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.

Artificial Intelligence Generated Content (AIGC) services based on text-to-image generation tasks have driven change in the AI industry in recent years. Diffusion models are widely used in AIGC services for generating high-quality images from complex prompts. However, the process of generating diffusion models requires a large number of autoregressive denoising steps, posing significant challenges related to service latency and privacy issues, especially in resource-constrained edge devices. To further improve the quality of service (QoS) of edge AIGC services, we design a multi-edge cooperative caching mechanism based on the idea of reusing early intermediate results of similar prompts to reduce denoising steps. First, a collaborative filtering algorithm based on prompt similarity is developed to analyze user preferences. The designed multi-agent reinforcement learning-based cache management (MARLCM) algorithm utilizes the preference data as input and determines the caches that need to be replaced in the cache space. Secondly, we propose an adaptive edge server cooperation strategy that instructs servers to build efficient cache pools based on server similarity and workload differences. The experiment conducted on a high-resolution generation task with multiple diffusion cache frameworks shows that the proposed method reduces task latency by 23.1 %, and achieves up to 1.57 times higher cache hit rate compared to existing methods.