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:
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
In the rapidly evolving landscape of AI and machine learning, Platform and Data Infrastructure Teams face critical challenges in building and managing large-scale AI platforms. Performance bottlenecks, scalability of the platform, and scarcity of GPUs pose significant challenges in supporting large-scale model training and serving.
In this talk, we introduce how Alluxio helps Platform and Data Infrastructure teams deliver faster, more scalable platforms to ML Engineering teams developing and training AI models. Alluxio’s highly-distributed cache accelerates AI workloads by eliminating data loading bottlenecks and maximizing GPU utilization. Customers report up to 4x faster training performance with high-speed access to petabytes of data spread across billions of files regardless of persistent storage type or proximity to GPU clusters. Alluxio’s architecture lowers data infrastructure costs, increases GPU utilization, and enables workload portability for navigating GPU scarcity challenges.
In this talk, Zhe Zhang (NVIDIA, ex-Anyscale) introduced Ray and its applications in the LLM and multi-modal AI era. He shared his perspective on ML infrastructure, noting that it presents more unstructured challenges, and recommended using Ray and Alluxio as solutions for increasingly data-intensive multi-modal AI workloads.
As large-scale machine learning becomes increasingly GPU-centric, modern high-performance hardware like NVMe storage and RDMA networks (InfiniBand or specialized NICs) are becoming more widespread. To fully leverage these resources, it’s crucial to build a balanced architecture that avoids GPU underutilization. In this talk, we will explore various strategies to address this challenge by effectively utilizing these advanced hardware components. Specifically, we will present experimental results from building a Kubernetes-native distributed caching layer, utilizing NVMe storage and high-speed RDMA networks to optimize data access for PyTorch training.