The demand for efficient Large Language Model (LLM) inference has surged with the rising adoption of Generative AI (GenAI) applications, particularly in areas such as agents and retrieval-augmented generation. Efficient inference serves two crucial purposes: it enables the deployment of LLM-centered applications that address critical business needs, while also facilitating rapid experimentation for researchers to extract valuable insights and new understandings. However, despite the field's rapid advancement and interdisciplinary nature, there remains a limited exchange of ideas and methodologies between production-facing practitioners and researchers seeking to experiment with new GenAI concepts quickly. To bridge this gap, we are introducing the first KDD workshop on Inference Optimization for Generative AI. Our goal is to create a collaborative platform where researchers and practitioners working across various use cases and stacks of efficient inference can come together to exchange research ideas, establish connections between different disciplines, and identify challenges and research questions that will shape future work.

There have been many proposals on how media-on-demand servers can effectively allow clients to share resources. In this paper, given a set of clients, we show how these clients may be partitioned into "sync-classes" sets of clients who can be serviced through allocation of a single set of resources. As a set of clients may be partitioned into sync-classes in many different ways, we show that a very large class of cost functions may be used to determine which partition to choose. We provide algorithms to compute such optimal splits. Our framework is very generic in the following ways: the system may plug-in any cost function whatsoever, as long as it satisfies four common-sense axioms that evaluate costs; and the system may evaluate the future anticipated requests of a user using any user model (e.g., a Markovian model) that has a specified I/O interface. Thus, a wide variety of predictive methods (of what the user will do) and a wide variety of costing methods may be used within our framework.

Though numerous multimedia systems exist in the commercial market today, relatively little work has been done on developing the mathematical foundation of multimedia technology. We attempt to take some initial steps towards the development of a theoretical basis for a multimedia information system. To do so, we develop the motion of a structured multimedia database system. We begin by defining a mathematical model of a media-instance. A media-instance may be thought of as “glue” residing on top of a specific physical media-representation (such as video, audio, documents, etc). Using this “glue”, it is possible to define a general purpose logical query language to query multimedia data. This glue consists of a set of “states” (e.g., video frames, audio tracks, etc.) and “features”, together with relationships between states and/or features. A structured multimedia database system imposes a certain mathematical structures on the set of features/states. Using this notion of a structure, we are able to define indexing structures for processing queries, methods to relax queries when answers do not exist to those queries, as well as sound, complete and terminating procedures to answer such queries (and their relaxations, when appropriate). We show how a media-presentation can be generated by processing a sequence of queries, and furthermore we show that when these queries are extended to include constraints , then these queries can not only generate presentations, but also generate temporal synchronization properties and spatial layout properties for such presentations. We describe the architecture of a prototype multimedia database system based on the principles described in this paper.

The declarative semantics of nonmonotonic logic programming has largely been based on propositional programs. However, the ground instantiation of a logic program may be very large, and likewise, a ground stable model may also be very large. We develop a non-ground semantic theory for non-monotonic logic programming. Its principal advantage is that stable models and well-founded models can be represented as sets of atoms, rather than as sets of ground atoms. A set SI of atoms may be viewed as a compact representation of the Herbrand interpretation consisting of all ground instances of atoms in SI. We develop generalizations of the stable and well-founded semantics based on such non-ground interpretations SI. The key notions for our theory are those of covers and anticovers. A cover as well as its anticover are sets of substitutions — non-ground in general — representing all substitutions obtained by ground instantiating some substitution in the (anti)cover, with the additional requirement that each ground substitution is represented either by the cover or by the anticover, but not by both. We develop methods for computing anticovers for a given cover, show that membership in so-called optimal covers is decidable, and investigate the complexity in the Datalog case.

Relatively little work has been done on developing the mathematical foundations of multimedia technology. S. Marcus and V.S. Subrahmanian (1995; 1994) have taken some initial steps toward the development of a theoretical basis for multimedia information systems. They define a mathematical model of a media instance. A media instance may be thought of as "glue" residing on top of a specific physical media representation (such as video, audio, documents, etc.) Using this "glue", it is possible to define a general purpose logical query language to query multimedia data. This glue consists of a set of "states" and "features", together with relationships between states and/or features. A structured multimedia database system imposes a certain mathematical structure on the set of features/states. Using this notion of a structure, they are able to define indexing structures for processing queries, methods to relax queries when answers do not exist to those queries, as well as sound, complete and terminating procedures to answer such queries. Using the S. Marcus and V.S. Subrahmanian (1995; 1994) work on multimedia database integration systems, we show how their logical based query language can be redefined as an SQL based query language. As there are numerous commercial SQL database systems, a wide and diverse population of users may access the Marcus and Subrahmanian work using their SQL interface. Also, there has been a great deal of study of query optimization of SQL based languages.