Mats Brorsson

Xinlin Wang, Mats Brorsson

Despite the impressive capabilities of large language models, their substantial computational costs, latency, and privacy risks hinder their widespread deployment in real-world applications. Small Language Models (SLMs) with fewer than 10 billion parameters present a promising alternative; however, their inherent limitations in knowledge and reasoning curtail their effectiveness. Existing research primarily focuses on enhancing SLMs through scaling laws or fine-tuning strategies while overlooking the potential of using agent paradigms, such as tool use and multi-agent collaboration, to systematically compensate for the inherent weaknesses of small models. To address this gap, this paper presents the first large-scale, comprehensive study of <10B open-source models under three paradigms: (1) the base model, (2) a single agent equipped with tools, and (3) a multi-agent system with collaborative capabilities. Our results show that single-agent systems achieve the best balance between performance and cost, while multi-agent setups add overhead with limited gains. Our findings highlight the importance of agent-centric design for efficient and trustworthy deployment in resource-constrained settings.

Karthick Panner Selvam, Mats Brorsson

Deep Learning (DL) has developed to become a corner-stone in many everyday applications that we are now relying on. However, making sure that the DL model uses the underlying hardware efficiently takes a lot of effort. Knowledge about inference characteristics can help to find the right match so that enough resources are given to the model, but not too much. We have developed a DL Inference Performance Predictive Model (DIPPM) that predicts the inference latency, energy, and memory usage of a given input DL model on the NVIDIA A100 GPU. We also devised an algorithm to suggest the appropriate A100 Multi-Instance GPU profile from the output of DIPPM. We developed a methodology to convert DL models expressed in multiple frameworks to a generalized graph structure that is used in DIPPM. It means DIPPM can parse input DL models from various frameworks. Our DIPPM can be used not only helps to find suitable hardware configurations but also helps to perform rapid design-space exploration for the inference performance of a model. We constructed a graph multi-regression dataset consisting of 10,508 different DL models to train and evaluate the performance of DIPPM, and reached a resulting Mean Absolute Percentage Error (MAPE) as low as 1.9%.