[KVConnector] Support worker -> scheduler metadata (#31964)

Signed-off-by: Or Ozeri <oro@il.ibm.com>
Co-authored-by: Nicolò Lucchesi <nlucches@redhat.com>

vllm/v1/outputs.py

@@ -14,9 +14,13 @@ from vllm.v1.core.sched.output import SchedulerOutput
 
 if TYPE_CHECKING:
     from vllm.distributed.kv_events import KVConnectorKVEvents
+    from vllm.distributed.kv_transfer.kv_connector.v1.base import (
+        KVConnectorWorkerMetadata,
+    )
     from vllm.distributed.kv_transfer.kv_connector.v1.metrics import KVConnectorStats
 else:
     KVConnectorStats = object
+    KVConnectorWorkerMetadata = object
     KVConnectorKVEvents = object
 
 
@@ -142,6 +146,7 @@ class KVConnectorOutput:
     finished_recving: set[str] | None = None
     kv_connector_stats: KVConnectorStats | None = None
     kv_cache_events: KVConnectorKVEvents | None = None
+    kv_connector_worker_meta: KVConnectorWorkerMetadata | None = None
     # IDs of externally computed KV blocks that failed to load.
     # Requests referencing these blocks should be rescheduled to recompute them
     invalid_block_ids: set[int] = field(default_factory=set)
@@ -159,6 +164,7 @@ class KVConnectorOutput:
             and not self.kv_connector_stats
             and not self.kv_cache_events
             and not self.invalid_block_ids
+            and not self.kv_connector_worker_meta
         )
 
     @classmethod

vllm/v1/worker/kv_connector_model_runner_mixin.py

@@ -123,6 +123,7 @@ class KVConnectorModelRunnerMixin:
 
             output.kv_connector_stats = kv_connector.get_kv_connector_stats()
             output.kv_cache_events = kv_connector.get_kv_connector_kv_cache_events()
+            output.kv_connector_worker_meta = kv_connector.build_connector_worker_meta()
 
             if not defer_finalize:
                 kv_connector.clear_connector_metadata()