allocmode

type Mode struct {
	// APIVersion is the served resource.k8s.io API version the DRA probe ran
	// against (per APIVersionPreference: v1, v1beta2, or v1beta1), or ""
	// when the cluster serves no version of the group. Consumers that create
	// DRA resources (e.g. the secure-access ResourceClaim) must use it so the
	// probe and the behavioral test operate at the same group-version.
	APIVersion string

	// DRAUsable is true when the gpu.nvidia.com DeviceClass exists AND at
	// least one complete, current-generation ResourceSlice pool from driver
	// gpu.nvidia.com advertises an untainted device reachable from a Ready,
	// schedulable node.
	DRAUsable bool
	// DRADetail explains why DRA is or is not usable.
	DRADetail string
	// DRANodes is the sorted set of Ready, schedulable node names with
	// usable full-GPU DRA devices.
	DRANodes []string
	// DRANodeDevices maps each DRANodes entry to its USABLE full-GPU DRA
	// device count (untainted devices in complete, current-generation pools).
	// Consumers that size DRA-mode workloads (e.g. inference-perf worker
	// count) must use this — NOT scalar allocatable nvidia.com/gpu, which on
	// dual-advertised nodes can differ from the DRA-usable device count.
	DRANodeDevices map[string]int
	// NodeLocalGPUSliceDevices maps node name → RAW node-local
	// gpu.nvidia.com device count: slices attributed by raw spec.nodeName
	// only, devices counted without pool/taint validation. This mirrors how
	// kai-scheduler v0.14.1 attributes slices to nodes
	// (kubernetes_lister.go:198, cluster_info.go:270 — kai then rejects
	// scalar device-plugin GPU pods on every counted node, node_info.go (the scalar-GPU-on-DRA-node rejection, upstream-marked "temporary fix"))
	// and doubles as the SUPPORTED-configuration detector: in AICR's
	// supported full-GPU DRA state the NVIDIA driver publishes one
	// node-local pool per node, so a node with usable validated devices
	// (DRANodes) always appears here too. A node present here but absent
	// from DRANodes carries slices kai counts but AICR cannot validate.
	// Populated whenever the DRA API is served — independent of DeviceClass
	// existence, exactly like kai.
	NodeLocalGPUSliceDevices map[string]int
	// ForeignGPUSliceDrivers lists "gpu"-named ResourceSlice drivers OTHER
	// than gpu.nvidia.com (kai's driver predicate is lowercase-contains
	// "gpu" — dra.go:149 — so e.g. gpu.amd.com slices consume kai's shared
	// GPU capacity vector). AICR's inference validation does not model mixed
	// GPU DRA drivers: consumers fail fast when this is non-empty (#1652).
	// ComputeDomain drivers contain no "gpu" substring and never appear.
	ForeignGPUSliceDrivers []string
	// NonNodeLocalGPUSlices lists gpu.nvidia.com ResourceSlice names using
	// nodeSelector/allNodes/perDeviceNodeSelection topologies. kai cannot
	// node-attribute them (invisible to its capacity vector while the DRA
	// driver can still bind them), which AICR's inference validation does
	// not model: consumers fail fast when this is non-empty (#1652). The
	// conformance checks' generic topology validation is unaffected.
	NonNodeLocalGPUSlices []string
	// GPUPoolNodes maps each node-local gpu.nvidia.com ResourceSlice POOL
	// name to the node (spec.nodeName) whose slices publish it. The K8s API
	// documents that a pool name "is often the node name, but this is not
	// required" (resource/v1 ResourcePool.Name) — so occupancy attribution
	// must resolve pool → node through this slice-derived map, never through
	// pool==node name equality, which mis-attributes whenever a pool is
	// named after a DIFFERENT existing node. Pools listed in
	// AmbiguousGPUPools are excluded.
	GPUPoolNodes map[string]string
	// AmbiguousGPUPools lists gpu.nvidia.com pool names that node-local
	// slices of MULTIPLE nodes publish — pool → node attribution would be a
	// guess: consumers fail fast when this is non-empty (#1652). The NVIDIA
	// driver publishes one pool per node, so this is empty in AICR's
	// supported configurations.
	AmbiguousGPUPools []string

	// DevicePluginUsable is true when a Ready, schedulable node has
	// allocatable nvidia.com/gpu > 0. Kubernetes semantics: a Ready node
	// advertising scalar nvidia.com/gpu allocatable is device-plugin-backed
	// even when a DeviceClass maps that extended resource to DRA (KEP-5004
	// DRAExtendedResource) — the scheduler routes the extended-resource
	// request to DRA only on nodes where scalar allocatable is absent or
	// zero. See ExtendedResourceDRABacked for the recorded attribution caveat.
	DevicePluginUsable bool
	// DevicePluginDetail explains why the device plugin is or is not usable.
	DevicePluginDetail string
	// DevicePluginNodes is the sorted set of Ready, schedulable node names
	// advertising allocatable nvidia.com/gpu.
	DevicePluginNodes []string

	// DualAdvertisedNodes lists Ready, schedulable nodes where BOTH
	// mechanisms advertise GPUs. Dual advertisement risks GPU over-admission
	// (the same physical GPU allocatable through both paths). Currently a
	// warning only; a later PR promotes this to an error.
	DualAdvertisedNodes []string

	// ExtendedResourceDRABacked is true when a DeviceClass maps the
	// nvidia.com/gpu extended resource to DRA via spec.extendedResourceName
	// (KEP-5004 DRAExtendedResource). Recorded as attribution evidence only —
	// it does NOT clear DevicePluginUsable: on nodes with scalar allocatable
	// nvidia.com/gpu the request is device-plugin-served by definition; the
	// DRA mapping applies only where scalar allocatable is absent/zero.
	// Consumers that need deterministic per-node attribution should pin their
	// workload to a node with scalar allocatable (see the secure-access check).
	ExtendedResourceDRABacked bool
	// ExtendedResourceDetail explains the extended-resource attribution state.
	ExtendedResourceDetail string
}