Disaggregate prefill and decode: prefill block, role-aware scheduler, and InferencePool routing

apis/modelreplicas/definition.yaml

@@ -118,7 +118,6 @@ spec:
                   count:
                     type: integer
                     minimum: 1
-                    default: 1
                   topology:
                     type: object
                     required: [tensor]
@@ -214,5 +213,185 @@ spec:
                               properties:
                                 name:
                                   type: string
+              prefill:
+                type: object
+                description: >-
+                  Prefill role placement for a disaggregated replica. Mirrors the
+                  top-level decode fields (workers, nodePoolName, deviceRequests),
+                  pinned to the prefill pool the scheduler chose. Absent for
+                  unified replicas. When present, all three fields are required:
+                  compose-model-deployment always populates them, and the llm-d
+                  backend reads them unconditionally.
+                required: [workers, nodePoolName, deviceRequests]
+                properties:
+                  workers:
+                    type: object
+                    required: [topology, template]
+                    properties:
+                      count:
+                        type: integer
+                        minimum: 1
+                      topology:
+                        type: object
+                        required: [tensor]
+                        properties:
+                          tensor:
+                            type: integer
+                            minimum: 1
+                          pipeline:
+                            type: integer
+                            minimum: 1
+                            default: 1
+                      template:
+                        type: object
+                        properties:
+                          metadata:
+                            type: object
+                            properties:
+                              labels:
+                                type: object
+                                additionalProperties:
+                                  type: string
+                              annotations:
+                                type: object
+                                additionalProperties:
+                                  type: string
+                          spec:
+                            type: object
+                            required: [containers]
+                            properties:
+                              containers:
+                                type: array
+                                minItems: 1
+                                maxItems: 1
+                                x-kubernetes-validations:
+                                - rule: "self.exists_one(c, c.name == 'engine')"
+                                  message: "the single container must be named 'engine'"
+                                items:
+                                  type: object
+                                  required: [name, image]
+                                  properties:
+                                    name:
+                                      type: string
+                                      minLength: 1
+                                    image:
+                                      type: string
+                                      minLength: 1
+                                    command:
+                                      type: array
+                                      items:
+                                        type: string
+                                    args:
+                                      type: array
+                                      items:
+                                        type: string
+                                    env:
+                                      type: array
+                                      items:
+                                        type: object
+                                        required: [name]
+                                        properties:
+                                          name:
+                                            type: string
+                                          value:
+                                            type: string
+                                          valueFrom:
+                                            type: object
+                                            properties:
+                                              secretKeyRef:
+                                                type: object
+                                                required: [name, key]
+                                                properties:
+                                                  name:
+                                                    type: string
+                                                  key:
+                                                    type: string
+                                                  optional:
+                                                    type: boolean
+                                              configMapKeyRef:
+                                                type: object
+                                                required: [name, key]
+                                                properties:
+                                                  name:
+                                                    type: string
+                                                  key:
+                                                    type: string
+                                                  optional:
+                                                    type: boolean
+                              imagePullSecrets:
+                                type: array
+                                items:
+                                  type: object
+                                  required: [name]
+                                  properties:
+                                    name:
+                                      type: string
+                  nodePoolName:
+                    type: string
+                    description: The prefill pool on this replica's cluster.
+                  deviceRequests:
+                    type: array
+                    description: >-
+                      Resolved DRA device requests for the matched pool. The parent
+                      ModelDeployment's compose function joins the nodeSelector
+                      requests with the matched InferenceClass devices and stamps the
+                      claim: DRA devices here. This function turns each into a
+                      DeviceRequest in a DRA ResourceClaim for the serving pods. At
+                      least one request is always present: the scheduler only pins a
+                      replica to a pool that yields a claimable device, so the
+                      serving workload always has a ResourceClaim to bind through.
+                    minItems: 1
+                    maxItems: 16
+                    x-kubernetes-list-type: map
+                    x-kubernetes-list-map-keys: [name]
+                    items:
+                      type: object
+                      required: [name, deviceClassName]
+                      properties:
+                        name:
+                          type: string
+                          description: Request name; becomes the DeviceRequest name.
+                          minLength: 1
+                          maxLength: 63
+                        deviceClassName:
+                          type: string
+                          description: >-
+                            Cluster-scoped DRA DeviceClass to claim through, from the
+                            matched InferenceClass device.
+                          minLength: 1
+                          maxLength: 253
+                        count:
+                          type: integer
+                          description: How many devices to claim.
+                          default: 1
+                          minimum: 1
+                          maximum: 64
+                        selectors:
+                          type: array
+                          description: >-
+                            DRA CEL selectors copied verbatim from the nodeSelector
+                            request, ANDed in the DeviceRequest.
+                          maxItems: 8
+                          x-kubernetes-list-type: atomic
+                          items:
+                            type: object
+                            # A selector must carry a selector kind (today only cel).
+                            # An empty {} selector would match any device, silently
+                            # widening the DRA claim the serving pod binds through.
+                            minProperties: 1
+                            properties:
+                              cel:
+                                type: string
+                                minLength: 1
+                                maxLength: 10240
+              routing:
+                type: object
+                properties:
+                  template:
+                    type: object
+                    properties:
+                      spec:
+                        type: object
+                        x-kubernetes-preserve-unknown-fields: true
           status:
             type: object