Can I test jobs on specific GPU hardware like H100s?

Yes, the skill provides specific workflows for testing large models on Slurm-managed H100 or H200 clusters when local A10 development machine memory is insufficient.

What is Ray Data used for in this skill?

Ray Data is used for large-scale ML workloads involving batch inference, data preprocessing, and distributed ETL with coordinated CPU/GPU resources.

How does this skill handle production deployments?

It includes instructions for submitting jobs to a production RayCluster using the Ray Client and setting up scheduled execution via Airflow BashOperators.

Does it support large language model inference?

Yes, it provides specific code patterns for using vLLM for batch inference and handling complex vision-language models like Qwen-VL.

How do I optimize job performance?

The skill includes best practices for batch sizing, concurrent processing, and managing ActorPoolStrategy to maximize GPU utilization and prevent memory errors.

Ray Data Development

Name: Ray Data Development
Author: SerendipityOneInc

bySerendipityOneInc

•

数据科学与机器学习

Streamlines the development, testing, and deployment of large-scale Ray Data jobs for distributed ML workloads.

关于

The Ray Data Development skill provides specialized guidance for building high-performance data processing pipelines using Ray Data. It supports the entire software lifecycle, from local development on A10 machines and testing on high-performance Slurm clusters (H100/H200) to production deployment on RayCluster and automation via Airflow. By offering pre-configured patterns for GPU/CPU coordination, model-based processing, and vLLM batch inference, it helps developers optimize GPU utilization, manage memory efficiently, and troubleshoot common scaling issues in massive ML datasets.

主要功能

Optimized GPU/CPU coordination patterns for high-throughput ML batch inference
Multi-environment workflows covering Local A10, Slurm clusters, and production RayCluster
1 GitHub stars
Pre-built implementation templates for image classification and vLLM pipelines
Performance tuning and troubleshooting strategies for distributed data workloads
Deployment and scheduling guides for production environments including Airflow and JuiceFS

使用场景

Processing large-scale datasets for machine learning training or distributed batch inference
Migrating local data processing scripts to multi-node Ray clusters for production scale
Deploying scalable ETL pipelines that require high-performance GPU acceleration

关于

主要功能

Optimized GPU/CPU coordination patterns for high-throughput ML batch inference
Multi-environment workflows covering Local A10, Slurm clusters, and production RayCluster
1 GitHub stars
Pre-built implementation templates for image classification and vLLM pipelines
Performance tuning and troubleshooting strategies for distributed data workloads
Deployment and scheduling guides for production environments including Airflow and JuiceFS

使用场景

Processing large-scale datasets for machine learning training or distributed batch inference
Migrating local data processing scripts to multi-node Ray clusters for production scale
Deploying scalable ETL pipelines that require high-performance GPU acceleration