Thesis abstract:

«p»Generative models have achieved remarkable success in creating realistic data samples, but their increasing size and complexity pose significant challenges for efficient deployment. This thesis explores algorithmic advancements and system design strategies to optimize large generative models for scalable and efficient implementation.«/p» «p»The research investigates key challenges in training and deploying large generative models, including computational complexity, memory constraints, and data throughput bottlenecks. It proposes a comprehensive optimization framework encompassing algorithmic innovations, such as model compression and advanced training strategies, and system design techniques, including efficient hardware architectures and optimized data pipelines.«/p» «p»The thesis presents evaluation metrics and benchmarks to assess the efficiency and scalability of the proposed optimization techniques, conducting experiments on diverse datasets and comparing performance against state-of-the-art baselines. The research highlights the potential impact of the optimization framework on real-world applications and discusses future research directions.«/p» «p»This thesis aims to bridge the gap between algorithmic innovations and system design techniques, providing a systematic approach to optimizing large generative models. The proposed framework has the potential to significantly enhance the efficiency and scalability of generative models, enabling their widespread adoption in practical applications and paving the way for future advancements in generative AI.«/p»