The growing industrial demand for customized and cost-efficient large language models (LLMs) is fueled by the rise of vertical, domain-specific tasks and the need to optimize performance under constraints such as latency and budget. Knowledge distillation, as an efficient model compression and transfer technique, offers a feasible solution. However, existing distillation frameworks often require manual intervention and struggle to meet such complex user-defined distillation requirements. To bridge this gap, we propose Stratos, an end-to-end LLM distillation pipeline that automates server/model selection, knowledge distillation, and deployment in distributed cloud environments. Given user-defined constraints on model performance and system budget, Stratos automatically selects Pareto-optimal servers, dynamically matches teacher–student pairs, and adapts distillation strategies based on task complexity to optimize cloud hosting. Experiments show that Stratos produces a student model that achieves four times the accuracy of its GPT-4o teacher baseline on a rare, domain-specific Mahjong reasoning task with reverse synthetic data and knowledge injection. Moreover, it achieves reduced latency and cost without compromising accuracy. These results highlight its promise for vertical-domain LLM deployment.

Pre-trained Transformer-based large models have revolutionized personal virtual assistants, but their deployment in cloud environments faces challenges related to data privacy and response latency. Deploying large models closer to the data and users has become a key research area to address these issues. However, applying these models directly often entails significant difficulties, such as model mismatching, resource constraints, and energy inefficiency. Automated design of customized models is necessary, but it faces three key challenges, namely, the high cost of centralized model customization, imbalanced performance from user heterogeneity, and suboptimal performance from data heterogeneity. In this paper, we propose ACME, an adaptive customization approach of Transformer-based large models via distributed systems. To avoid the low cost-efficiency of centralized methods, ACME employs a bidirectional single-loop distributed system to progressively achieve fine-grained collaborative model customization. In order to better match user heterogeneity, it begins by customizing the backbone generation and identifying the Pareto Front under model size constraints to ensure optimal resource utilization. Subsequently, it performs header generation and refines the model using data distribution-based personalized architecture aggregation to match data heterogeneity. Evaluation on different datasets shows that ACME achieves cost-efficient models under model size constraints. Compared to centralized systems, data transmission volume is reduced to 6%. Additionally, the average accuracy improves by 10% compared to the baseline, with the trade-off metrics increasing by nearly 30%.

In edge computing, energy-limited distributed edge clients present challenges such as heterogeneity, high energy consumption, and security risks. Traditional blockchain-based federated learning (BFL) struggles to address all three of these challenges simultaneously. This article proposes a Graph-Partitioning Multi-Granularity Federated Learning method on a consortium blockchain, namely GP-MGFL. To reduce the overall communication overhead, we adopt a balanced graph partitioning algorithm while introducing observer and consensus nodes. This method groups clients to minimize high-cost communications and focuses on the guidance effect within each group, thereby ensuring effective guidance with reduced overhead. To fully leverage heterogeneity, we introduce a cross-granularity guidance mechanism. This mechanism involves fine-granularity models guiding coarse-granularity models to enhance the accuracy of the latter models. We also introduce a credit model to adjust the contribution of models to the global model dynamically and to dynamically select leaders responsible for model aggregation. Finally, we implement a prototype system on real physical hardware and compare it with several baselines. Experimental results show that the accuracy of the GP-MGFL algorithm is 5.6% higher than that of ordinary BFL algorithms. In addition, compared to other grouping methods, such as greedy grouping, the accuracy of the proposed method improves by about 1.5%. In scenarios with malicious clients, the maximum accuracy improvement reaches 11.1%. We also analyze and summarize the impact of grouping and the number of clients on the model, as well as the impact of this method on the inherent security of the blockchain itself.

Federated Learning (FL) has become a common method for edge devices. Due to the limited energy capacity of edge devices, and the vulnerability of FL to malicious attacks from edge devices, vanilla FL still faces several challenges in edge computing, including energy consumption, model heterogeneity, and malicious behavior. To address these challenges, we propose a multi-granularity grouping-based federated learning (MG2FL), which groups and aggregates edge devices with low communication energy consumption and latency to reduce communication costs. Additionally, we introduce a multi-granularity guidance mechanism and a credit model to enhance model accuracy while ensuring security. Experimental results show that compared to the traditional FL algorithms, MG2FL achieves a 5.6% increase in accuracy, with the highest accuracy improvement reaching 11.1% in the presence of malicious edge devices.