multi-cloud storage gpu-scheduling hugging-face cost-optimization

SkyPilot mounts Hugging Face storage across clouds

Mount Hub models and datasets via hf:// URLs in SkyPilot jobs; no egress charges means compute placement decouples from storage location.

Summary

Teams running multi-cloud GPU clusters can now read models/datasets from a single Hugging Face Bucket without egress bills, eliminating the storage-location tax that forces compute to follow data. Lazy mounting keeps GPUs working while files stream in.

Why it matters

Implementation verdict

Replaces per-cloud data copies and S3/GCS mounts for teams already on Hub. Requires HF_TOKEN in environment; use `store: hf` with `hf://` URLs in SkyPilot YAML. Ready now—benchmark shows 30-second model load, 112–168 MB/s checkpoint writes across AWS/GCP/Lambda with identical config.

Sources

1.Mount a Hugging Face Bucket or any Hub repo into a SkyPilot job with one hf:// URL and the HF_TOKEN you already have, then launch it wherever capacity is
2.Hugging Face charges no egress, so reading your data onto those GPUs costs nothing, on any cloud
3.The model loaded free on every cloud. Lazy reads pull only what from_pretrained touches, so it was ready to train in about 30 seconds (up to ~500 MB/s)
4.storage at $12-18/TB/month (versus AWS S3 at roughly $23/TB plus egress)

Dev Signal

Get briefs like this in your inbox — free, every weekday.

100+ sources compressed into one 4-minute read. Ranked, cited, implementation-ready.

Read the full issue →All briefs