diff --git a/python/openai/openai_frontend/engine/triton_engine.py b/python/openai/openai_frontend/engine/triton_engine.py
index 0d735be5c0..d194fc2e1f 100644
--- a/python/openai/openai_frontend/engine/triton_engine.py
+++ b/python/openai/openai_frontend/engine/triton_engine.py
@@ -1,4 +1,4 @@
-# Copyright 2024-2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# Copyright 2024-2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 #
 # Redistribution and use in source and binary forms, with or without
 # modification, are permitted provided that the following conditions
@@ -991,6 +991,11 @@ def _validate_completion_request(
         if not isinstance(request.prompt, str):
             raise ClientError("only single string input is supported")
 
+        if "best_of" in request.model_fields_set and metadata.backend == "vllm":
+            raise ClientError(
+                "best_of is no longer supported in vLLM backend, removed from vLLM V1 engine"
+            )
+
         if request.n and request.n > 1:
             raise ClientError(
                 f"Received n={request.n}, but only single choice (n=1) is currently supported"
diff --git a/python/openai/openai_frontend/engine/utils/triton.py b/python/openai/openai_frontend/engine/utils/triton.py
index 344ad374c9..8dd7663f3a 100644
--- a/python/openai/openai_frontend/engine/utils/triton.py
+++ b/python/openai/openai_frontend/engine/utils/triton.py
@@ -1,4 +1,4 @@
-# Copyright 2024-2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# Copyright 2024-2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 #
 # Redistribution and use in source and binary forms, with or without
 # modification, are permitted provided that the following conditions
@@ -100,6 +100,8 @@ def _create_vllm_generate_request(
         "max_tokens",
         "logprobs",
         "top_logprobs",
+        # not supported for vLLM backend (removed from vLLM V1) but supported for TRT-LLM/Python backend
+        "best_of",
     }
 
     # NOTE: The exclude_none is important, as internals may not support
diff --git a/python/openai/requirements.txt b/python/openai/requirements.txt
index cf63446574..4b4f631f9b 100644
--- a/python/openai/requirements.txt
+++ b/python/openai/requirements.txt
@@ -1,4 +1,4 @@
-# Copyright 2024-2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# Copyright 2024-2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 #
 # Redistribution and use in source and binary forms, with or without
 # modification, are permitted provided that the following conditions
@@ -31,6 +31,9 @@ fastapi==0.121.2
 httpx==0.27.2
 openai==1.107.3
 partial-json-parser # used for parsing partial JSON outputs
+
+# FIXME [TRI-641]: The latest stable version of scipy is 1.17.0 which caused segfault during tests. See TRI-620.
+scipy==1.16.3
 # Minimum starlette version needed to address CVE(s):
 # https://github.com/advisories/GHSA-f96h-pmfr-66vw
 # https://github.com/advisories/GHSA-7f5h-v6xp-fcq8
diff --git a/python/openai/tests/conftest.py b/python/openai/tests/conftest.py
index cff670f3ec..a1195000ed 100644
--- a/python/openai/tests/conftest.py
+++ b/python/openai/tests/conftest.py
@@ -1,4 +1,4 @@
-# Copyright 2024-2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# Copyright 2024-2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 #
 # Redistribution and use in source and binary forms, with or without
 # modification, are permitted provided that the following conditions
@@ -37,6 +37,7 @@ def pytest_configure(config):
     config.addinivalue_line(
         "markers", "openai: mark test to run with OpenAI server (subprocess)"
     )
+    config.addinivalue_line("markers", "asyncio: mark test as an asyncio test")
 
 
 ### TEST ENVIRONMENT SETUP ###
diff --git a/python/openai/tests/test_completions.py b/python/openai/tests/test_completions.py
index 4dcad4c2d2..9d496476e1 100644
--- a/python/openai/tests/test_completions.py
+++ b/python/openai/tests/test_completions.py
@@ -1,4 +1,4 @@
-# Copyright 2024-2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# Copyright 2024-2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 #
 # Redistribution and use in source and binary forms, with or without
 # modification, are permitted provided that the following conditions
@@ -57,7 +57,6 @@ def test_completions_defaults(self, client, model: str, prompt: str):
             ("top_p", 0.9),
             ("frequency_penalty", 0.5),
             ("presence_penalty", 0.2),
-            ("best_of", 1),
             ("n", 1),
             # logprobs is an integer for completions
             ("logprobs", 5),
@@ -359,7 +358,12 @@ def test_no_prompt(self, client, model: str):
         ],
     )
     def test_completions_multiple_choices(
-        self, client, sampling_parameter_dict: dict, model: str, prompt: str
+        self,
+        client,
+        sampling_parameter_dict: dict,
+        backend: str,
+        model: str,
+        prompt: str,
     ):
         response = client.post(
             "/v1/completions",
@@ -370,7 +374,11 @@ def test_completions_multiple_choices(
         # FIXME: Add support and test for success
         # Expected to fail when n or best_of > 1, only single choice supported for now
         assert response.status_code == 400
-        assert "only single choice" in response.json()["detail"]
+        if backend == "vllm" and "best_of" in sampling_parameter_dict:
+            error_message = "best_of is no longer supported in vLLM backend"
+        else:
+            error_message = "only single choice"
+        assert error_message in response.json()["detail"]
 
     @pytest.mark.skip(reason="Not Implemented Yet")
     def test_lora(self):