Deploying large language models (LLMs) often requires specialized hardware and complex frameworks, creating barriers for CPU-based environments with resource constraints. These systems, common in air-gapped or edge scenarios, lack support for maintenance due to security, budget, or technical limits. To address this, we introduce TranSQL⁺, a compiler that translates LLM inference into SQL queries, enabling deployment on relational databases. By converting transformer operations into relational algebra, TranSQL⁺ generates vector-oriented SQL queries that leverage native database features (buffer management, indexing) to manage computations without hardware accelerators or deep learning frameworks. Demonstrated with the LLaMA3.1 8B model on DuckDB, results show relational databases can effectively serve LLMs, reducing deployment barriers and expanding access to advanced AI.

The rapid growth of large-scale machine learning (ML) models has led numerous commercial companies to utilize ML models for generating predictive results to help business decision-making. As two primary components in traditional predictive pipelines, data processing, and model predictions often operate in separate execution environments, leading to redundant engineering and computations. Additionally, the diverging mathematical foundations of data processing and machine learning hinder cross-optimizations by combining these two components, thereby overlooking potential opportunities to expedite predictive pipelines. In this paper, we propose an operator fusion method based on GPU-accelerated linear algebraic evaluation of relational queries. Our method leverages linear algebra computation properties to merge operators in machine learning predictions and data processing, significantly accelerating predictive pipelines by up to 317x. We perform a complexity analysis to deliver quantitative insights into the advantages of operator fusion, considering various data and model dimensions. Furthermore, we extensively evaluate linear algebra query processing and operator fusion utilizing the widely-used Star Schema and TPC-DI benchmarks. Through comprehensive evaluations, we demonstrate the effectiveness and potential of our approach in improving the efficiency of data processing and machine learning workloads on modern hardware.

Large language models (LLMs) have become central to many applications, but their deployment often requires high-performance hardware, specialized libraries, and complex engineering, limiting accessibility for smaller organizations. Meanwhile, relational database systems (RDBMS) are widely used for portability, efficiency, and native support for managing large-scale data operations. This paper presents TranSQL1, a toolkit that enables transformerbased LLM inference within RDBMS. By translating neural operations into SQL queries and representing model weights as relational tables, TranSQL leverages database features like dynamic disk-to-memory data management and caching to reduce hardware and engineering demands for serving LLMs. Using the LLaMA3 8B model, we demonstrate TranSQL's ability to implement attention layers, KV-cache, and end-to-end text generation through SQL queries. TranSQL offers a cost-effective, portable, and scalable approach to making advanced AI technologies more accessible.

Machine learning (ML) training data is often scattered across disparate collections of datasets, called <italic>data silos</italic>. This fragmentation poses a major challenge for data-intensive ML applications: integrating and transforming data residing in different sources demand a lot of manual work and computational resources. With data privacy constraints, data often cannot leave the premises of data silos; hence model training should proceed in a decentralized manner. In this work, we present a vision of bridging traditional data integration (DI) techniques with the requirements of modern machine learning systems. We explore the possibilities of utilizing metadata obtained from data integration processes for improving the effectiveness, efficiency, and privacy of ML models. Towards this direction, we analyze ML training and inference over data silos. Bringing data integration and machine learning together, we highlight new research opportunities from the aspects of systems, representations, factorized learning, and federated learning.

The proliferation of pre-trained ML models in public Web-based model zoos facilitates the engineering of ML pipelines to address complex inference queries over datasets and streams of unstructured content. Constructing optimal plan for a query is hard, especially when constraints (e.g. accuracy or execution time) must be taken into consideration, and the complexity of the inference query increases. To address this issue, we propose a method for optimizing ML inference queries that selects the most suitable ML models to use, as well as the order in which those models are executed. We formally define the constraint-based ML inference query optimization problem, formulate it as a Mixed Integer Programming (MIP) problem, and develop an optimizer that maximizes accuracy given constraints. This optimizer is capable of navigating a large search space to identify optimal query plans on various model zoos.

Machine learning (ML) training data is often scattered across disparate collections of datasets, called data silos. This fragmentation poses a major challenge for data-intensive ML applications: integrating and transforming data residing in different sources demand a lot of manual work and computational resources. With data privacy and security constraints, data often cannot leave the premises of data silos, hence model training should proceed in a decentralized manner. In this work, we present a vision of how to bridge the traditional data integration (DI) techniques with the requirements of modern machine learning. We explore the possibilities of utilizing metadata obtained from data integration processes for improving the effectiveness and efficiency of ML models. Towards this direction, we analyze two common use cases over data silos, feature augmentation and federated learning. Bringing data integration and machine learning together, we highlight new research opportunities from the aspects of systems, representations, factorized learning and federated learning.

Recent advances in Graphic Processing Units (GPUs) have facilitated a significant performance boost for database operators, in particular, joins. It has been intensively studied how conventional join implementations, such as hash joins, benefit from the massive parallelism of GPUs. With the proliferation of machine learning, more databases have started to provide native support for the basic building blocks of ML algorithms, i.e., linear algebra operators such as matrix multiplication (MM). Despite the recent increasing interest in processing relational joins using matrix multiplication (MM-join), two crucial questions still remain open: i) how efficient are current MM-join implementations compared to the GPU-based join algorithms; ii) how should practitioners choose among MM-join and conventional GPU-based joins given different data characteristics.In this paper, we compare the execution time, and memory I/O of MM-join against multiple GPU hash joins. An empirical analysis of our experimental results reveals that the state-of-the-art hash join implementation shows substantial scalability for various data characteristics. In contrast, MM-join outperforms the SOTA hash join in low join selectivity and low table cardinality but shows unsatisfactory scalability due to synchronous data movement and computation.

The rapid growth of large-scale machine learning (ML) models has led numerous commercial companies to utilize ML models for generating predictive results to help business decision-making. As two primary components in traditional predictive pipelines, data processing, and model predictions often operate in separate execution environments, leading to redundant engineering and computations. Additionally, the diverging mathematical foundations of data processing and machine learning hinder cross-optimizations by combining these two components, thereby overlooking potential opportunities to expedite predictive pipelines. In this paper, we propose an operator fusing method based on GPU-accelerated linear algebraic evaluation of relational queries. Our method leverages linear algebra computation properties to merge operators in machine learning predictions and data processing, significantly accelerating predictive pipelines by up to 317x. We perform a complexity analysis to deliver quantitative insights into the advantages of operator fusion, considering various data and model dimensions. Furthermore, we extensively evaluate matrix multiplication query processing utilizing the widely-used Star Schema Benchmark. Through comprehensive evaluations, we demonstrate the effectiveness and potential of our approach in improving the efficiency of data processing and machine learning workloads on modern hardware.