Scale document OCR batch inference for RAG on multiple clouds and Kubernetes clusters using SkyPilot Pool.

SkyPilot

DeepSeek-OCR processes scanned documents and legacy PDFs into clean markdown for RAG systems, but single-GPU processing is too slow for enterprise archives. SkyPilot Pools enables parallel batch inference across multiple clouds and Kubernetes clusters by creating a unified pool of GPU workers that stay warm and ready. Workers share setup costs (models, datasets, dependencies) and automatically distribute jobs using environment variables for work partitioning. The system handles automatic recovery, dynamic scaling, and cloud storage integration, making it practical to process hundreds of thousands of documents by utilizing idle GPU capacity across AWS, GCP, Nebius, and on-premise clusters.

Batch Inference for Documents with DeepSeek-OCR using a Pool of Workers on any Clouds

Scaling batch inference with SkyPilot pools #

Why pools work well for batch inference #

Beyond OCR: other batch inference use cases #