Compute routing

Require workload specification, hardware/memory, data locality, network and secret policy, quota, latency, cost ceiling, persistence, and cleanup approval.

Workflow

1. Preflight local or remote hardware, driver/framework compatibility, quota, expected duration, cost, and failure modes.
2. Select local CPU/GPU, HF Jobs/Sandbox/Spaces, Kaggle, Modal, SSH, or Slurm by constraints; record why alternatives were rejected.
3. Stage immutable code, data, model, environment, and secrets-by-reference. Define submit, status, logs, inspect, cancel, artifact, cleanup, heartbeat, retry, and resume actions.
4. Require explicit approval before paid, remote, secret-bearing, destructive, or publication actions.
5. Emit a bounded provider plan or executable job handoff to ml-stack-training and diagnostics to ml-stack-tracking; never claim submission when no registry operation exists.

Outputs and recovery

Return provider, resource profile, commands or API intent, estimated cost, approval state, run identity, artifact paths, heartbeat and orphan recovery policy. Abstain when quota, cost, data access, or resume point is unknown. Next owner: ml-stack-training for job content, ml-stack-tracking for runtime diagnostics, and ml-stack-audit for evidence reconciliation.

See provider routing, sandbox and jobs, resource profiles, resume and orphans, and cost controls.