Fix issue #2843: Exclude stop parameter for models that don't support it

src/crewai/llm.py

@@ -92,6 +92,8 @@ def suppress_warnings():
 
 
 class LLM:
+    MODELS_WITHOUT_STOP_SUPPORT = ["o3", "o3-mini", "o4-mini"]
+
     def __init__(
         self,
         model: str,
@@ -155,7 +157,7 @@ def call(self, messages: List[Dict[str, str]], callbacks: List[Any] = []) -> str
                     "temperature": self.temperature,
                     "top_p": self.top_p,
                     "n": self.n,
-                    "stop": self.stop,
+                    "stop": self.stop if self.supports_stop_words() else None,
                     "max_tokens": self.max_tokens or self.max_completion_tokens,
                     "presence_penalty": self.presence_penalty,
                     "frequency_penalty": self.frequency_penalty,
@@ -193,6 +195,19 @@ def supports_function_calling(self) -> bool:
             return False
 
     def supports_stop_words(self) -> bool:
+        """
+        Determines whether the current model supports the 'stop' parameter.
+
+        This method checks if the model is in the list of models known not to support
+        stop words, and if not, it queries the litellm library to determine if the
+        model supports the 'stop' parameter.
+
+        Returns:
+            bool: True if the model supports stop words, False otherwise.
+        """
+        if any(self.model.startswith(model) for model in self.MODELS_WITHOUT_STOP_SUPPORT):
+            return False
+
         try:
             params = get_supported_openai_params(model=self.model)
             return "stop" in params

tests/llm_test.py

@@ -28,3 +28,41 @@ def test_llm_callback_replacement():
     assert usage_metrics_1.successful_requests == 1
     assert usage_metrics_2.successful_requests == 1
     assert usage_metrics_1 == calc_handler_1.token_cost_process.get_summary()
+
+
+class TestLLMStopWords:
+    """Tests for LLM stop words functionality."""
+
+    def test_supports_stop_words_for_o3_model(self):
+        """Test that supports_stop_words returns False for o3 model."""
+        llm = LLM(model="o3")
+        assert not llm.supports_stop_words()
+
+    def test_supports_stop_words_for_o4_mini_model(self):
+        """Test that supports_stop_words returns False for o4-mini model."""
+        llm = LLM(model="o4-mini")
+        assert not llm.supports_stop_words()
+
+    def test_supports_stop_words_for_supported_model(self):
+        """Test that supports_stop_words returns True for models that support stop words."""
+        llm = LLM(model="gpt-4")
+        assert llm.supports_stop_words()
+
+    @pytest.mark.vcr(filter_headers=["authorization"])
+    def test_llm_call_excludes_stop_parameter_for_unsupported_models(self, monkeypatch):
+        """Test that the LLM.call method excludes the stop parameter for models that don't support it."""
+        def mock_completion(**kwargs):
+            assert 'stop' not in kwargs, "Stop parameter should be excluded for o3 model"
+            assert 'model' in kwargs, "Model parameter should be included"
+            assert 'messages' in kwargs, "Messages parameter should be included"
+            return {"choices": [{"message": {"content": "Hello, World!"}}]}
+
+        monkeypatch.setattr("litellm.completion", mock_completion)
+
+        llm = LLM(model="o3")
+        llm.stop = ["STOP"]
+
+        messages = [{"role": "user", "content": "Say 'Hello, World!'"}]
+        response = llm.call(messages)
+
+        assert response == "Hello, World!"