Overview
Long-term reservations of specific GPU units on a node for a set of organizations. Reserved capacity is exclusive to its audience and hidden from everyone else.
Reservations
Long-term reservations of specific GPU units on a node for a set of organizations. Reserved capacity is exclusive to its audience and hidden from everyone else.
Amend a reservation (platform admin)
End a reservation now (platform admin)
Get one node reservation
List a node's reservable GPU units
List a node's reservations
List GPU capacity reserved to this org
Reclaim released GPU units from spot
Release GPU units to the spot market
Reserve GPU units for orgs (platform admin)
Terminate spot workloads on a node POST
Terminates spot-tier workloads on a node your organization owns, or on reservations it manages there, and returns the VM ids it terminated and left running. A managing org (managingOrgIds) reaches only the released units of the reservations it manages, while each is active and inside its term. Scoped strictly to workloads bought at the spot tier: a spot instance is sold on the explicit contract that the capacity owner can terminate it at any time. Non-spot workloads on the node are never touched. Every termination is audited and the affected VM records capacity reclaim as its reason, so the customer is not told it was a platform fault.
List a node's reservable GPU units GET
Returns the allocatable GPU units on one node, with the unit ids that create, release, and reclaim take, and whether each unit can be reserved. Without this those calls are unusable: they are unit-scoped by design, and nothing else on the API returns a unit id, so the first reservation on a node could not be written at all. Each unit is one IOMMU group and is indivisible: a paired unit sells its GPUs together. Sellable units of a reservable class (vfio-group for VMs, device-request for pods) are reservable, and the response says which, and why not, rather than leaving the caller to infer it from a rejection. Readable by a platform operator, the org that owns the node, and any org managing a reservation on it. Anyone else gets 404, so the fleet stays unenumerable.