[Tool Parser] Normalize nested arguments in DeepSeek DSML tool calls

vllm/tool_parsers/deepseekv32_tool_parser.py

@@ -58,15 +58,19 @@ def __init__(self, tokenizer: TokenizerLike, tools: list[Tool] | None = None):
         self.current_tool_index: int = 0
         self._sent_content_idx: int = 0
 
-        # Regex patterns for complete parsing
+        # Regex patterns for complete parsing.
+        #
+        # The wrapper tokens are class attributes so subclasses such as
+        # DeepSeekV4ToolParser can override them.
         self.tool_call_complete_regex = re.compile(
             re.escape(self.tool_call_start_token)
             + r"(.*?)"
             + re.escape(self.tool_call_end_token),
             re.DOTALL,
         )
         self.invoke_complete_regex = re.compile(
-            r'<｜DSML｜invoke\s+name="([^"]+)"\s*>(.*?)</｜DSML｜invoke>', re.DOTALL
+            r'<｜DSML｜invoke\s+name="([^"]+)"\s*>(.*?)</｜DSML｜invoke>',
+            re.DOTALL,
         )
         self.parameter_complete_regex = re.compile(
             r'<｜DSML｜parameter\s+name="([^"]+)"\s+string="(?:true|false)"\s*>(.*?)</｜DSML｜parameter>',
@@ -90,7 +94,7 @@ def adjust_request(
         if request.tools and request.tool_choice != "none":
             # Ensure tool call tokens
             # (e.g. <｜DSML｜function_calls>, </｜DSML｜function_calls>)
-            # are not skippedduring decoding.
+            # are not skipped during decoding.
             # Even though they are not marked as special tokens,
             # setting skip_special_tokens=False ensures proper handling in
             # transformers 5.x where decoding behavior may have changed.
@@ -101,8 +105,8 @@ def _generate_tool_call_id(self) -> str:
         """Generate a unique tool call ID."""
         return f"call_{uuid.uuid4().hex[:24]}"
 
-    def _parse_invoke_params(self, invoke_str: str) -> dict:
-        param_dict = dict()
+    def _parse_invoke_params(self, invoke_str: str) -> dict[str, str]:
+        param_dict: dict[str, str] = {}
         for param_name, param_val in self.parameter_complete_regex.findall(invoke_str):
             param_dict[param_name] = param_val
         return param_dict
@@ -130,8 +134,11 @@ def _convert_param_value_checked(self, value: str, param_type: str) -> Any:
         else:
             return json.loads(value)
 
-    def _convert_param_value(self, value: str, param_type: str | list[str]) -> Any:
+    def _convert_param_value(self, value: Any, param_type: str | list[str]) -> Any:
         """Convert parameter value to the correct type."""
+        if not isinstance(value, str):
+            return value
+
         if not isinstance(param_type, list):
             param_type = [param_type]
         for current_type in param_type:
@@ -142,6 +149,45 @@ def _convert_param_value(self, value: str, param_type: str | list[str]) -> Any:
         # return value as fallback
         return value
 
+    def _normalize_arguments_wrapper(
+        self,
+        converted: dict[str, Any],
+    ) -> dict[str, Any]:
+        """Normalize model-generated nested arguments wrapper.
+
+        DeepSeek V4 Flash may generate DSML parameters like:
+
+            <｜DSML｜parameter name="arguments" string="false">
+            {"path": "/tmp/a", "content": "hello"}
+            </｜DSML｜parameter>
+
+        The parser would otherwise produce:
+
+            {"arguments": {"path": "/tmp/a", "content": "hello"}}
+
+        OpenAI-compatible function.arguments should be:
+
+            {"path": "/tmp/a", "content": "hello"}
+        """
+        if set(converted.keys()) != {"arguments"}:
+            return converted
+
+        wrapped = converted.get("arguments")
+
+        if isinstance(wrapped, dict):
+            return wrapped
+
+        if isinstance(wrapped, str):
+            try:
+                parsed = json.loads(wrapped)
+            except Exception:
+                return converted
+
+            if isinstance(parsed, dict):
+                return parsed
+
+        return converted
+
     def _convert_params_with_schema(
         self,
         function_name: str,
@@ -167,7 +213,8 @@ def _convert_params_with_schema(
             if name in param_config and isinstance(param_config[name], dict):
                 param_type = param_config[name].get("type", "string")
             converted[name] = self._convert_param_value(value, param_type)
-        return converted
+
+        return self._normalize_arguments_wrapper(converted)
 
     def extract_tool_calls(
         self,
@@ -191,12 +238,16 @@ def extract_tool_calls(
                     tool_call_match
                 ):
                     param_dict = self._parse_invoke_params(invoke_content)
+                    converted = self._convert_params_with_schema(
+                        invoke_name,
+                        param_dict,
+                    )
                     tool_calls.append(
                         ToolCall(
                             type="function",
                             function=FunctionCall(
                                 name=invoke_name,
-                                arguments=json.dumps(param_dict, ensure_ascii=False),
+                                arguments=json.dumps(converted, ensure_ascii=False),
                             ),
                         )
                     )
@@ -268,15 +319,29 @@ def _extract_delta_tool_calls(
 
         return delta_tool_calls
 
-    def _extract_content(self, current_text: str) -> str | None:
+    def _extract_content(
+        self,
+        current_text: str,
+        *,
+        is_final: bool = False,
+    ) -> str | None:
         """Return unsent non-tool-call text, or None.
 
         Holds back any suffix that could be a partial start marker
         so that split markers are never leaked as content.
+
+        On final streaming step, flush the held-back suffix because it
+        cannot form a complete tool-call start marker anymore.
         """
         if self.tool_call_start_token not in current_text:
-            overlap = partial_tag_overlap(current_text, self.tool_call_start_token)
-            sendable_idx = len(current_text) - overlap
+            if is_final:
+                sendable_idx = len(current_text)
+            else:
+                overlap = partial_tag_overlap(
+                    current_text,
+                    self.tool_call_start_token,
+                )
+                sendable_idx = len(current_text) - overlap
         else:
             sendable_idx = current_text.index(self.tool_call_start_token)
 
@@ -307,15 +372,19 @@ def extract_tool_calls_streaming(
         if not previous_text:
             self._reset_streaming_state()
 
-        content = self._extract_content(current_text)
+        # Empty delta with token ids means EOS or a skipped/closing token.
+        # Treat it as final for content flushing purposes.
+        is_final = not delta_text and bool(delta_token_ids)
+
+        content = self._extract_content(current_text, is_final=is_final)
         delta_tool_calls = self._extract_delta_tool_calls(current_text, request)
 
         if delta_tool_calls or content:
             return DeltaMessage(content=content, tool_calls=delta_tool_calls)
 
         # Empty delta with token ids means EOS or closing tag; return
         # non-None so the serving framework can finalize finish_reason.
-        if not delta_text and delta_token_ids and self.prev_tool_call_arr:
+        if is_final and self.prev_tool_call_arr:
             return DeltaMessage(content="")
 
         return None