S3-compatible storage with NFS and SMB for AI training

Mount object storage as a real filesystem on your GPU clusters. Regional caching gateways, managed buckets, and bring-your-own S3, GCS, or Azure.

Start for free
bucketfs

Try it out

# ── Use Training Pipes storage ──$ npx bucketfs buckets create --name datasetscreated bucket: datasets# ── Mount it locally with NFS near your compute ──$ npx bucketfs fs create --bucket datasets --region us-east-1 --name datasetscreated file system: datasets$ npx bucketfs fs connect datasetsmounted at ./datasets$ ls ./datasetsdata-000.parquet  data-001.parquet  data-002.parquet# ── Mount your existing S3/GCS buckets ──$ npx bucketfs fs connect my-s3-bucket <your credentials>mounted at ./my-s3-bucket$ ls ./my-s3-bucket/checkpointsepoch-08.pt  epoch-09.pt

What is Training Pipes?

Training Pipes lets you mount any S3-compatible bucket as a regular file system over NFS or SMB. Point your training jobs at a path and read your data at local-disk speeds — no rewrites, no custom SDKs, no copying datasets onto your GPU nodes.

How it works

1

Connect a bucket

Bring your own AWS S3, GCS, Azure Blob, or R2 bucket — or use ours. Read-only or read-write, your credentials never leave your account.

2

Create a file system

Spin up an NFS or SMB file system in the region closest to your compute. Configure cache size and write-back behavior in a few clicks.

3

Mount and train

Mount the share on your GPU nodes with a single command and start reading data. We handle caching, prefetching, and consistency.