The KV Cache Bottleneck: Architectural Alternatives

This event will take place in Bengaluru on October 21, 2026, at 6:30 PM. Abhay Kumar, co-founder of FrontiersMind, will explore the challenges of cache size in autoregressive decoding and its implications for AI models. The session will cover architectural alternatives that address the KV cache bottleneck, including grouped-query attention and sliding-window attention, among others. Participants will gain insights into novel approaches like Grouped Value Attention, along with their respective advantages and limitations in real-world deployments.