feat: add prefill worker metrics support for vLLM (#3949)

Signed-off-by: Keiven Chang <keivenchang@users.noreply.github.com>

components/src/dynamo/vllm/handlers.py

@@ -12,6 +12,7 @@ from vllm.inputs import TokensPrompt
 from vllm.sampling_params import SamplingParams
 from vllm.v1.engine.exceptions import EngineDeadError
 
+from dynamo.llm import ZmqKvEventPublisher
 from dynamo.runtime.logging import configure_dynamo_logging
 
 from .engine_monitor import VllmEngineMonitor
@@ -62,7 +63,7 @@ class BaseWorkerHandler(ABC):
         self.component = component
         self.engine_client = engine
         self.default_sampling_params = default_sampling_params
-        self.kv_publishers = None
+        self.kv_publishers: list[ZmqKvEventPublisher] | None = None
         self.engine_monitor = VllmEngineMonitor(runtime, engine)
 
     @abstractmethod

components/src/dynamo/vllm/main.py

@@ -274,6 +274,11 @@ async def init_prefill(runtime: DistributedRuntime, config: Config):
     if kv_publishers:
         handler.kv_publishers = kv_publishers
 
+    if config.engine_args.disable_log_stats is False:
+        register_engine_metrics_callback(
+            endpoint=generate_endpoint, registry=REGISTRY, metric_prefix_filter="vllm:"
+        )
+
     # Register prefill model with ModelType.Prefill
     if not config.engine_args.data_parallel_rank:  # if rank is 0 or None then register
         await register_vllm_model(