Speculative decoding is a powerful technique that accelerates token generation by leveraging both small and large models in tandem. Existing work focuses on either draft model speed or target model verification efficiency. Nevertheless, the inherent constraint of the draft model’s limited parameter size necessitates a compromise between drafting precision and decoding speed. This makes it difficult to achieve high accuracy across diverse downstream tasks. To address this challenge, we propose TaskSpec, a speculative decoding framework optimized for downstream tasks that improves drafting accuracy while maintaining the inference latency of the draft model. TaskSpec classifies incoming requests, automatically identifies downstream tasks, and constructs a heterogeneous set of draft models that are dynamically updated based on the input stream. Each draft model is aligned with the target model using dynamically collected task-specific prompt-response pairs, maximizing inference consistency for its designated task. Additionally, TaskSpec integrates a lightweight online classifier to dynamically route requests to appropriate draft models. Experimental results demonstrate that TaskSpec achieves a drafting accuracy improvement of 45% on LLaMA-2-13B and 13% on Qwen3-14B compared with vanilla speculative decoding, along with a 2.55 × and 1.41 × speedup for LLM inference, respectively.
This paper discusses the design, implementation, and evaluation of edge-federation-as-a-service where master-slave and peer-to-peer operation models rely on policies to manage data flows transfer across edge federations. These policies include rules that define permissions, prohibitions, and duties regulating interactions between edge nodes in terms of how data flows are transferred and which edge node is responsible for what during transfer. To illustrate the operation models, 2 scenarios are designed, homogeneous and heterogeneous. While the former focuses on interactions across all edge federations of the same type, i.e., either master-slave or peer-to-peer, the later mixes both. The paper also examines data flows from an offloading perspective. 3 offloading stages are defined, collection, offloading, and distribution. To demonstrate the technical feasibility of the edge-federation-as-a-service, an in-house testbed is implemented and evaluated in terms of latency, bandwidth, data volume, and scalability. Finally, future work is presented revolving around service deployment over edge federations and testbed’s performance improvement.
For scalable PDSs (parallel-and-distributed systems), dynamic MINs (multistage interconnection networks) with advantages of recursive construction and correct (S, D)-routing are studied over static interconnection networks, according to dynamic configurations, flexible-and-efficient communication, identical latency, and scalability. While recursive construction of the dynamic MINs is crucial, the (S, D)-routing of the existing recursive MINs is not stable. In the past, the original MINs (omega and flip) were proposed with the correct (S, D)-routing but their structures were not recursive. Later, the baseline network was presented with the recursive construction but its (S, D)-routing was not stable, while the correct (S, D)-routing was necessary for parallel communication. Recently, the baseline+ (one of semi-recursive MINs+) was fulfilled by post-port adjusting for the correct (S, D)-routing. However, for the expanded MIN+(N), old-port adjusting of sub-nets+(N/2) must be reset before integrated plus new-port adjusting, which is not consistent during construction. In this study, we propose a unified class of novel recursive MINs (reMINs) for d × d switches along with the fully recursive construction and the correct (S, D)-routing plus the complete partitioning (to support parallel tasks on independent partitions). Key innovation of our reMINs is novel dual-ISC (inter-stage connection) balancing to solve limitations of the existing MINs and MINs+. In network analysis, the correct (S, D)-routing of each reMIN was proven. In experiments, each reMIN-structure was verified and parallel (S, D)-routing of all-to-all personalized exchange communication was applied on scalable networks (N=dm≤65536) and flexible switches (d=2k).
The dynamic nature of Wireless Sensor Networks (WSNs) has led to a raise in data transfer. WSNs have concerns with data transmission, including lower energy efficiency, security and network longevity. In this paper, a Blockchain Framework based Secure Multipath Routing utilizing Conditional Self-Attention Generative Adversarial Network and Efficient Clustering in Edge-Assisted Wireless Networks (BFSMR-CSAGAN-EC-EAWN) is proposed. At first, network development authentication of sensor model clustered using Multi-Objective Generalized Normal Distribution Optimization (MOGNDO) that choose optimum cluster head (CH). The model increases the network life span by performing duty cycling utilizing the Improved Twin Delayed Deep Deterministic Policy Gradient (ITD3) method for lower power use. A last, a safe multiple path routing is carried out through Conditional Self-Attention Generative Adversarial Network (CSAGAN). In this, every transaction is stored on the blockchain (BC) technology using Fair Proof-of-Reputation (PoM) for increased security. The BFSMR-CSAGAN-EC-EAWN is implemented and the performance metrics, like Throughput, Packet delivery ratio (PDR), Energy consumption, network life time, No. of Alive nodes and latency are analyzed. Finally, the performance of BFSMR-CSAGAN-EC-EAWN method provides 16.60%, 24.01% and 26.60% high throughput, 16.57%, 14.16% and 12.01% high packet delivery ratio and 16.78%, 22.45% and 26.50% less energy consumption when compared to the existing methods respectively.
Diabetic retinopathy (DR) is a leading cause of preventable vision loss, and deep learning models can automate its screening from retinal fundus images. In networks such as the Brazilian Unified Health System, specialist shortages delay diagnosis, and both diagnostic accuracy and computational efficiency matter. Two questions shape practical deployment: which convolutional neural network (CNN) architecture to pick, and how to run it efficiently on modern hardware. This paper extends our previous work with two additions: a distributed multi-node training experiment complemented by a TensorFlow XPlane execution profile, and a SimGrid discrete-event simulation calibrated on the physical runs. Predictive performance is now reported through AUC together with the specificity obtained at a fixed 95% sensitivity operating point, which brings the evaluation closer to how a screening pipeline is actually deployed. We benchmark 38 pretrained CNN architectures on a Brazilian DR dataset. MobileNet offers the best overall trade-off: relative to InceptionV3, it improves AUC by 3%, achieves a specificity of 0.67 versus 0.58, and reduces energy, model size, and training time by 77%, 85%, and 83%, respectively. A multi-objective hyperparameter study over activation function, optimizer, and input resolution further reduces energy and time without hurting predictive quality. It confirms an input size of 300 × 300 as the efficiency sweet spot. We then study the tuned MobileNet on the Grace Superchip. A thread sweep on one node identifies an 18–36-thread band where speedup and energy jointly optimize, achieving about 20 × speedup and 88% energy savings before scheduling and memory pressure dominate. Distributed training across two Grace nodes with TensorFlow MultiWorkerMirroredStrategy on a 6.13 × larger dataset reaches a 1.93 × speedup at 96.5% scaling efficiency, halves training time, adds less than 1% to total energy, and preserves both AUC and specificity. A TensorFlow XPlane profile of the distributed runs shows that 97% of the active time is spent in computation and only 0.74% in inter-node communication, which explains the near-linear scaling. A SimGrid simulation calibrated on the two physical configurations projects efficiency to about 74% at four nodes and 54% at eight nodes and beyond, and identifies the gRPC/10 Gbps Ethernet transport as the main bottleneck at larger scale. These results indicate that efficient DL for DR depends jointly on model choice, the clinical operating point at which the model is evaluated, and the system-level execution strategy chosen for the target hardware.
