The traditional approach of transmitting sensing data to ground stations for training large AI models (LAMs) is becoming increasingly impractical due to unstable ground-to-satellite links (GSLs) and growing privacy concerns. Fortunately, the rapid expansion of Low Earth Orbit (LEO) satellite constellations offers a new solution. The increasing number of LEO satellites constitutes a distributed computing network, enabling on-orbit computation without the need for data transmission to ground stations. However, the limited computational resources on satellites result in substantial latency when training edge LAMs directly on a single satellite, and the memory capacity is insufficient to accommodate large models for training. To address these challenges, we propose a latency-optimized decentralized training framework that enables collaborative in-orbit LAM pretraining across heterogeneous LEO satellites without the need for downloading data to the ground. In the proposed architecture, we adopt a looped Transformer architecture that reduces memory usage through cross-layer parameter sharing. To improve training efficiency, we introduce a hybrid parallel training strategy that combines intra-group pipeline parallelism and inter-group data parallelism through satellite grouping. Furthermore, to accommodate heterogeneous satellite capabilities, we formulate an optimization problem for workload balancing and develop a dynamic programming–based strategy to minimize overall system latency. Extensive experiments on a heterogeneous satellite testbed demonstrate that our method significantly outperforms existing baselines in both training speed and memory efficiency.

The rapid advancements in computing and communication capabilities of Low Earth Orbit (LEO) satellites have made it feasible to execute complex and collaborative inorbit computation missions. Transformer-based large AI models (LAMs), known for their exceptional performance in in-context learning (ICL) and prompt-based reasoning, have attracted significant attention, providing powerful intelligence across sectors such as industry and aerospace. However, the significant parameter volume of LAMs poses a substantial challenge for direct deployment on satellites with constrained computing power and energy provision. To address this, the looped Transformer model reduces parameter requirements through layerwise parameter sharing, achieving performance comparable to vanilla Transformer-based LAMs in ICL tasks. Despite this efficiency, the limited and heterogeneous space-borne computing and storage capabilities complicate the orchestration for balanced workload allocation during multi-satellite cooperation. In this paper, we propose SAI, a collaborative multi-satellite space AI system that exploits the memory efficiency of the looped Transformer and the inherent parallelism in batch data processing. SAI enables accelerated on-satellite inference by integrating heterogeneous onboard resources and introducing a novel hybrid approach combining data and pipeline parallelism. This approach supports cross-satellite cooperation with parallelism planning and asynchronous inter-batch overlapping, significantly reducing inference latency and enhancing resource efficiency. Furthermore, SAI optimizes inference latency by formulating it as a shortest-path problem, effectively solved via Dijkstras algorithm. Extensive evaluations demonstrate SAIs superior performance in reducing inference latency and runtime memory usage compared to existing baselines.

Microservice architecture has revolutionized web service development by facilitating loosely coupled and independently developable components distributed as containers or virtual machines. While existing studies emphasize end-to-end latency, this paper investigates energy-efficient quality-of-service (QoS)-oriented microservice provisioning, focusing on both QoS satisfaction and power consumption (PC) conservation. We propose the Green and Intelligent Resource Provision (GIRP) architecture, integrating a data-driven energy-latency-aware resource allocation and scheduling manager to balance latency and PC. To reconcile the trade-offs involved, a dual-objective optimization problem is formulated to minimize latency and energy use by selecting proper servers, allocating CPU cores, and determining service replicas. To address challenges with discrete variables, dual objectives, and implicit mappings, we leverage a model-free deep deterministic policy gradient-based reinforcement learning algorithm. Specifically, we develop a multi-task agent via the Multi-gate Mixture-of-Experts model to simultaneously make two separate actions regarding CPU core numbers and service replica numbers, followed by a single-task agent to determine service scheduling. Extensive experiments on the DeathStarBenchmark testbed validate GIRP’s effectiveness, demonstrating approximately 52% resource savings and a 43% reduction in PC compared to leading methods like Sinan, Firm, and heuristic-based algorithms. These results highlight GIRP’s capability to optimize microservice orchestration by balancing end-to-end latency and power efficiency.