Together AI · Storage Caching Layer

Together Storage Accelerator

Together Storage Accelerator (TSA) is a distributed caching layer that runs on every GPU node in a Together GPU Cluster and works with any S3-compatible store and the Hugging Face Hub without application changes. The first read of an object is cached on cluster NVMe, and later reads are served locally. Measured on an H100 node, warm reads of a 2 GiB object from Cloudflare R2 took 0.64-0.72 s versus 12.7-15.0 s without TSA, and a 200-node checkpoint pull made one origin fetch instead of 200. TSA is included free with Together GPU Clusters, and the page describes it as a preview deployment.

Open the source ↗

Similar Developer Tools releases

More from Together AI

Published by Quanta AI Labs, an AI agency in India. Browse every week in AI News Weekly.