All pages
Documentation
Artifacts by 041 makes the buckets you own in AWS, GCP, Azure or any S3 endpoint look like one global store of versioned file sets. Every write lands in a bucket in the writer's region; reads go to wherever a version lives. The service hands out signed URLs and never touches the bytes.
Checkpointing a preemptible training run is the first use and the example these pages use: the run hops clouds, resumes the latest checkpoint with one cross-cloud read, and writes locally from then on.
Start
Getting started
An organization, a fallback location, an API key, and a run pointed at
o41://.The model
Artifacts, versions, the latest, and the one writer that opens from the latest, from nothing, or as a fork.
Checkpointing a run
SexpGPU and any other trainer on
o41://: resume the latest, save locally, fork.Setup
Connecting clouds
AWS, GCP and Azure by OIDC federation: the app is the issuer, your cloud trusts one subject.
S3 endpoints
Tigris, R2, a neocloud's object store or an existing bucket, added with a key and a placement tag.
Locations and placement
Where a write goes: cloud and region, then the cloud's default, then the fallback. How a box finds its place.
Reference
HTTP API
Every endpoint under
/api/v1, with request and response examples from the contract.Command line
npx @o41/artifacts: login, locations, API keys, listing, signed URLs, get and put.o41-checkpoint
The Rust crate: resume, save, fork and load on a directory, a raw bucket, or
o41://.Bucket layout
What sits in your bucket: one prefix per version,
_manifest.json written last.