fix: resolve sgl be E/P/D multimodal routing issues (#5500)

Signed-off-by: davilu <davilu@nvidia.com>

components/src/dynamo/sglang/main.py

@@ -487,23 +487,13 @@ async def init_multimodal_encode_worker(runtime: DistributedRuntime, config: Con
 
     await pd_worker_client.wait_for_instances()
 
-    ready_event = asyncio.Event()
-
     try:
-        await asyncio.gather(
-            generate_endpoint.serve_endpoint(
+        # Encode Worker is an internal component, should not register with Frontend
+        # Only needs to provide internal service endpoint for Processor to call
+        await generate_endpoint.serve_endpoint(
             handler.generate,
             graceful_shutdown=True,
             metrics_labels=[("model", server_args.served_model_name)],
-            ),
-            register_llm_with_readiness_gate(
-                None,  # encode worker doesn't have engine
-                generate_endpoint,
-                server_args,
-                dynamo_args,
-                input_type=ModelInput.Text,
-                readiness_gate=ready_event,
-            ),
         )
     except Exception as e:
         logging.error(f"Failed to serve endpoints: {e}")
@@ -542,6 +532,17 @@ async def init_multimodal_worker(runtime: DistributedRuntime, config: Config):
     ready_event = asyncio.Event()
 
     try:
+        if config.serving_mode == DisaggregationMode.DECODE:
+            # Decode Worker is an internal component, should not register with Frontend
+            # Only needs to provide internal service endpoint for Processor to call
+            await generate_endpoint.serve_endpoint(
+                handler.generate,
+                metrics_labels=[("model", server_args.served_model_name)],
+                graceful_shutdown=True,
+                health_check_payload=health_check_payload,
+            )
+        else:
+            # In aggregated mode, need to register with Frontend
             await asyncio.gather(
                 generate_endpoint.serve_endpoint(
                     handler.generate,
@@ -580,23 +581,15 @@ async def init_multimodal_prefill_worker(runtime: DistributedRuntime, config: Co
     await handler.async_init()
 
     health_check_payload = SglangPrefillHealthCheckPayload(engine).to_dict()
-    ready_event = asyncio.Event()
 
     try:
-        await asyncio.gather(
-            generate_endpoint.serve_endpoint(
+        # Prefill Worker is an internal component, should not register with Frontend
+        # Only needs to provide internal service endpoint for Decode Worker to call
+        await generate_endpoint.serve_endpoint(
             handler.generate,
             graceful_shutdown=True,
             metrics_labels=[("model", server_args.served_model_name)],
             health_check_payload=health_check_payload,
-            ),
-            register_llm_with_readiness_gate(
-                engine,
-                generate_endpoint,
-                server_args,
-                dynamo_args,
-                readiness_gate=ready_event,
-            ),
         )
     except Exception as e:
         logging.error(f"Failed to serve endpoints: {e}")

components/src/dynamo/sglang/multimodal_utils/multimodal_encode_utils.py

@@ -82,7 +82,18 @@ def is_model_supported(model_name: str, supported_model: str) -> bool:
     normalized_name = normalize_model_name(model_name).lower()
     normalized_supported = normalize_model_name(supported_model).lower()
 
-    return normalized_name == normalized_supported
+    # Exact match
+    if normalized_name == normalized_supported:
+        return True
+
+    # Handle local path case: compare only the model name part (without organization)
+    # e.g., "qwen2.5-vl-7b-instruct" matches "qwen/qwen2.5-vl-7b-instruct"
+    if "/" in normalized_supported:
+        model_part = normalized_supported.split("/")[-1]
+        if normalized_name == model_part:
+            return True
+
+    return False
 
 
 def get_qwen_image_features(

components/src/dynamo/sglang/request_handlers/multimodal/worker_handler.py

@@ -237,7 +237,7 @@ class MultimodalWorkerHandler(BaseWorkerHandler):
         config: Config,
         prefill_client: Client = None,
     ):
-        super().__init__(component, engine, config, None, prefill_client)
+        super().__init__(component, engine, config, None)
 
         # Initialize processors
         self.embeddings_processor = EmbeddingsProcessor()