diff --git a/flexeval/core/language_model/vllm_serve_lm.py b/flexeval/core/language_model/vllm_serve_lm.py
index 0e2562be..524bd914 100644
--- a/flexeval/core/language_model/vllm_serve_lm.py
+++ b/flexeval/core/language_model/vllm_serve_lm.py
@@ -193,6 +193,7 @@ def __init__(
         if api_headers is None:
             api_headers = {}
         api_headers["base_url"] = self.manager.base_url
+        api_headers["api_key"] = "EMPTY"  # OpenAI client requires an api_key, but vLLM does not use it.
         super().__init__(
             model=model,
             api_headers=api_headers,
diff --git a/tests/core/language_model/vllm/test_vllm_serve_lm.py b/tests/core/language_model/vllm/test_vllm_serve_lm.py
index 96667d6e..72a3700b 100644
--- a/tests/core/language_model/vllm/test_vllm_serve_lm.py
+++ b/tests/core/language_model/vllm/test_vllm_serve_lm.py
@@ -1,5 +1,6 @@
 from __future__ import annotations
 
+import os
 from unittest import mock
 
 import pytest
@@ -24,6 +25,7 @@ def readline(self) -> str:
 
 @pytest.fixture(scope="module")
 def chat_lm() -> VLLMServeLM:
+    openai_api_key = os.environ.pop("OPENAI_API_KEY", None)
     llm = VLLMServeLM(
         model="sbintuitions/tiny-lm-chat",
         model_kwargs={
@@ -36,6 +38,8 @@ def chat_lm() -> VLLMServeLM:
     )
     yield llm
     llm.manager.stop()
+    if openai_api_key is not None:
+        os.environ["OPENAI_API_KEY"] = openai_api_key
 
 
 def test_port_is_auto_assigned() -> None: