LLM inference demands substantial GPU resources, but its highly variable workload characteristics create efficiency challenges at both inter-GPU and intra-GPU levels. To meet Service Level Objectives (SLOs), existing systems typically overprovision resources in two ways: allocating excess GPUs to handle peak loads and reserving excessive memory per request to prevent out-of-memory during token generation. We introduce PiLLM (Predictable inference for LLMs), a system that addresses these inefficiencies through accurate workload prediction and dynamic resource allocation.

Large-scale multi-agent systems face two core challenges: inefficient policy learning and the explosion of state dimensions. Existing methods often rely on manually designed task sequences to guide agents’ learning in stages, but these designs lack adaptability to agents’ learning abilities, making it difficult to ensure the rationality of task difficulty. Moreover, the representation capability of current network structures is limited, making it challenging to efficiently handle high-dimensional state information and complex interaction relationships. To address these issues, we propose a Progressive Multi-Agent Reinforcement Learning (PMARL) framework. PMARL introduces a task adapter that adaptively selects task difficulty based on agents’ learning abilities, eliminating reliance on manual experience. Additionally, a Dynamic Dimension Adaptive Network (DDAN) is designed, incorporating hypernetwork and self-attention mechanisms to achieve adaptive feature extraction of high-dimensional states and efficient representation of agent interaction relationships. Experimental results demonstrate that PMARL exhibits higher efficiency and better adaptability compared to existing methods when addressing large-scale multi-agent tasks.

Junyi Chen, Shihao Bai, Zaijun Wang, Siyu Wu, Chuheng Du, Hailong Yang, Ruihao Gong, Shengzhong Liu, Fan Wu, Guihai Chen. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

The exploration and application of Large Language Models (LLMs) is thriving. To reduce deployment costs, continuous batching has become an essential feature in current service frameworks. The effectiveness of continuous batching relies on an accurate estimate of the memory requirements of requests. However, due to the diversity in request output lengths, existing frameworks tend to adopt aggressive or conservative schedulers, which often result in significant overestimation or underestimation of memory consumption. Consequently, they suffer from harmful request evictions or prolonged queuing times, failing to achieve satisfactory throughput under strict Service Level Agreement (SLA) guarantees (a.k.a. goodput), across various LLM application scenarios with differing input-output length distributions. To address this issue, we propose a novel Past-Future scheduler that precisely estimates the peak memory resources required by the running batch via considering the historical distribution of request output lengths and calculating memory occupancy at each future time point. It adapts to applications with all types of input-output length distributions, balancing the trade-off between request queuing and harmful evictions, thereby consistently achieving better goodput. Furthermore, to validate the effectiveness of the proposed scheduler, we developed a high-performance LLM serving framework, LightLLM, that implements the Past-Future scheduler. Compared to existing aggressive or conservative schedulers, LightLLM demonstrates superior goodput, achieving up to 2-3× higher goodput than other schedulers under heavy loads. LightLLM is open source to boost the research in such direction (https://github.com/ModelTC/lightllm).