In large language model inference, efficient utilization of GPU memory is of utmost importance. Current systems suffer from unreasonable GPU memory allocation: excessive memory is idle under low load, while memory shortages occur under high load. Moreover, existing GPU memory management methods do not consider the different memory requirements for prefill and decode stages in the inference process. This paper proposes a new solution - DynamicAttention. It allocates a continuous virtual GPU memory space at startup, but does not actually allocate physical GPU memory. Instead, it dynamically allocates memory according to the demand during runtime. Different memory consumption prediction algorithms are designed specifically for the Prefill and Decode stages. Experimental results show that compared with the existing SOTA solutions, DynamicAttention improves GPU memory utilization by 4 times under low load and 15% under high load. Furthermore, it increases throughput by 1.6 times.

Reinforcement learning (RL) has become a trending paradigm for training large language models (LLMs), particularly for reasoning tasks. Effective RL for LLMs requires massive parallelization and poses an urgent need for efficient training systems. Most existing large-scale RL systems for LLMs are synchronous by alternating generation and training in a batch setting, where the rollouts in each training batch are generated by the same (or latest) model. This stabilizes RL training but suffers from severe system-level inefficiency. Generation must wait until the longest output in the batch is completed before model update, resulting in GPU underutilization. We present AReaL, a fully asynchronous RL system that completely decouples generation from training. Rollout workers in AReaL continuously generate new outputs without waiting, while training workers update the model whenever a batch of data is collected. AReaL also incorporates a collection of system-level optimizations, leading to substantially higher GPU utilization. To stabilize RL training, AReaL balances the workload of rollout and training workers to control data staleness, and adopts a staleness-enhanced PPO variant to better handle outdated training samples. Extensive experiments on math and code reasoning benchmarks show that AReaL achieves up to 2.77x training speedup compared to synchronous systems with the same number of GPUs and matched or even improved final performance. The code of AReaL is available at https://github.com/inclusionAI/AReaL/.

Satellite Federated Learning (SFL) has emerged as a transformative paradigm for distributed machine learning in Low Earth Orbit (LEO) mega-constellations, enabling real-time processing of space-acquired data and enhancing remote sensing missions. The integration of Inter-Satellite Links (ISLs) into SFL alleviates synchronization delays due to intermittent connectivity between LEO satellites and ground-based parameter server (PS). However, the traffic generated from satellites creates communication bandwidth bottlenecks at the PS in the global model aggregation procedure. To address this issue, this paper proposes a novel SFL framework that leverages in-network model aggregation through ISLs to improve communication efficiency. Furthermore, two aggregation strategies, i.e., Chain Aggregation (CA) and Ring Aggregation (RA), are discussed in detail. Through system latency analysis and comprehensive simulations across constellation scales and data distributions, we demonstrate that: (1) in-network model aggregation fundamentally transforms communication load growth from ${\mathcal{O}}\left({{N^2}}\right)$ to ${\mathcal{O}}\left(N\right)$, and (2) the convergence time improves with increasing number of satellites, but degrades beyond a threshold.

Data centers struggle with growing cluster sizes and rising submissions of short-lived, high-frequency tasks that cause performance bottlenecks in task scheduling. Existing centralized and distributed scheduling systems fall short in meeting performance requirements due to computational overload on the scheduler, cluster state management overhead, and scheduling conflicts. To address these challenges, this paper introduces Beehive, a novel lightweight decentralized scheduling framework. In Beehive, each cluster node can schedule tasks within its local neighborhood, effectively reducing resource management overhead and scheduling conflicts. Moreover, all nodes are interconnected in a small-world network, an efficient structure that allows tasks to access resources across the entire cluster through global routing. This lightweight design enables Beehive to scale efficiently, supporting over 10,000 nodes and up to 80,000 task submissions per second without causing single-node scheduling bottlenecks. Experimental results demonstrate that Beehive significantly reduces scheduling latency. Specifically, 99% of tasks are scheduled within 100 milliseconds, and scheduling throughput can increase linearly with the number of nodes. Compared to existing centralized and distributed scheduling frameworks, Beehive substantially alleviates scheduling bottlenecks, particularly for high-frequency, short-lived tasks