Prefill in LLM inference is known to be resource-intensive, especially for long LLM inputs. While better scheduling can mitigate prefill’s impact, it would be fundamentally better to avoid (most of) prefill. This talk introduces our preliminary effort towards drastically minimizing prefill delay for LLM inputs that naturally reuse text chunks, such as in retrieval-augmented generation. While keeping the KV cache of all text chunks in memory is difficult, we show that it is possible to store them on cheaper yet slower storage. By improving the loading process of the reused KV caches, we can still significantly speed up prefill delay while maintaining the same generation quality.
Prefill in LLM inference is known to be resource-intensive, especially for long LLM inputs. While better scheduling can mitigate prefill’s impact, it would be fundamentally better to avoid (most of) prefill. This talk introduces our preliminary effort towards drastically minimizing prefill delay for LLM inputs that naturally reuse text chunks, such as in retrieval-augmented generation. While keeping the KV cache of all text chunks in memory is difficult, we show that it is possible to store them on cheaper yet slower storage. By improving the loading process of the reused KV caches, we can still significantly speed up prefill delay while maintaining the same generation quality.
Video:
Presentation slides:
Videos:
Presentation Slides:
Complete the form below to access the full overview:
.png)
Videos
Pratik Mishra delivered insights on architecting scalable, deployable, and resilient AI infrastructure at scale. His discussion on fault tolerance, checkpoint optimization, and the democratization of AI compute through AMD's open ecosystem resonated strongly with the challenges teams face in production ML deployments.
In this talk, Bin Fan, VP of Technology at Alluxio, presents on building tiered architectures that bring sub-millisecond latency to S3-based workloads. The comparison showing Alluxio's 45x performance improvement over S3 Standard and 5x over S3 Express One Zone demonstrated the critical role the performance & caching layer plays in modern AI infrastructure.
In this talk, Greg Lindstrom shared how Blackout Power Trading achieved double-digit millisecond offline feature store performance using Alluxio, a game-changer for real-time power trading where every millisecond counts. The 60x latency reduction for inference queries was particularly impressive.