Olivier Tardieu

Pol G. Recasens, Ferran Agullo, Yue Zhu et al.

Large language models have been widely adopted across different tasks, but their auto-regressive generation nature often leads to inefficient resource utilization during inference. While batching is commonly used to increase throughput, performance gains plateau beyond a certain batch size, especially with smaller models, a phenomenon that existing literature typically explains as a shift to the compute-bound regime. In this paper, through an in-depth GPU-level analysis, we reveal that large-batch inference remains memory-bound, with most GPU compute capabilities underutilized due to DRAM bandwidth saturation as the primary bottleneck. To address this, we propose a Batching Configuration Advisor (BCA) that optimizes memory allocation, reducing GPU memory requirements with minimal impact on throughput. The freed memory and underutilized GPU compute capabilities can then be leveraged by concurrent workloads. Specifically, we use model replication to improve serving throughput and GPU utilization. Our findings challenge conventional assumptions about LLM inference, offering new insights and practical strategies for improving resource utilization, particularly for smaller language models. The code is publicly available at https://github.com/FerranAgulloLopez/vLLMBatchingMemoryGap.

Wenqing Yu, Neel Karia, Tanvi Hisaria et al.

The emergence of Multi-Instance GPU (MIG) technology enables us to run smaller machine learning models on partitions of a GPU rather than the entire device, thus improving utilization and reducing energy consumption, albeit with potential performance trade-offs. Meanwhile, the growing energy demands of GPU-equipped data centers motivate the development of online partitioning and scheduling schemes that not only ensure fast job processing but also achieve high energy efficiency. However, achieving energy-tardiness efficiency with manageable algorithmic complexity in large-scale scheduling remains a great challenge, due to the dual objectives of deciding on the GPU partitions and scheduling jobs onto the slices of the heterogeneous partitions. To address this challenge, we propose SMART-MIG, a parallel computing system that combines Mean-Field Multi-Agent Reinforcement Learning (MF-MARL) for large-scale MIG repartitioning with tailored heuristic algorithms for job scheduling. We demonstrate that the complexity of the repartitioning component remains constant even as the number of jobs and GPUs increases. We also establish theoretical lower bounds on energy consumption and tardiness to rigorously benchmark system performance. Finally, extensive experiments show that SMART-MIG improves the energy-tardiness efficiency by $18\%$ compared to its corresponding static-partitioning counterpart, while being only $27\%$ above the theoretical lower bound on energy consumption.

Pol G. Recasens, Yue Zhu, Chen Wang et al.

Large language models (LLMs) have revolutionized the state-of-the-art of many different natural language processing tasks. Although serving LLMs is computationally and memory demanding, the rise of Small Language Models (SLMs) offers new opportunities for resource-constrained users, who now are able to serve small models with cutting-edge performance. In this paper, we present a set of experiments designed to benchmark SLM inference at performance and energy levels. Our analysis provides a new perspective in serving, highlighting that the small memory footprint of SLMs allows for reaching the Pareto-optimal throughput within the resource capacity of a single accelerator. In this regard, we present an initial set of findings demonstrating how model replication can effectively improve resource utilization for serving SLMs.

Ferran Agullo, Joan Oliveras, Chen Wang et al.

Large Language Model (LLM) adapters enable low-cost model specialization, but introduce complex caching and scheduling challenges in distributed serving systems where hundreds of adapters must be hosted concurrently. While prior work has largely focused on latency minimization, resource efficiency through throughput maximization remains underexplored. This paper presents a data-driven pipeline that, for a given workload, computes an adapter placement that serves the workload with the minimum number of GPUs while avoiding request starvation and GPU memory errors. To that end, the approach identifies the maximum feasible throughput attainable on each GPU by leveraging accurate performance predictions learned from real serving behavior. The proposed pipeline integrates three components: (i) a Digital Twin (DT) tailored to LLM-adapter serving, (ii) a distilled machine learning (ML) model trained on DT-generated data, and (iii) a greedy placement algorithm that exploits ML-based performance estimates to maximize GPU efficiency. The DT emulates real system dynamics with high fidelity, achieving below 5% throughput estimation error while executing up to 90 times faster than full LLM benchmarking across both predictable and unpredictable workloads. The learned ML models further accelerate performance estimation with marginal accuracy degradation, enabling scalable optimization. Experimental results demonstrate that the pipeline substantially improves GPU efficiency by reducing the number of GPUs required to sustain target workloads. Beyond GPU efficiency, the pipeline can be adapted to alternative objectives, such as latency minimization, highlighting its versatility for future large-scale LLM serving infrastructures.

Martin Hirzel, Rodric Rabbah, Philippe Suter et al.

We discuss the suitability of spreadsheet processors as tools for programming streaming systems. We argue that, while spreadsheets can function as powerful models for stream operators, their fundamental boundedness limits their scope of application. We propose two extensions to the spreadsheet model and argue their utility in the context of programming streaming systems.