Skip to main content

Module placement

Module placement 

Source
Expand description

DeploymentPlacementActor — picks nodes for new deployments. Doc §7.2.

Reads transport_kind() to decide whether the deployment needs a GPU node (LocalGpu) or any egress-capable node (RemoteNetwork { provider }). This actor operates at the deployment → node level; once a node is chosen for a LocalGpu deployment, the which-GPU-on-this-node decision is delegated to the upstream rakka_accel::cuda::placement::PlacementActor (under the local-gpu feature). Renamed from the doc’s plain PlacementActor to make the abstraction-level distinction visible at the call site.

Without the local-gpu feature this actor still runs but the GPU ordinals it returns are a naïve 0..gpus_per_replica slice — fine for remote-only builds and for tests that don’t actually allocate CUDA contexts. Operators wanting topology-aware GPU choice (NVLink islands, MIG slicing, etc.) enable the feature and the deployment-level placement starts asking the upstream actor for per-node device choices.

Structs§

DeploymentPlacementActor
NodeAssignment
PlacementConstraints
PlacementResult

Enums§

PlacementError
PlacementMsg