Hanxi Fang

Haocheng Xia, Mihir Pamnani, Hanxi Fang et al.

Key-value (KV) caching accelerates inference of large language models (LLMs) by reusing past computations for generated tokens. Its importance becomes even greater in long-context applications such as retrieval-augmented generation (RAG) and in-context learning (ICL). However, conventional KV caching embeds positional information directly into the cache, limiting its reusability. Existing solutions either restrict reuse to prefixes or require expensive memory materialization for positional re-encoding. We introduce LazyAttention, a novel attention mechanism that kernelizes deferred positional encoding to enable zero-copy, position-agnostic KV reuse. By adjusting positional encoding within attention kernels on-the-fly, LazyAttention resolves the materialization bottleneck, allowing a single physical KV copy to serve multiple logical requests at arbitrary positions. Leveraging attention kernels tailored for prefilling and decoding, our system achieves significant efficiency improvements: under skewed document distributions, it reduces time-to-first-token (TTFT) by 1.37$\times$ and increases inference throughput by 1.40$\times$ compared to the state-of-the-art Block-Attention, while maintaining comparable output quality.

Danlei Hu, Ziquan Fang, Hanxi Fang et al.

Transportation mode classification, the process of predicting the class labels of moving objects transportation modes, has been widely applied to a variety of real world applications, such as traffic management, urban computing, and behavior study. However, existing studies of transportation mode classification typically extract the explicit features of trajectory data but fail to capture the implicit features that affect the classification performance. In addition, most of the existing studies also prefer to apply RNN-based models to embed trajectories, which is only suitable for classifying small-scale data. To tackle the above challenges, we propose an effective and scalable framework for transportation mode classification over GPS trajectories, abbreviated Estimator. Estimator is established on a developed CNN-TCN architecture, which is capable of leveraging the spatial and temporal hidden features of trajectories to achieve high effectiveness and efficiency. Estimator partitions the entire traffic space into disjointed spatial regions according to traffic conditions, which enhances the scalability significantly and thus enables parallel transportation classification. Extensive experiments using eight public real-life datasets offer evidence that Estimator i) achieves superior model effectiveness (i.e., 99% Accuracy and 0.98 F1-score), which outperforms state-of-the-arts substantially; ii) exhibits prominent model efficiency, and obtains 7-40x speedups up over state-of-the-arts learning-based methods; and iii) shows high model scalability and robustness that enables large-scale classification analytics.