Skip to content

hopsworks.gpu_availability #

GpuAvailability #

GPUs a project can use, as reported by the Hopsworks cluster.

When NVIDIA Run:ai manages the cluster's GPUs, the numbers come from Run:ai for the project's Run:ai project, its department and its default node pool. Otherwise only the cluster's GPU inventory is filled in. Counts are in whole GPUs; None means unknown, or unlimited for a quota or limit.

Use Project.get_gpu_availability to get an instance of this class.

Returned by

available_non_preemptible property #

available_non_preemptible: float | None

GPUs that notebooks, terminals and model deployments can still get within the project's quota.

available_preemptible property #

available_preemptible: float | None

Idle GPUs a batch job could borrow over quota right now.

This is an estimate: a job asking for more is still submitted and waits in the project's queue.

department property #

department: dict | None

The Run:ai department's name, quota, limit and allocation.

expected names the department the project should be in when a Run:ai administrator has to move it.

node_pool property #

node_pool: str | None

The Run:ai node pool the numbers are for.

node_pool_gpus property #

node_pool_gpus: dict | None

The node pool's total, allocated and unallocated (idle) GPUs, with the sample time sampled_at.

project property #

project: dict | None

The Run:ai project's name, phase, quota, limit, allocated, allocated_non_preemptible and requested GPUs.

provider property #

provider: str | None

Where the numbers come from: RUN_AI or KUBERNETES.

warnings property #

warnings: list[str]

Conditions that make the numbers incomplete, such as RUN_AI_UNREACHABLE, RUN_AI_PROJECT_NOT_BOUND or RUN_AI_METRICS_MISSING.