fix: patch o1 and o3-mini (#1657)

letta/llm_api/openai.py

@@ -135,7 +135,7 @@ def build_openai_chat_completions_request(
             tool_choice=tool_choice,
             user=str(user_id),
             max_completion_tokens=llm_config.max_tokens,
-            temperature=llm_config.temperature,
+            temperature=1.0 if llm_config.enable_reasoner else llm_config.temperature,
         )
     else:
         data = ChatCompletionRequest(
@@ -145,7 +145,7 @@ def build_openai_chat_completions_request(
             function_call=function_call,
             user=str(user_id),
             max_completion_tokens=llm_config.max_tokens,
-            temperature=llm_config.temperature,
+            temperature=1.0 if llm_config.enable_reasoner else llm_config.temperature,
         )
         # https://platform.openai.com/docs/guides/text-generation/json-mode
         # only supported by gpt-4o, gpt-4-turbo, or gpt-3.5-turbo
@@ -168,7 +168,6 @@ def build_openai_chat_completions_request(
                 tool.function = FunctionSchema(**structured_output_version)
             except ValueError as e:
                 warnings.warn(f"Failed to convert tool function to structured output, tool={tool}, error={e}")
-
     return data
 
 
@@ -488,4 +487,6 @@ def prepare_openai_payload(chat_completion_request: ChatCompletionRequest):
     #         except ValueError as e:
     #             warnings.warn(f"Failed to convert tool function to structured output, tool={tool}, error={e}")
 
+    if "o3-mini" in chat_completion_request.model or "o1" in chat_completion_request.model:
+        data.pop("parallel_tool_calls", None)
     return data

letta/schemas/llm_config.py

@@ -74,6 +74,13 @@ class LLMConfig(BaseModel):
     # FIXME hack to silence pydantic protected namespace warning
     model_config = ConfigDict(protected_namespaces=())
 
+    @model_validator(mode="before")
+    @classmethod
+    def set_default_enable_reasoner(cls, values):
+        if any(openai_reasoner_model in values.get("model", "") for openai_reasoner_model in ["o3-mini", "o1"]):
+            values["enable_reasoner"] = True
+        return values
+
     @model_validator(mode="before")
     @classmethod
     def set_default_put_inner_thoughts(cls, values):
@@ -100,6 +107,9 @@ class LLMConfig(BaseModel):
                 logger.warning("max_tokens must be greater than max_reasoning_tokens (thinking budget)")
             if self.put_inner_thoughts_in_kwargs:
                 logger.warning("Extended thinking is not compatible with put_inner_thoughts_in_kwargs")
+        elif self.max_reasoning_tokens and not self.enable_reasoner:
+            logger.warning("model will not use reasoning unless enable_reasoner is set to True")
+
         return self
 
     @classmethod

letta/schemas/openai/chat_completion_request.py

@@ -133,6 +133,7 @@ class ChatCompletionRequest(BaseModel):
     temperature: Optional[float] = 1
     top_p: Optional[float] = 1
     user: Optional[str] = None  # unique ID of the end-user (for monitoring)
+    parallel_tool_calls: Optional[bool] = False
 
     # function-calling related
     tools: Optional[List[Tool]] = None