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:
Complete the form below to access the full overview:
Videos
TorchTitan is a proof-of-concept for Large-scale LLM training using native PyTorch. It is a repo that showcases PyTorch's latest distributed training features in a clean, minimal codebase.
In this talk, Tianyu will share TorchTitan’s design and optimizations for the Llama 3.1 family of LLMs, spanning 8 billion to 405 billion parameters, and showcase its performance, composability, and scalability.
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.