On-Demand 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.
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.
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.
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.
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.
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
Running AI/ML workloads in different clouds present unique challenges. The key to a manageable multi-cloud architecture is the ability to seamlessly access data across environments with high performance and low cost.
This webinar is designed for data platform engineers, data infra engineers, data engineers, and ML engineers who work with multiple data sources in hybrid or multi-cloud environments. Chanchan and Bin will guide the audience through using Alluxio to greatly simplify data access and make model training and serving more efficient in these environments.
You will learn:
- How to access data in multi-region, hybrid, and multi-cloud like accessing a local file system
- How to run PyTorch to read datasets and write checkpoints to remote storage with Alluxio as the distributed data access layer
- Real-world examples and insights from tech giants like Uber, AliPay and more
Cloud-native model training jobs require fast data access to achieve shorter training cycles. Accessing data can be challenging when your datasets are distributed across different regions and clouds. Additionally, as GPUs remain scarce and expensive resources, it becomes more common to set up remote training clusters from where data resides. This multi-region/cloud scenario introduces the challenges of losing data locality, resulting in operational overhead, latency and expensive cloud costs.
In the third webinar of the multi-cloud webinar series, Chanchan and Shawn dive deep into:
- The data locality challenges in the multi-region/cloud ML pipeline
- Using a cloud-native distributed caching system to overcome these challenges
- The architecture and integration of PyTorch/Ray+Alluxio+S3 using POSIX or RESTful APIs
- Live demo with ResNet and BERT benchmark results showing performance gains and cost savings analysis
In this presentation, Bin Fan (VP of Open Source @ Alluxio) will address a critical challenge of optimizing data loading for distributed Python applications within AI/ML workloads in the cloud, focusing on popular frameworks like Ray and Hugging Face. Integration of Alluxio’s distributed caching for Python applications is accomplished using the fsspec interface, thus greatly improving data access speeds. This is particularly useful in machine learning workflows, where repeated data reloading across slow, unstable or congested networks can severely affect GPU efficiency and escalate operational costs.
Attendees can look forward to practical, hands-on demonstrations showcasing the tangible benefits of Alluxio’s caching mechanism across various real-world scenarios. These demos will highlight the enhancements in data efficiency and overall performance of data-intensive Python applications. This presentation is tailored for developers and data scientists eager to optimize their AI/ML workloads. Discover strategies to accelerate your data processing tasks, making them not only faster but also more cost-efficient.
As GenAI and AI continue to transform businesses, scaling these workloads requires optimized underlying infrastructure. A multi-cloud architecture allows organizations to leverage different cloud services to meet diverse workload demands while maximizing efficiency, reducing costs, and avoiding vendor lock-in. However, achieving a multi-cloud vision can be challenging.
In this webinar, Tarik will share how an agonistic data layer, like Alluxio, allows you to embrace the separation of storage from compute and simplify the adoption of multi-cloud for AI.
- Learn why leveraging multiple cloud providers is critical for balancing performance, scalability, and cost of your AI platform
- Discover how an agnostic data layer like Alluxio provides seamless data access in multi-cloud that bridges storage and compute without data replication
- Gain insights into real-world examples and best practices for deploying AI across on-prem, hybrid, and multi-cloud environments
2024 is gearing up to be an impactful year for AI and analytics. Join us on January 30, as Kevin Petrie (VP of Research at Eckerson Group) and Omid Razavi (SVP of Customer Success at Alluxio) share key trends that data and AI leaders should know. This event will efficiently guide you with market data and expert insights to drive successful business outcomes.
- Assess current and future trends in data and AI with industry experts
- Discover valuable insights and practical recommendations
- Learn best practices to make your enterprise data more accessible for both analytics and AI applications
As a cache eviction algorithm, FIFO has a lot of attractive properties, such as simplicity, speed, scalability, and flash-friendliness. The most prominent criticism of FIFO is its low efficiency (high miss ratio). In this talk, Juncheng Yangb describes a simple, scalable FIFO-based algorithm with three static queues (S3-FIFO). Evaluated on 6594 cache traces from 14 datasets, we show that S3- FIFO has lower miss ratios than state-of-the-art algorithms across traces. Moreover, S3-FIFO’s efficiency is robust — it has the lowest mean miss ratio on 10 of the 14 datasets. FIFO queues enable S3-FIFO to achieve good scalability with 6× higher throughput compared to optimized LRU at 16 threads. Our insight is that most objects in skewed workloads will only be accessed once in a short window, so it is critical to evict them early (also called quick demotion). The key of S3-FIFO is a small FIFO queue that filters out most objects from entering the main cache, which provides a guaranteed demotion speed and high demotion precision.
In this session, Jingwen presents an overview of using Alluxio Edge caching to accelerate Trino or Presto queries. She offers practical best practices for using distributed caching with compute engines. In addition, this session also features insights from real-world examples.
In this session, cloud optimization specialists Chunxu and Siyuan break down the challenges and present a fresh architecture designed to optimize I/O across the data pipeline, ensuring GPUs function at peak performance. The integrated solution of PyTorch/Ray + Alluxio + S3 offers a promising way forward, and the speakers delve deep into its practical applications. Attendees will not only gain theoretical insights but will also be treated to hands-on instructions and demonstrations of deploying this cutting-edge architecture in Kubernetes, specifically tailored for Tensorflow/PyTorch/Ray workloads in the public cloud.
Shengxuan Liu from ByteDance presents the new ByteDance’s native Parquet Reader. The talk covers the architecture and key features of the Reader, and how the new Reader is able to facilitate data processing efficiency.