Partial Typing for Asynchronous Multiparty Sessions: Ensuring Partial Communication Properties
The paper presents a partial type system for multiparty sessions with asynchronous communication, ensuring partial versions of lock-freedom and orphan-message-freedom for specified subsets of participants.
Efficient Tensor Parallel Inference of Large Language Models on Low-Resource Edge Devices
TPI-LLM, a compute- and memory-efficient tensor parallel inference framework, enables serving 70B-scale large language models on low-resource edge devices by leveraging a sliding window memory scheduler and a star-based allreduce algorithm.
Comprehensive Benchmark for Evaluating Large Language Models on Conversational Question Answering with Instructional Documents
The InsCoQA benchmark assesses the ability of large language models to retrieve, interpret, and accurately summarize procedural guidance from multiple instructional documents in a conversational setting, reflecting the intricate and multi-faceted nature of real-world instructional tasks.
比特幣每年消耗的電力超過阿根廷全國的分析
比特幣的"挖礦"過程非常耗電,據劍橋大學的分析,每年消耗約121.36太瓦時的電力,這個數字很難下降,除非比特幣的價值大幅下跌。
캠브리지 대학 분석에 따르면 비트코인은 아르헨티나 전체 전력 소비량을 초과하는 전력을 사용한다
비트코인 채굴 과정은 막대한 전력을 소비하며, 이는 비트코인 가치 하락이 없다면 감소하기 어려운 것으로 나타났다.
Distributed Monitoring of Timed Properties with Imprecise Clocks
This paper proposes an online monitoring algorithm for timed properties in a distributed setting where each component has its own local clock, resulting in imprecise timestamps and partial information across components.
Influence Propagation in Signed Networks: A Graphical Characterization of Opinion Dynamics
The core message of this paper is to study and quantify the influence that individuals in a heterogeneous network exert on each other when their opinions evolve according to a Friedkin-Johnsen (FJ) based model with signed interactions.
Comprehensive Spatiotemporal Mobility Datasets from Four Low- and Middle-Income Countries
The NetMob23 dataset provides comprehensive spatiotemporal data on population density and origin-destination matrices across four low- and middle-income countries (India, Mexico, Indonesia, and Colombia) over 2019-2020, enabling researchers to study human mobility patterns and their applications.
대규모 언어 모델 서빙을 위한 계층 단위 KV 캐시 관리 최적화
대규모 언어 모델의 문맥 창 확장으로 인한 지연 시간 증가 문제를 해결하기 위해 LayerKV를 제안한다. LayerKV는 계층 단위 KV 캐시 할당, 관리 및 오프로딩을 통해 GPU KV 캐시 사용을 최적화하여 대기 시간을 줄이고, SLO 요구 사항을 충족시킨다.
Optimizing Large Language Model Serving by Reducing Queuing Delays through Layer-wise KV Cache Management
LayerKV, a simple yet effective plug-in method, significantly reduces Time to First Token (TTFT) latency in large language model serving by introducing layer-wise KV cache allocation, management, and offloading, without requiring additional hardware or compromising output performance.
