fix: address review feedback for multi-model-name support

- Deduplicate aliases fully (not just primary) using ordered set logic
- Document gRPC single-model limitation in docstring

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: nvyutwu

tensorrt_llm/commands/serve.py

@@ -335,7 +335,9 @@ def launch_grpc_server(host: str,
         host: Host to bind to
         port: Port to bind to
         llm_args: Arguments for LLM initialization (from get_llm_args)
-        served_model_name: Custom model name for API responses (defaults to model path)
+        served_model_name: Model name(s) for API responses (defaults to model path).
+            Note: the gRPC server only uses the first (primary) name. Multiple
+            aliases are supported by the HTTP/OpenAI server only.
     """
     import grpc
 

tensorrt_llm/serve/openai_server.py

@@ -205,9 +205,12 @@ def __init__(
         if model_dir.exists() and model_dir.is_dir():
             primary = model_dir.name
         self.model = primary
-        self.served_model_names: List[str] = [primary] + [
-            n for n in names[1:] if n != primary
-        ]
+        seen = {primary}
+        self.served_model_names: List[str] = [primary]
+        for n in names[1:]:
+            if n not in seen:
+                seen.add(n)
+                self.served_model_names.append(n)
         self.metrics_collector = None
         self.perf_metrics = None
         self.perf_metrics_lock = None