fix: Add model registration to SGLang multimodal workers fixing bug #4486 (#4512)

Signed-off-by: Nancy Agarwal <nancya@nvidia.com>
Co-authored-by: Kris Hung <krish@nvidia.com>
Co-authored-by: Ryan McCormick <rmccormick@nvidia.com>

components/src/dynamo/sglang/main.py

@@ -430,16 +430,24 @@ async def init_multimodal_encode_worker(runtime: DistributedRuntime, config: Con
 
     await pd_worker_client.wait_for_instances()
 
-    tasks = [
+    ready_event = asyncio.Event()
+
+    try:
+        await asyncio.gather(
             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,
+            ),
         )
-    ]
-
-    try:
-        await asyncio.gather(*tasks)
     except Exception as e:
         logging.error(f"Failed to serve endpoints: {e}")
         raise
@@ -473,11 +481,24 @@ async def init_multimodal_worker(runtime: DistributedRuntime, config: Config):
 
     await handler.async_init()
 
+    health_check_payload = SglangHealthCheckPayload(engine).to_dict()
+    ready_event = asyncio.Event()
+
     try:
-        await generate_endpoint.serve_endpoint(
+        await asyncio.gather(
+            generate_endpoint.serve_endpoint(
                 handler.generate,
                 metrics_labels=[("model", server_args.served_model_name)],
                 graceful_shutdown=True,
+                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}")
@@ -502,6 +523,7 @@ 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(
@@ -510,7 +532,14 @@ async def init_multimodal_prefill_worker(runtime: DistributedRuntime, config: Co
                 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}")