diff --git a/tests/tool_parsers/test_qwen3coder_tool_parser.py b/tests/tool_parsers/test_qwen3coder_tool_parser.py index 7db1b6857..8f92d5b37 100644 --- a/tests/tool_parsers/test_qwen3coder_tool_parser.py +++ b/tests/tool_parsers/test_qwen3coder_tool_parser.py @@ -429,6 +429,208 @@ hello world assert args["obj_param"] == {"key": "value"} +def test_extract_tool_calls_anyof_type_conversion(qwen3_tool_parser): + """Test type conversion for anyOf/oneOf nullable schemas (Pydantic v2). + + Pydantic v2 emits anyOf for Optional[T] fields, e.g.: + Optional[int] -> {"anyOf": [{"type": "integer"}, {"type": "null"}]} + The parser must extract the non-null type and apply the correct + conversion (int(), float(), etc.) instead of returning a raw string. + """ + tools = [ + ChatCompletionToolsParam( + type="function", + function={ + "name": "test_anyof", + "parameters": { + "type": "object", + "properties": { + "anyof_int": { + "anyOf": [ + {"type": "integer"}, + {"type": "null"}, + ], + "default": 5, + }, + "anyof_str": { + "anyOf": [ + {"type": "string"}, + {"type": "null"}, + ], + }, + "anyof_array": { + "anyOf": [ + {"type": "array", "items": {"type": "string"}}, + {"type": "null"}, + ], + }, + "anyof_obj": { + "anyOf": [ + {"type": "object"}, + {"type": "null"}, + ], + }, + "type_as_array": { + "type": ["integer", "null"], + }, + "multi_non_null": { + "anyOf": [ + {"type": "string"}, + {"type": "integer"}, + {"type": "null"}, + ], + }, + "ref_param": { + "$ref": "#/$defs/ToolInput", + }, + }, + }, + }, + ) + ] + + model_output = """ + + +5 + + +hello + + +["a", "b", "c"] + + +{"key": "value"} + + +42 + + +some text + + +{"city": "Paris"} + + +""" + + request = ChatCompletionRequest(model=MODEL, messages=[], tools=tools) + extracted = qwen3_tool_parser.extract_tool_calls(model_output, request=request) + + args = json.loads(extracted.tool_calls[0].function.arguments) + assert args["anyof_int"] == 5 + assert isinstance(args["anyof_int"], int) + assert args["anyof_str"] == "hello" + assert isinstance(args["anyof_str"], str) + assert args["anyof_array"] == ["a", "b", "c"] + assert isinstance(args["anyof_array"], list) + assert args["anyof_obj"] == {"key": "value"} + assert isinstance(args["anyof_obj"], dict) + assert args["type_as_array"] == 42 + assert isinstance(args["type_as_array"], int) + # Multi non-null: anyOf[string, integer, null] → first non-null is string + assert args["multi_non_null"] == "some text" + assert isinstance(args["multi_non_null"], str) + # $ref: treated as object, parsed via json.loads + assert args["ref_param"] == {"city": "Paris"} + assert isinstance(args["ref_param"], dict) + + +def test_extract_tool_calls_anyof_type_conversion_streaming( + qwen3_tool_parser, qwen3_tokenizer +): + """Test streaming e2e for anyOf/oneOf nullable schemas (Pydantic v2). + + Verifies that the full streaming pipeline — tokenize, incrementally + decode, extract_tool_calls_streaming — correctly resolves types from + anyOf schemas and produces valid JSON with properly typed values. + """ + tools = [ + ChatCompletionToolsParam( + type="function", + function={ + "name": "search_web", + "parameters": { + "type": "object", + "properties": { + "query": { + "anyOf": [ + {"type": "string"}, + {"type": "null"}, + ], + }, + "count": { + "anyOf": [ + {"type": "integer"}, + {"type": "null"}, + ], + "default": 5, + }, + "verbose": { + "anyOf": [ + {"type": "boolean"}, + {"type": "null"}, + ], + }, + "filters": { + "$ref": "#/$defs/SearchFilters", + }, + }, + }, + }, + ) + ] + + model_output = """ + + +vllm tool parser + + +10 + + +true + + +{"lang": "en", "year": 2025} + + +""" + + request = ChatCompletionRequest(model=MODEL, messages=[], tools=tools) + + tool_states = {} + for delta_message in stream_delta_message_generator( + qwen3_tool_parser, qwen3_tokenizer, model_output, request + ): + if delta_message.tool_calls: + for tool_call in delta_message.tool_calls: + idx = tool_call.index + if idx not in tool_states: + tool_states[idx] = {"name": None, "arguments": ""} + if tool_call.function: + if tool_call.function.name: + tool_states[idx]["name"] = tool_call.function.name + if tool_call.function.arguments is not None: + tool_states[idx]["arguments"] += tool_call.function.arguments + + assert len(tool_states) == 1 + assert tool_states[0]["name"] == "search_web" + assert tool_states[0]["arguments"] is not None + args = json.loads(tool_states[0]["arguments"]) + assert args["query"] == "vllm tool parser" + assert isinstance(args["query"], str) + assert args["count"] == 10 + assert isinstance(args["count"], int) + assert args["verbose"] is True + assert isinstance(args["verbose"], bool) + # $ref: treated as object, parsed via json.loads + assert args["filters"] == {"lang": "en", "year": 2025} + assert isinstance(args["filters"], dict) + + @pytest.mark.parametrize( ids=[ "no_tools", diff --git a/vllm/tool_parsers/qwen3coder_tool_parser.py b/vllm/tool_parsers/qwen3coder_tool_parser.py index ea25ea2be..8a5b4f5a2 100644 --- a/vllm/tool_parsers/qwen3coder_tool_parser.py +++ b/vllm/tool_parsers/qwen3coder_tool_parser.py @@ -131,11 +131,58 @@ class Qwen3CoderToolParser(ToolParser): logger.debug("Tool '%s' is not defined in the tools list.", func_name) return {} + @staticmethod + def _first_non_null_type(type_value: Any) -> str | None: + """Extract the first non-null type from a type value. + + Handles both scalar types ("integer") and type-as-array + (["integer", "null"]) per JSON Schema spec. + """ + if isinstance(type_value, list): + return next( + ( + str(t).strip().lower() + for t in type_value + if t is not None and str(t).lower() != "null" + ), + None, + ) + if type_value is not None and str(type_value).lower() != "null": + return str(type_value).strip().lower() + return None + + def _resolve_param_type(self, param_def: dict) -> str: + """Resolve the effective type string from a parameter definition. + + Handles direct "type" fields (including type-as-array), + anyOf/oneOf schemas emitted by Pydantic v2 for Optional[T], + and $ref schemas from Pydantic model inputs. + """ + if "type" in param_def: + resolved = self._first_non_null_type(param_def["type"]) + return resolved or "string" + + if "anyOf" in param_def or "oneOf" in param_def: + variants = param_def.get("anyOf") or param_def.get("oneOf", []) + for v in variants: + if not isinstance(v, dict): + continue + resolved = self._first_non_null_type(v.get("type")) + if resolved: + return resolved + + # $ref points to a schema definition (e.g. a Pydantic model). + # The referenced type is almost always an object, so treat it + # as such to route through json.loads. + if "$ref" in param_def: + return "object" + + return "string" + def _convert_param_value( self, param_value: str, param_name: str, param_config: dict, func_name: str ) -> Any: """Convert parameter value based on its type in the schema.""" - # Handle null value for any type if param_value.lower() == "null": return None @@ -150,19 +197,10 @@ class Qwen3CoderToolParser(ToolParser): ) return param_value - if ( - isinstance(param_config[param_name], dict) - and "type" in param_config[param_name] - ): - param_type = str(param_config[param_name]["type"]).strip().lower() - elif ( - isinstance(param_config[param_name], dict) - and "anyOf" in param_config[param_name] - ): - # anyOf has no top-level "type"; treat as object to trigger json.loads. - param_type = "object" - else: - param_type = "string" + if not isinstance(param_config[param_name], dict): + return param_value + + param_type = self._resolve_param_type(param_config[param_name]) if param_type in ["string", "str", "text", "varchar", "char", "enum"]: return param_value elif (