feat: add --endpoint flag support to dynamo.vllm (#6360)

Signed-off-by: alec-flowers <aflowers@nvidia.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

components/src/dynamo/common/configuration/groups/runtime_args.py

@@ -74,7 +74,7 @@ class DynamoRuntimeArgGroup(ArgGroup):
             flag_name="--endpoint",
             env_var="DYN_ENDPOINT",
             default=None,
-            help="Dynamo endpoint string in 'dyn://namespace.component.endpoint' format. Example: dyn://dynamo.backend.generate. Currently used only by TRT-LLM and SGLang backends.",
+            help="Dynamo endpoint string in 'dyn://namespace.component.endpoint' format. Example: dyn://dynamo.backend.generate.",
         )
         add_argument(
             g,

components/src/dynamo/vllm/args.py

@@ -21,6 +21,7 @@ from dynamo.common.configuration.groups.runtime_args import (
     DynamoRuntimeArgGroup,
     DynamoRuntimeConfig,
 )
+from dynamo.common.utils.runtime import parse_endpoint
 from dynamo.vllm.backend_args import DynamoVllmArgGroup, DynamoVllmConfig
 
 from . import envs
@@ -152,6 +153,9 @@ def update_dynamo_config_with_engine(
     else:
         dynamo_config.served_model_name = None
 
+    # Capture user-provided --endpoint before defaults overwrite it
+    user_endpoint = dynamo_config.endpoint
+
     # TODO: move to "disaggregation_mode" as the other engines.
     if dynamo_config.route_to_encoder:
         dynamo_config.component = "processor"
@@ -178,6 +182,13 @@ def update_dynamo_config_with_engine(
         dynamo_config.component = "backend"
         dynamo_config.endpoint = "generate"
 
+    # If user provided --endpoint, override namespace/component/endpoint
+    if user_endpoint is not None:
+        parsed_ns, parsed_comp, parsed_ep = parse_endpoint(user_endpoint)
+        dynamo_config.namespace = parsed_ns
+        dynamo_config.component = parsed_comp
+        dynamo_config.endpoint = parsed_ep
+
     if dynamo_config.custom_jinja_template is not None:
         expanded_template_path = os.path.expanduser(
             os.path.expandvars(dynamo_config.custom_jinja_template)

components/src/dynamo/vllm/tests/test_vllm_unit.py

@@ -140,3 +140,56 @@ def test_model_express_url_none_for_default_load_format(mock_vllm_cli):
     mock_vllm_cli("--model", "Qwen/Qwen3-0.6B")
     config = parse_args()
     assert config.model_express_url is None
+
+
+# --endpoint flag tests
+
+
+def test_endpoint_overrides_defaults(mock_vllm_cli):
+    """Test that --endpoint overrides default namespace/component/endpoint."""
+    mock_vllm_cli(
+        "--model",
+        "Qwen/Qwen3-0.6B",
+        "--endpoint",
+        "dyn://mynamespace.mycomponent.myendpoint",
+    )
+    config = parse_args()
+    assert config.namespace == "mynamespace"
+    assert config.component == "mycomponent"
+    assert config.endpoint == "myendpoint"
+
+
+def test_endpoint_not_provided_preserves_defaults(mock_vllm_cli):
+    """Test that without --endpoint, defaults are preserved."""
+    mock_vllm_cli("--model", "Qwen/Qwen3-0.6B")
+    config = parse_args()
+    assert config.namespace == "dynamo"
+    assert config.component == "backend"
+    assert config.endpoint == "generate"
+
+
+def test_endpoint_overrides_with_prefill_worker(mock_vllm_cli):
+    """Test that --endpoint overrides even with --is-prefill-worker."""
+    mock_vllm_cli(
+        "--model",
+        "Qwen/Qwen3-0.6B",
+        "--endpoint",
+        "dyn://custom.worker.serve",
+        "--is-prefill-worker",
+    )
+    config = parse_args()
+    assert config.namespace == "custom"
+    assert config.component == "worker"
+    assert config.endpoint == "serve"
+
+
+def test_endpoint_invalid_format_raises(mock_vllm_cli):
+    """Test that invalid --endpoint format raises ValueError."""
+    mock_vllm_cli(
+        "--model",
+        "Qwen/Qwen3-0.6B",
+        "--endpoint",
+        "invalid-endpoint",
+    )
+    with pytest.raises(ValueError, match="Invalid endpoint format"):
+        parse_args()