distributed compute

Run one job across many GPUs

Split a single job across GPUs on different machines, wired into one cluster over the VPN (torchrun/NCCL). Gang-scheduled — one rank per machine — and escrowed all-or-nothing: a cluster that can't fully form is refused and refunded.

Checking available machines…

Your cluster

GPUs — one per machine Max runtime (hours) Collective backend Private network (VPN) — get a WireGuard tunnel into your cluster Cluster self-test first — every rank runs a real all-reduce to prove the cluster communicates and reduces correctly (no image needed) Container image Command (each node runs it under torchrun)

Advanced — GPU class / region

Estimated cost shown once machines are available.

You prepay all N GPUs into escrow up-front. If the full cluster can't be reserved, you're charged nothing.

Already on Slurm / MPI / Ray?

Don't change your stack — Petabyte is another provider. Every rank registers its VPN address, then the cluster exports as the artifacts your launcher already reads:

GET /jobs/{id}/hostfile MPI / torchrun
GET /jobs/{id}/cluster nodes + launch cmds

Full recipes on the Developer API.