Redefining Caching for Generative AI - Approximate Caching in Text-to-Image Models and LLMs

Learn about revolutionary approaches to caching in Generative AI systems through this technical talk that explores novel approximate caching techniques for text-to-image diffusion models and large language models. Discover NIRVANA, an innovative system that reduces computational costs in diffusion models by reusing intermediate noise states, and Cache-Craft, which optimizes Retrieval-Augmented Generation (RAG) in LLMs through strategic reuse of precomputed key-value pairs. Gain insights from Dr. Subrata Mitra, a Senior Research Scientist at Adobe Research, as he presents research findings published in USENIX NSDI 2024 and upcoming in ACM SIGMOD 2025, demonstrating significant improvements in GPU compute efficiency, generation latency, and system throughput for Generative AI applications.