[Bugfix] fix tool_parser error handling when serve a model not support it (#8709)

vllm/entrypoints/openai/serving_chat.py

@@ -302,10 +302,6 @@ class OpenAIServingChat(OpenAIServing):
         finish_reason_sent = [False] * num_choices
         num_prompt_tokens = 0
 
-        tool_parsers: List[Optional[ToolParser]] = [
-            self.tool_parser(tokenizer) if self.tool_parser else None
-        ] * num_choices
-
         if isinstance(request.tool_choice, ChatCompletionNamedToolChoiceParam):
             tool_choice_function_name = request.tool_choice.function.name
         else:
@@ -324,6 +320,21 @@ class OpenAIServingChat(OpenAIServing):
         else:
             previous_texts, all_previous_token_ids = None, None
 
+        # Prepare the tool parser if it's needed
+        try:
+            if tool_choice_auto and self.tool_parser:
+                tool_parsers: List[Optional[ToolParser]] = [
+                    self.tool_parser(tokenizer)
+                ] * num_choices
+            else:
+                tool_parsers = [None] * num_choices
+        except RuntimeError as e:
+            logger.error("Error in tool parser creation: %s", e)
+            data = self.create_streaming_error_response(str(e))
+            yield f"data: {data}\n\n"
+            yield "data: [DONE]\n\n"
+            return
+
         try:
             async for res in result_generator:
                 if res.prompt_token_ids is not None:
@@ -704,7 +715,12 @@ class OpenAIServingChat(OpenAIServing):
                     or request.tool_choice is None) and self.enable_auto_tools \
                     and self.tool_parser:
 
-                tool_parser = self.tool_parser(tokenizer)
+                try:
+                    tool_parser = self.tool_parser(tokenizer)
+                except RuntimeError as e:
+                    logger.error("Error in tool parser creation: %s", e)
+                    return self.create_error_response(str(e))
+
                 tool_call_info = tool_parser.extract_tool_calls(
                     output.text, request=request)
                 tools_called = tool_call_info.tools_called

vllm/entrypoints/openai/tool_parsers/hermes_tool_parser.py

@@ -50,10 +50,10 @@ class Hermes2ProToolParser(ToolParser):
             raise ValueError(
                 "The model tokenizer must be passed to the ToolParser "
                 "constructor during construction.")
-        self.tool_call_start_token_id: int = self.model_tokenizer.vocab[
-            self.tool_call_start_token]
-        self.tool_call_end_token_id: int = self.model_tokenizer.vocab[
-            self.tool_call_end_token]
+        self.tool_call_start_token_id: int = self.model_tokenizer.vocab.get(
+            self.tool_call_start_token, None)
+        self.tool_call_end_token_id: int = self.model_tokenizer.vocab.get(
+            self.tool_call_end_token, None)
         if not self.tool_call_start_token_id or not self.tool_call_end_token_id:
             raise RuntimeError(
                 "Hermes 2 Pro Tool parser could not locate tool call start/end "

vllm/entrypoints/openai/tool_parsers/mistral_tool_parser.py

@@ -61,8 +61,13 @@ class MistralToolParser(ToolParser):
         self.streamed_args_for_tool: List[str] = [
         ]  # map what has been streamed for each tool so far to a list
         self.bot_token = "[TOOL_CALLS]"
-        self.bot_token_id = self.model_tokenizer.get_vocab()[self.bot_token]
+        self.bot_token_id = self.model_tokenizer.get_vocab().get(
+            self.bot_token, None)
         self.tool_call_regex = re.compile(r"\[{.*?}\]", re.DOTALL)
+        if not self.bot_token_id:
+            raise RuntimeError(
+                "Mistral Tool Parser could not locate the tool call token in "
+                "the tokenizer!")
 
     def extract_tool_calls(
         self,