Skip to content
Log in
#edge computing Open access

Ultralytics YOLO

Glenn Jocher Jing Qiu Ayush Chaurasia
Aug 2026 · Zenodo (CERN European Organization for Nuclear Research)

Abstract

🌟 Summary Ultralytics 8.4.133 improves hyperparameter tuning convergence, speeds up inference preprocessing, expands detection metrics, and simplifies edge-device setup. 🚀 📊 Key Changes Smarter hyperparameter tuning — PR #25984 by @glenn-jocher Replaces coordinate-by-coordinate crossover with fitness-weighted selection of complete, high-performing configurations. Preserves useful relationships between hyperparameters instead of mixing them independently. Mutates approximately half of the parameters in normalized search-space coordinates, allowing parameters that start at zero—such as degrees or shear—to evolve more effectively. Gradually reduces mutation size when tuning stops finding better results, encouraging refinement after broad exploration. Prevents duplicate candidates after clipping, rounding, or integer conversion, including small and discrete search spaces. Ray Tune now defaults to Optuna multivariate TPE, with parallel-aware suggestions rather than independent random search. Faster predictor preprocessing — PR #25982 by @jahsef ⚡ Moves image channel reordering and tensor-contiguity operations from CPU-side NumPy processing to the inference device. Preserves output values while reducing unnecessary CPU copies. Reported benchmarks show approximately 2.2–3.1× faster preprocessing on an RTX 5080, with additional gains on CPU. Automatic channels-last CPU inference — PR #25983 by @JESUSROYETH Enables channels-last memory layout automatically for native PyTorch inference and standalone validation on supported x86 Linux and Windows CPUs with oneDNN. Keeps training defaults and unsupported platforms unchanged. Explicit channels_last=True remains available for supported CPU and CUDA paths. Saved models are converted back to a safe contiguous format and stale EMA data is cleared to improve compatibility. More accurate INT8 calibration subsets — PR #25978 by @JESUSROYETH Fixes fraction handling during classification and detection INT8 export calibration. Scalar fractions now apply directly to the selected calibration split, while list-based fractions retain train/validation/test behavior. Prevents exports from unintentionally calibrating on an entire dataset when only a subset was requested. Size-specific mAP for custom detection datasets — PR #25981 by @fcakyon 📈 Custom detection datasets can now report small-, medium-, and large-object mAP when using save_json=True. Builds temporary COCO-format annotations internally while preserving existing native metrics and prediction files. Applies consistently during training validation, final-model validation, and standalone validation. Simpler edge-device installation Raspberry Pi, Jetson, DGX Spark, DeepStream, and related guides now install the base ultralytics package instead of the larger [export] extra. Export dependencies are installed automatically when an export is requested, reducing installation size and dependency conflicts. Improved Weights & Biases artifact control — PR #25985 by @glenn-jocher W&B model artifact uploads now follow the existing training save argument. save=False skips uploading the best checkpoint while retaining metrics and plots. Default behavior remains unchanged with save=True. Package update Version bumped to 8.4.133. 🎯 Purpose & Impact Better tuning results: Hyperparameter searches are more likely to preserve successful configurations, explore meaningful alternatives, and avoid wasting trials on duplicates. 🎯 Faster inference: Device-side preprocessing can reduce latency, particularly for batched inference and CPU-bound pipelines. Broader performance optimization: Supported x86 CPU users may benefit from channels-last inference without changing their existing commands. More reliable model export: INT8 calibration now honors requested dataset fractions, improving calibration speed and reducing unexpected resource usage. Richer evaluation: Custom detection datasets can now receive object-size performance breakdowns similar to COCO evaluations. Easier edge deployment: Base installations are smaller and less prone to dependency conflicts, while export workflows remain available when needed. More control over experiment storage: W&B users can keep experiment tracking lightweight by disabling checkpoint saving with the standard save setting. What's Changed Allow skipping W&B model artifacts by @fcakyon in https://github.com/ultralytics/ultralytics/pull/25979 Remove [export] from edge-device install guides by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/25977 Accelerate predictor preprocessing on inference devices by @jahsef in https://github.com/ultralytics/ultralytics/pull/25982 Fix INT8 export calibration fractions by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25978 Use save argument for W&B model artifacts by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25985 Compute size-specific mAP for custom detection datasets by @fcakyon in https://github.com/ultralytics/ultralytics/pull/25981 Enable channels-last by default for x86 CPU inference by @JESUSROYETH in https://github.com/ultralytics/ultralytics/pull/25983 Improve hyperparameter Tuner mutation convergence by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25984 Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.132...v8.4.133

DOI
View source

Similar papers

#federated learning Sep 2026

AoI Minimization in Heterogeneous MEC Networks: A Federated Learning-Assisted Hybrid DRL and Convex Approach

This paper investigates a dynamic heterogeneous mobile edge computing network (HMECN), where mobile devices (MDs) could offload their full tasks to a small base station (SBS) directly or the macro base station (MBS) in direct or relay mode. As age of information (AoI) is a comprehensive and accurate metric to capture the freshness of computation results, we formulate a long-term weighted sum AoI (LWSA) minimization problem in the HMECN by jointly optimizing the offloading decisions of MDs as well as the bandwidth and computation resource allocation of all base stations, subject to energy, delay and peak AoI constraints. To address the formulated non-convex mixed integer nonlinear programming problem, we decompose it into the offloading decision optimization (ODO) top-problem and the resource allocation optimization (RAO) sub-problem. Based on the decomposition, we propose a federated learning (FL)-assisted hybrid DRL and convex approach that is comprised of a safe multi-agent DRL algorithm, convex optimization and FL. The ODO top-problem is solved by the safe multi-agent DRL algorithm, which strictly ensures that the actions of each agent do not exceed its energy constraint and then paves the way for using convex optimization to solve the RAO sub-problem. FL is used to alleviate the training instability problem aggravated by multi-agent settings via breaking the limitation of partial knowledge for each individual agent. Simulation results demonstrate the superiority of the proposed approach in terms of the LWSA, convergence, scalability and robustness in dynamic environments.

Xiaoying Liu, Junhao Zheng, Kechen Zheng et al. · 8 citations
#edge computing Sep 2026

DRUDM-CFG: A Fairness-Aware Multi-Agent DRL Algorithm for AMEC-Assisted Task Offloading in Post-Disaster Scenarios

High-altitude airships (HAS) and uncrewed aerial vehicles (UAVs) equipped with Multiaccess Edge Computing (MEC) servers have emerged as promising aerial MEC nodes for providing task offloading (TO) services to intelligent mobile devices (IMDs) in post-disaster scenarios. HAS offers robust computing and energy resources, while UAVs provide flexible, low-altitude coverage for rapid deployment. However, direct task offloading from IMDs to HAS often leads to task failures due to high transmission delays. UAVs with limited onboard resources require to minimize resource waste. Additionally, IMDs in sparse areas face insufficient TO services due to unfair UAV coverage. This paper defines these challenges as a joint optimization problem involving TO, RA, and UAV coverage fairness. It proposes a cooperative aerial Multiaccess Edge Computing (AMEC) framework integrating HAS and UAVs to address the issue. Within this framework, a hybrid TO scheme is first developed to mitigate the high transmission delay between IMDs and HAS. Second, a Distance, Resource, Urgency-based Decision Mechanism (DRUDM) is designed to enhance the accuracy of UAVs in selecting target IMDs for TO services. Third, a Coverage Fairness Guarantee (CFG) strategy is proposed to optimize UAV flight trajectories, ensuring IMDs in sparse areas receive fair TO services. Finally, the joint optimization problem is modeled as a Multi-Agent Partially Observable Markov Decision Process (MA-POMDP), and a DRUDM–CFG algorithm is presented to efficiently solve this complex non-convex optimization problem. Experimental results demonstrate that the proposed algorithm outperforms other compared algorithms in task completion rate and average delay, benefiting from the DRUDM mechanism. Meanwhile, the CFG strategy effectively improves TO service fairness for IMDs in sparse areas.

Xiting Peng, Chuanqi Qin, Xiaoyu Zhang et al. · 4 citations
#edge computing Preprint Aug 2026

Computing an e-net of a closed hyperbolic surface

Hyperbolic surfaces are a fundamental object in mathematics and play an increasingly important role in computational geometry and topology. A key ingredient in the design of efficient algorithms on such surfaces is the availability of a geometric discretization of controlled complexity. In this paper, we present the first algorithm for constructing e-nets on hyperbolic surfaces starting from a fundamental polygon representation. Our approach is based on Delaunay refinement and relies on maintaining Delaunay triangulations through edge flips. The size of an e-net cannot be bounded solely as a function of the genus because of the presence of arbitrarily long collars around short geodesics. To overcome this difficulty, we introduce the notion of a pseudo e-net, which decomposes the surface into e-thin cylinders together with a Delaunay triangulation over an e-net of the remaining thick part. As applications, we obtain algorithms for computing the length spectrum of an e-thick hyperbolic surface and for computing the systole from a pseudo log(sqrt(2))-net. These results demonstrate that Delaunay-based discretizations provide a practical and versatile framework for algorithmic computations on hyperbolic surfaces.

V. Delecroix, Vincent Despré, Camille Lanuel et al. · 3 citations
#edge computing Sep 2026

Toward 6G Edge Intelligence: Lightweight LLMs for Intent-Driven Network Automation

Future 6G networks are envisaged to tightly integrate communication, sensing, and computing, demanding real-time, intent-driven intelligence at the edge. While large language models (LLMs) excel in intent recognition and semantic reasoning, their application to real-time network lifecycle management at the edge is limited by heterogeneous application intents (APPIs), dynamic network conditions, and severe resource constraints. This paper proposes a novel lightweight LLM architecture, KGLlama-KD, that synergizes knowledge graphs (KGs) with knowledge distillation (KD) to enable intent-driven networking and enhance 6G edge intelligence. Specifically, a KG is constructed to formally describe the relationships among application scenarios, functional primitives, performance requirements within APPIs, and the correspondences between APPIs and network service requests (NSRs), thereby producing a structured intent training dataset. Building upon the Llama 3 foundation model, a two-phase optimization framework is designed to support lightweight edge deployment while preserving translation fidelity. The LLM is first fine-tuned with KG guidance and compressed via KD in the cloud, and then deployed on resource-constrained edge nodes to perform real-time, accurate, and efficient APPIs interpretation. Experiments validate that KGLlama-KD achieves 95% accuracy for APPI understanding, surpassing DeepSeek and Qwen by an average of 8%. The distilled model reduces inference latency by 60% compared to full-scale LLMs, fulfilling the sub-100 ms requirement for 6G latency-sensitive services.

Bing Wu, Sai Zou, Minghui Liwang et al. · 3 citations
#edge computing Sep 2026

Joint Latency and Charge Cost Minimization for Reliable Task Offloading in Dispersed Computing: A Multi-Objective Optimization Approach

Dispersed computing has emerged as a promising paradigm that leverages underutilized resources from massive Internet of Things devices (IoTDs) to enhance the computing capacity at the network edge. However, existing works about the dispersed computing overlook the heterogeneous computing environment with parallel and serial computations and task reliability requirements for the hardware-constrained IoTDs, and they lack multi-objective optimization approaches to optimize the task offloading. To address the challenges, we propose a comprehensive scheme to achieve a delay-aware and economic-aware dispersed computing paradigm by using a multi-objective optimization approach. Particularly, we consider parallel processing at an edge server and serial processing at the lightweight IoTDs, and leverage the task redundancy to satisfy the task reliability requirements on the IoTD side. We further formulate a constrained multi-objective optimization problem (CMOP) aiming at jointly optimizing the task assignment, bandwidth allocation, and CPU frequency allocation to simultaneously minimize the total delay cost and the total charge cost of the tasks. To address the CMOP, we propose an improved constrained multi-objective evolutionary algorithm that employs a dual-population cooperative mechanism between two populations and a repairing constraint-handling technique. The dual-population cooperative mechanism can balance convergence toward Pareto optimality and solution diversity maintenance. The repairing constraint-handling technique is designed to guide solutions toward feasible regions, achieving efficient exploration of complex constrained search spaces. Simulation results demonstrate the superiority of our algorithm in seeking the better-converged and better-distributed Pareto optimal solutions to well address the tradeoffs between the two objectives.

Xumin Huang, Zexiong Wu, Chaoda Peng et al. · 2 citations
#edge computing Sep 2026

Service Enhancement and Reliability Assurance in 6G Vehicular Networks via a Stackelberg Game-Theoretic Approach

With the rapid development of 6G and Internet of Vehicles (IoV) technologies, the volume of computation-intensive tasks generated by intelligent vehicles is growing exponentially. Given limited onboard processing capabilities, vehicles increasingly rely on edge servers deployed by service providers (SPs) at roadside units to offload tasks. Vehicle clients can offload the tasks to SPs to mitigate their onboard computation load, while SPs derive economic benefits through the provision of computation resources. However, this interaction introduces a conflict of interest, as vehicles aim to minimize their offloading costs, while SPs seek to maximize revenue. To address this problem, we propose SPOR, a Stackelberg game-based service priority-aware computation offloading and resource pricing scheme in IoV. SPOR is a hierarchical game-theoretic framework in which SPs act as leaders setting prices, while vehicles act as followers determining their offloading strategies. A novel service prioritization function is introduced, incorporating booking price, system load, and reputation to ensure fair and balanced resource allocation. We provide a theoretical proof of the existence and uniqueness of a Nash equilibrium. Extensive experiments on a real-world vehicle edge computing dataset show that SPOR outperforms baseline methods in delay, energy consumption, average load, and task completion rate. Notably, SPOR maintains task completion rates above 97% even under heavy workloads, demonstrating its effectiveness in enhancing system reliability and overall performance.

Kai Peng, Yuanlin Lin, Shuai Zhao et al. · 2 citations

Related blog posts