On-Demand Videos

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.

In this talk, Sandeep Manchem discussed big data and AI, covering typical platform architecture and data challenges. We had engaging discussions about ensuring data safety and compliance in Big Data and AI applications.

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.

‍

AI training workloads running on compute engines like PyTorch, TensorFlow, and Ray require consistent, high-throughput access to training data to maintain high GPU utilization. However, with the decoupling of compute and storage and with today’s hybrid and multi-cloud landscape, AI Platform and Data Infrastructure teams are struggling to cost-effectively deliver the high-performance data access needed for AI workloads at scale.

Join Tom Luckenbach, Alluxio Solutions Engineering Manager, to learn how Alluxio enables high-speed, cost-effective data access for AI training workloads in hybrid and multi-cloud architectures, while eliminating the need to manage data copies across regions and clouds.

What Tom will share:

• AI data access challenges in cross-region, cross-cloud architectures.​
• The architecture and integration of Alluxio with frameworks like PyTorch, TensorFlow, and Ray using POSIX, REST, or Python APIs across AWS, GCP and Azure.
• A live demo of an AI training workload accessing cross-cloud datasets leveraging Alluxio's distributed cache, unified namespace, and policy-driven data management.
• MLPerf and FIO benchmark results and cost-savings analysis.

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

OpenAI’s developer Developer Experience Engineer, Ankit Khare, provides practical insights for AI enthusiasts on effectively customizing and leveraging LLMs in various applications through preference tuning and fine-tuning.

In today’s AI-driven world, organizations face unprecedented demands for powerful AI infrastructure to fuel their model training and serving workloads. Performance bottlenecks, cost inefficiencies, and management complexities pose significant challenges for AI platform teams supporting large-scale model training and serving. On July 9, 2024, we introduced Alluxio Enterprise AI 3.2, a groundbreaking solution designed to address these critical issues in the ever-evolving AI landscape.

In this webinar, Shouwei Chen introduced exciting new features of Alluxio Enterprise AI 3.2:

• Leveraging GPU resources anywhere accessing remote data with the same local performance
• Enhanced I/O performance with 97%+ GPU utilization for popular language model training benchmarks
• Achieving the same performance as HPC storage on existing data lake without additional HPC storage infrastructure
• New Python FileSystem API to seamlessly integrate with Python applications like Ray
• Other new features, include advanced cache management, rolling upgrades, and CSI failover

As Trino users increasingly rely on cloud object storage for retrieving data, speed and cloud cost have become major challenges. The separation of compute and storage creates latency challenges when querying datasets; scanning data between storage and compute tiers becomes I/O bound. On the other hand, cloud API costs related to GET/LIST operations and cross-region data transfer add up quickly.

The newly introduced Trino file system cache by Alluxio aims to overcome the above challenges. In this session, Jianjian will dive into Trino data caching strategies, the latest test results, and discuss the multi-level caching architecture. This architecture makes Trino 10x faster for data lakes of any scale, from GB to EB.

What you will learn:

• Challenges relating to the speed and costs of running Trino in the cloud
• The new Trino file system cache feature overview, including the latest development status and test results
• A multi-level cache framework for maximized speed, including Trino file system cache and Alluxio distributed cache
• Real-world cases, including a large online payment firm and a top ridesharing company
• The future roadmap of Trino file system cache and Trino-Alluxio integration

Speed and efficiency are two requirements for the underlying infrastructure for machine learning model development. Data access can bottleneck end-to-end machine learning pipelines as training data volume grows and when large model files are more commonly used for serving. For instance, data loading can constitute nearly 80% of the total model training time, resulting in less than 30% GPU utilization. Also, loading large model files for deployment to production can be slow because of slow network or storage read operations. These challenges are prevalent when using popular frameworks like PyTorch, Ray, or HuggingFace, paired with cloud object storage solutions like S3 or GCS, or downloading models from the HuggingFace model hub.

In this presentation, Lu and Siyuan will offer comprehensive insights into improving speed and GPU utilization for model training and serving. You will learn:

• The data loading challenges hindering GPU utilization
• The reference architecture for running PyTorch and Ray jobs while reading data from S3, with benchmark results of training ResNet50 and BERT
• Real-world examples of boosting model performance and GPU utilization through optimized data access

From Caffe to MXNet, to PyTorch, and more, Xiande Cao, Senior Deep Learning Software Engineer Manager, will share his perspective on the evolution of deep learning frameworks.

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.

Uber has numerous deep learning models, most of which are highly complex with many layers and a vast number of features. Understanding how these models work is challenging and demands significant resources to experiment with various training algorithms and feature sets. With ML explainability, the ML team aims to bring transparency to these models, helping to clarify their predictions and behavior. This transparency also assists the operations and legal teams in explaining the reasons behind specific prediction outcomes.

In this talk, Eric Wang will discuss the methods Uber used for explaining deep learning models and how we integrated these methods into the Uber AI Michelangelo ecosystem to support offline explaining.