work leverages dynamic sparse attention algorithms (DSAes) to mitigate the KV cache overhead, but these algorithms rely on top-$k$ KV cache selection, which results in a trade-off between accuracy and efficiency. A larger $k$ improves accuracy but decreases efficiency, while a smaller $k$ boosts efficiency but compromises accuracy. To overcome this trade-off, this paper presents PSA, a $\underline{P}$rogressive $\underline{S}$parse $\underline{A}$ttention mechanism that integrates algorithmic innovations with system co-design to achieve both high inference accuracy and improved efficiency in LLM serving. The PSA algorithm adaptively adjusts the KV cache budget of different tokens and layers according to their real attention weight distributions, rather than relying on a fixed budget $k$. This enables high accuracy while minimizing KV cache usage. To further enhance execution efficiency, we introduce a pipelined iteration scheme that reduces CPU-GPU interleaving and synchronization overhead during PSA computation. Additionally, we implement unified GPU memory management that optimizes PSA's memory utilization by accounting for uneven memory requirements across different model layers. Extensive experimental results demonstrate that PSA reduces KV cache usage for attention computation by up to 2.4$\times$ and 8.8$\times$, and increases end-to-end serving throughput by up to 1.4$\times$ and 2.0$\times$, compared to state-of-the-art DSAes and systems without sparse attention, respectively.

cs.LG↗

arXiv2023

SPT: Fine-Tuning Transformer-based Language Models Efficiently with Sparsification

Transformer-based large language models (e.g., BERT and GPT) achieve great success, and fine-tuning, which tunes a pre-trained model on a task-specific dataset, is the standard practice to utilize these models for downstream tasks. However, Transformer fine-tuning has long running time and high memory consumption due to the large size of the models. We propose the SPT system to fine-tune Transformer-based models efficiently by introducing sparsity. We observe that the memory consumption of Transformer mainly comes from storing attention weights for multi-head attention (MHA), and the majority of running time is spent on feed-forward network (FFN). Thus, we design the sparse MHA module, which computes and stores only large attention weights to reduce memory consumption, and the routed FFN module, which dynamically activates a subset of model parameters for each token to reduce computation cost. We implement SPT on PyTorch and customize CUDA kernels to run sparse MHA and routed FFN efficiently. Specifically, we use product quantization to identify the large attention weights and compute attention via sparse matrix multiplication for sparse MHA. For routed FFN, we batch the tokens according to their activated model parameters for efficient computation. We conduct extensive experiments to evaluate SPT on various model configurations. The results show that SPT consistently outperforms well-optimized baselines, reducing the peak memory consumption by up to 50% and accelerating fine-tuning by up to 2.2x.

cs.DC↗

arXiv2022

DGI: Easy and Efficient Inference for GNNs

While many systems have been developed to train Graph Neural Networks (GNNs), efficient model inference and evaluation remain to be addressed. For instance, using the widely adopted node-wise approach, model evaluation can account for up to 94% of the time in the end-to-end training process due to neighbor explosion, which means that a node accesses its multi-hop neighbors. On the other hand, layer-wise inference avoids the neighbor explosion problem by conducting inference layer by layer such that the nodes only need their one-hop neighbors in each layer. However, implementing layer-wise inference requires substantial engineering efforts because users need to manually decompose a GNN model into layers for computation and split workload into batches to fit into device memory. In this paper, we develop Deep Graph Inference (DGI) -- a system for easy and efficient GNN model inference, which automatically translates the training code of a GNN model for layer-wise execution. DGI is general for various GNN models and different kinds of inference requests, and supports out-of-core execution on large graphs that cannot fit in CPU memory. Experimental results show that DGI consistently outperforms layer-wise inference across different datasets and hardware settings, and the speedup can be over 1,000x.

cs.LG↗