refactor: add explicit non-leader node handling in vLLM (#5597)

Signed-off-by: alec-flowers <aflowers@nvidia.com>

.github/workflows/pr.yaml

@@ -148,7 +148,7 @@ jobs:
           - { major_minor: '12.9', major: '12' }
     name: vllm-build-test (cuda${{ matrix.cuda_version.major_minor}}, ${{ matrix.platform.arch }})
     runs-on: ${{ matrix.platform.runner }}
-    timeout-minutes: 90
+    timeout-minutes: 240
     env:
       FRAMEWORK: vllm
     steps: &runtime-container-build-push-test

components/src/dynamo/vllm/main.py

@@ -48,6 +48,19 @@ configure_dynamo_logging()
 logger = logging.getLogger(__name__)
 
 
+async def _handle_non_leader_node(dp_rank: int) -> None:
+    """
+    Handle non-leader node (data_parallel_rank >= 1) in multi-node deployments.
+    Non-leader nodes run vLLM workers but don't serve Dynamo endpoints.
+    """
+    logger.info(
+        f"Non-leader node detected (data_parallel_rank={dp_rank}). "
+        "Skipping endpoint serving."
+    )
+    # Wait indefinitely - process terminated via signal handlers
+    await asyncio.Event().wait()
+
+
 async def graceful_shutdown(runtime):
     """
     Shutdown dynamo distributed runtime.
@@ -452,11 +465,13 @@ async def init_prefill(runtime: DistributedRuntime, config: Config):
     runtime.register_engine_route("wake", handler.wake)
     logger.info("Registered engine routes: /engine/sleep, /engine/wake")
 
+    # Handle non-leader nodes - don't serve endpoints
+    if config.engine_args.data_parallel_rank:
+        await _handle_non_leader_node(config.engine_args.data_parallel_rank)
+        return
+
     # Register prefill model with ModelType.Prefill
-    if not config.engine_args.data_parallel_rank:  # if rank is 0 or None then register
-        model_input = (
-            ModelInput.Text if config.use_vllm_tokenizer else ModelInput.Tokens
-        )
+    model_input = ModelInput.Text if config.use_vllm_tokenizer else ModelInput.Tokens
     await register_vllm_model(
         model_input,
         ModelType.Prefill,
@@ -575,16 +590,16 @@ async def init(runtime: DistributedRuntime, config: Config):
     runtime.register_engine_route("wake", handler.wake)
     logger.info("Registered engine routes: /engine/sleep, /engine/wake")
 
-    if not config.engine_args.data_parallel_rank:  # if rank is 0 or None then register
+    # Handle non-leader nodes - don't serve endpoints
+    if config.engine_args.data_parallel_rank:
+        await _handle_non_leader_node(config.engine_args.data_parallel_rank)
+        return
+
     # Parse endpoint types from --dyn-endpoint-types flag
     model_type = parse_endpoint_types(config.dyn_endpoint_types)
-        logger.info(
-            f"Registering model with endpoint types: {config.dyn_endpoint_types}"
-        )
+    logger.info(f"Registering model with endpoint types: {config.dyn_endpoint_types}")
 
-        model_input = (
-            ModelInput.Text if config.use_vllm_tokenizer else ModelInput.Tokens
-        )
+    model_input = ModelInput.Text if config.use_vllm_tokenizer else ModelInput.Tokens
 
     # Warn if custom template provided but chat endpoint not enabled
     if config.custom_jinja_template and "chat" not in config.dyn_endpoint_types: