December 2021 Summaries

This summary explains how to use Ray to speed up Deep Learning forecasting models for time series prediction by utilizing data and model parallelism. The process involves installing the necessary libraries, initializing Ray and its plugin for PyTorch Lightning, reading in sample data, converting it to PyTorch tensors and defining data loaders, training a PyTorch Forecasting model with the Ray Lightning plugin, and running the code on a laptop or any cloud using Anyscale. By leveraging parallel computing capabilities of Ray, developers can significantly reduce training times for these models, making them more efficient and scalable.

Miha, a data scientist at Ceneje, joined the company in March 2020 to explore business process optimization opportunities in content management, particularly through machine learning adoption. The team developed a system that uses machine learning for categorizing products on their website, which involves automatically classifying products by textual description and images. To improve this process, they considered adding another classifier, but it required careful architectural design to avoid technical debt and ensure scalability. They adopted Ray, a distributed processing framework, as the solution due to its scalability, library-agnostic model serving capabilities, and Python-first approach. The team implemented autoscaling using Ray Serve's experimental API, which initially lacked this feature, and also explored cloud cost optimization strategies. To maintain code maintainability, they utilized tools like Ray Tune for hyperparameter search and MLflow integration, allowing other teams to serve their models directly from MLflow.

XGBoost is a widely used gradient boosting algorithm that can be slow to train due to its computational requirements. To accelerate training, one approach is to change the tree construction method, such as using the GPU implementation of the histogram-based algorithm (`gpu_hist`) which can significantly reduce training time on larger datasets. Another approach is to leverage cloud computing by utilizing more resources than available locally, with XGBoost having support for GPUs on some operating systems. Additionally, distributing XGBoost model training with XGBoost-Ray, a distributed execution framework that leverages the actor model, can also speed up training by scaling applications and leveraging state-of-the-art machine learning libraries.

Artificial intelligence has a problem that's causing companies to struggle with scale, production, and expertise in building and managing AI infrastructure. To address this challenge, Ray and Anyscale have developed a managed cloud platform that enables developers and teams to successfully deploy and manage AI models without worrying about the underlying infrastructure. The platform provides an "infinite laptop experience" for development and experimentation, effortless deployment of AI models, and management of the end-to-end AI lifecycle on a single unified compute platform. With this platform, companies can accelerate development, scale their applications, and deploy machine learning in production with minimal hassle. Anyscale has also secured $100 million in Series C funding to further support its mission to enable every developer and organization to succeed with AI.

Ray version 1.9 has been released! The beta version of Ray Train includes usability improvements for distributed PyTorch training and checkpoint management, support for Ray Client, and an integration with Ray Datasets for distributed data ingest. Ray Datasets now supports groupby and aggregations, including multi-column/multi-lambda aggregations. Docker images for multiple CUDA versions are also available, allowing users to specify a `-cuXXX` suffix to pick a specific version. A new Ray Job Submission server + CLI & SDK clients is being launched, enabling users to package, deploy, and manage their Ray application as Jobs, which can be submitted by a Job manager of their choice.