feat: dynamic endpoint registration

components/src/dynamo/sglang/main.py

@@ -128,6 +128,9 @@ async def register_model():
 
     health_check_payload = SglangHealthCheckPayload(engine).to_dict()
 
+    test_endpoint = component.endpoint("test")
+    test_endpoint.register_custom_endpoint("/test")
+
     try:
         # Start endpoint immediately and register model concurrently
         # Requests queue until ready_event is set
@@ -139,6 +142,12 @@ async def register_model():
                 health_check_payload=health_check_payload,
             ),
             register_model(),
+            test_endpoint.serve_endpoint(
+                handler.generate,
+                graceful_shutdown=True,
+                metrics_labels=metrics_labels,
+                health_check_payload=health_check_payload,
+            ),
         )
     except Exception as e:
         logging.error(f"Failed to serve endpoints: {e}")

components/src/dynamo/sglang/request_handlers/multimodal_encode_worker_handler.py

@@ -5,7 +5,7 @@
 from typing import AsyncIterator
 
 import torch
-from sglang.srt.conversation import chat_templates
+from sglang.srt.parser.conversation import chat_templates
 from transformers import AutoImageProcessor, AutoModel, AutoTokenizer
 
 import dynamo.nixl_connect as connect

components/src/dynamo/sglang/utils/multimodal_chat_processor.py

@@ -3,7 +3,7 @@
 
 import logging
 
-from sglang.srt.conversation import chat_templates
+from sglang.srt.parser.conversation import chat_templates
 
 logger = logging.getLogger(__name__)
 

lib/bindings/python/rust/lib.rs

@@ -644,6 +644,39 @@ impl Component {
 
 #[pymethods]
 impl Endpoint {
+    #[pyo3(signature = (endpoint_path))]
+    fn register_custom_endpoint<'p>(
+        &self,
+        py: Python<'p>,
+        endpoint_path: &str,
+    ) -> PyResult<Bound<'p, PyAny>> {
+        // validate that the endpoint path looks like "/<string>" and does not end with a slash
+        if !endpoint_path.starts_with("/") || endpoint_path.ends_with("/") {
+            return Err(PyErr::new::<pyo3::exceptions::PyValueError, _>(
+                "Endpoint path must start with a slash and not end with a slash",
+            ));
+        }
+
+        let endpoint_path = endpoint_path.to_string();
+
+        let drt = self.inner.drt();
+        let etcd_client = drt.etcd_client();
+
+        pyo3_async_runtimes::tokio::future_into_py(py, async move {
+            if let Some(etcd_client) = etcd_client {
+                etcd_client
+                    .kv_create(
+                        endpoint_path.as_str(),
+                        serde_json::to_vec_pretty(&serde_json::Value::Null).unwrap(),
+                        None,
+                    )
+                    .await
+                    .map_err(to_pyerr)?;
+            }
+            Ok(())
+        })
+    }
+
     #[pyo3(signature = (generator, graceful_shutdown = true, metrics_labels = None, health_check_payload = None))]
     fn serve_endpoint<'p>(
         &self,

lib/bindings/python/src/dynamo/_core.pyi

@@ -115,6 +115,12 @@ class Endpoint:
 
     ...
 
+    def register_custom_endpoint(self, endpoint_path: str) -> None:
+        """
+        Register a custom endpoint path to our discovery plane by dynamo ingress
+        """
+        ...
+
     async def serve_endpoint(self, handler: RequestHandler, graceful_shutdown: bool = True, metrics_labels: Optional[List[Tuple[str, str]]] = None, health_check_payload: Optional[Dict[str, Any]] = None) -> None:
         """
         Serve an endpoint discoverable by all connected clients at