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:
.png)
Videos
Scaling experimentation in digital marketplaces is crucial for driving growth and enhancing user experiences. However, varied methodologies and a lack of experiment governance can hinder the impact of experimentation leading to inconsistent decision-making, inefficiencies, and missed opportunities for innovation.
At Poshmark, we developed a homegrown experimentation platform, Lightspeed, that allowed us to make reliable and confident reads on product changes, which led to a 10x growth in experiment velocity and positive business outcomes along the way.
This session will provide a deep dive into the best practices and lessons learned from successful implementations of large-scale experiments. We will explore the importance of experimentation, overcome scalability challenges, and gain insights into the frameworks and technologies that enable effective testing.
In the rapidly evolving world of e-commerce, visual search has become a game-changing technology. Poshmark, a leading fashion resale marketplace, has developed Posh Lens – an advanced visual search engine that revolutionizes how shoppers discover and purchase items.
Under the hood of Posh Lens lies Milvus, a vector database enabling efficient product search and recommendation across our vast catalog of over 150 million items. However, with such an extensive and growing dataset, maintaining high-performance search capabilities while scaling AI infrastructure presents significant challenges.
In this talk, Mahesh Pasupuleti shares:
- The architecture and strategies to scale Milvus effectively within the Posh Lens infrastructure
- Key considerations include optimizing vector indexing, managing data partitioning, and ensuring query efficiency amidst large-scale data growth
- Distributed computing principles and advanced indexing techniques to handle the complexity of Poshmark’s diverse product catalog
As machine learning and deep learning models grow in complexity, AI platform engineers and ML engineers face significant challenges with slow data loading and GPU utilization, often leading to costly investments in high-performance computing (HPC) storage. However, this approach can result in overspending without addressing the core issues of data bottlenecks and infrastructure complexity.
A better approach is adding a data caching layer between compute and storage, like Alluxio, which offers a cost-effective alternative through its innovative data caching strategy. In this webinar, Jingwen will explore how Alluxio's caching solutions optimize AI workloads for performance, user experience and cost-effectiveness.
What you will learn:
- The I/O bottlenecks that slow down data loading in model training
- How Alluxio's data caching strategy optimizes I/O performance for training and GPU utilization, and significantly reduces cloud API costs
- The architecture and key capabilities of Alluxio
- Using Rapid Alluxio Deployer to install Alluxio and run benchmarks in AWS in just 30 minutes