| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import json |
| 5 | |
| 6 | import openai # use the official client for correctness check |
| 7 | import pytest |
| 8 | |
| 9 | MODEL_NAME = "Qwen/Qwen3-1.7B" |
| 10 | tools = [ |
| 11 | { |
| 12 | "type": "function", |
| 13 | "name": "get_current_weather", |
| 14 | "description": "Get the current weather in a given location", |
| 15 | "parameters": { |
| 16 | "type": "object", |
| 17 | "properties": { |
| 18 | "city": { |
| 19 | "type": "string", |
| 20 | "description": "The city to find the weather for, e.g. 'Vienna'", |
| 21 | "default": "Vienna", |
| 22 | }, |
| 23 | "country": { |
| 24 | "type": "string", |
| 25 | "description": "The country that the city is in, e.g. 'Austria'", |
| 26 | }, |
| 27 | "unit": { |
| 28 | "type": "string", |
| 29 | "description": "The unit to fetch the temperature in", |
| 30 | "enum": ["celsius", "fahrenheit"], |
| 31 | }, |
| 32 | "options": { |
| 33 | "$ref": "#/$defs/WeatherOptions", |
| 34 | "description": "Optional parameters for weather query", |
| 35 | }, |
| 36 | }, |
| 37 | "required": ["country", "unit"], |
| 38 | "$defs": { |
| 39 | "WeatherOptions": { |
| 40 | "title": "WeatherOptions", |
| 41 | "type": "object", |
| 42 | "additionalProperties": False, |
| 43 | "properties": { |
| 44 | "unit": { |
| 45 | "type": "string", |
| 46 | "enum": ["celsius", "fahrenheit"], |
| 47 | "default": "celsius", |
| 48 | "description": "Temperature unit", |
| 49 | "title": "Temperature Unit", |
| 50 | }, |
| 51 | "include_forecast": { |
| 52 | "type": "boolean", |
| 53 | "default": False, |
| 54 | "description": "Whether to include a 24-hour forecast", |
| 55 | "title": "Include Forecast", |
| 56 | }, |
| 57 | "language": { |
| 58 | "type": "string", |
| 59 | "default": "zh-CN", |
| 60 | "description": "Language of the response", |
| 61 | "title": "Language", |
| 62 | "enum": ["zh-CN", "en-US", "ja-JP"], |
| 63 | }, |
| 64 | }, |
| 65 | }, |
| 66 | }, |
| 67 | }, |
| 68 | }, |
| 69 | { |
| 70 | "type": "function", |
| 71 | "name": "get_forecast", |
| 72 | "description": "Get the weather forecast for a given location", |
| 73 | "parameters": { |
| 74 | "type": "object", |
| 75 | "properties": { |
| 76 | "city": { |
| 77 | "type": "string", |
| 78 | "description": "The city to get the forecast for, e.g. 'Vienna'", |
| 79 | "default": "Vienna", |
| 80 | }, |
| 81 | "country": { |
| 82 | "type": "string", |
| 83 | "description": "The country that the city is in, e.g. 'Austria'", |
| 84 | }, |
| 85 | "days": { |
| 86 | "type": "integer", |
| 87 | "description": "Number of days to get the forecast for (1-7)", |
| 88 | }, |
| 89 | "unit": { |
| 90 | "type": "string", |
| 91 | "description": "The unit to fetch the temperature in", |
| 92 | "enum": ["celsius", "fahrenheit"], |
| 93 | }, |
| 94 | }, |
| 95 | "required": ["country", "days", "unit"], |
| 96 | }, |
| 97 | }, |
| 98 | ] |
| 99 | |
| 100 | |
| 101 | @pytest.mark.asyncio |
| 102 | @pytest.mark.parametrize("model_name", [MODEL_NAME]) |
| 103 | @pytest.mark.parametrize("tool_choice", ["auto", "required"]) |
| 104 | async def test_function_tool_use( |
| 105 | client: openai.AsyncOpenAI, model_name: str, tool_choice: str |
| 106 | ): |
| 107 | prompt = [ |
| 108 | { |
| 109 | "role": "user", |
| 110 | "content": "Can you tell me what the current weather is in Berlin and the " |
| 111 | "forecast for the next 5 days, in fahrenheit?", |
| 112 | }, |
| 113 | ] |
| 114 | response = await client.responses.create( |
| 115 | model=model_name, |
| 116 | input=prompt, |
| 117 | tools=tools, |
| 118 | tool_choice=tool_choice, |
| 119 | temperature=0.0, |
| 120 | ) |
| 121 | assert len(response.output) >= 1 |
| 122 | tool_call = None |
| 123 | reasoning = None |
| 124 | for out in response.output: |
| 125 | if out.type == "function_call": |
| 126 | tool_call = out |
| 127 | if out.type == "reasoning": |
| 128 | reasoning = out |
| 129 | if response.incomplete_details is None: |
| 130 | assert tool_call is not None |
| 131 | assert tool_call.type == "function_call" |
| 132 | assert json.loads(tool_call.arguments) is not None |
| 133 | assert reasoning is not None |
| 134 | assert reasoning.type == "reasoning" |
| 135 | else: |
| 136 | print(response.model_dump_json(indent=2)) |
| 137 | assert response.incomplete_details.reason == "max_output_tokens" |
| 138 | |
| 139 | |
| 140 | @pytest.mark.asyncio |
| 141 | @pytest.mark.parametrize("model_name", [MODEL_NAME]) |
| 142 | async def test_max_tokens_with_tool_choice_required( |
| 143 | client: openai.AsyncOpenAI, model_name: str |
| 144 | ): |
| 145 | prompt = [ |
| 146 | { |
| 147 | "role": "user", |
| 148 | "content": "Can you tell me what the current weather is in Berlin and the " |
| 149 | "forecast for the next 5 days, in fahrenheit?", |
| 150 | }, |
| 151 | ] |
| 152 | response = await client.responses.create( |
| 153 | model=model_name, |
| 154 | input=prompt, |
| 155 | tools=tools, |
| 156 | tool_choice="required", |
| 157 | max_output_tokens=10, |
| 158 | ) |
| 159 | assert len(response.output) >= 1 |
| 160 | for out in response.output: |
| 161 | # When `tool_choice="required"` and the tokens of `tools` |
| 162 | # exceed `max_output_tokens`,`function_call` should be empty. |
| 163 | # This behavior should be consistent with OpenAI |
| 164 | assert out.type != "function_call" |
| 165 | assert response.incomplete_details.reason == "max_output_tokens" |
| 166 | |
| 167 | |
| 168 | @pytest.mark.asyncio |
| 169 | async def test_named_tool_use(client: openai.AsyncOpenAI): |
| 170 | def get_weather(latitude: float, longitude: float) -> str: |
| 171 | """Mock function to simulate getting weather data. |
| 172 | In a real application, this would call an external weather API. |
| 173 | """ |
| 174 | return f"Current temperature at ({latitude}, {longitude}) is 20°C." |
| 175 | |
| 176 | tools = [ |
| 177 | { |
| 178 | "type": "function", |
| 179 | "name": "get_weather", |
| 180 | "description": ( |
| 181 | "Get current temperature for provided coordinates in celsius." |
| 182 | ), |
| 183 | "parameters": { |
| 184 | "type": "object", |
| 185 | "properties": { |
| 186 | "latitude": {"type": "number"}, |
| 187 | "longitude": {"type": "number"}, |
| 188 | }, |
| 189 | "required": ["latitude", "longitude"], |
| 190 | "additionalProperties": False, |
| 191 | }, |
| 192 | "strict": True, |
| 193 | } |
| 194 | ] |
| 195 | |
| 196 | input_messages = [ |
| 197 | {"role": "user", "content": "What's the weather like in Paris today?"} |
| 198 | ] |
| 199 | |
| 200 | response = await client.responses.create( |
| 201 | model=MODEL_NAME, |
| 202 | input=input_messages, |
| 203 | tools=tools, |
| 204 | tool_choice={"type": "function", "name": "get_weather"}, |
| 205 | ) |
| 206 | assert len(response.output) >= 1 |
| 207 | for out in response.output: |
| 208 | if out.type == "function_call": |
| 209 | tool_call = out |
| 210 | assert tool_call is not None |
| 211 | assert tool_call.type == "function_call" |
| 212 | assert tool_call.name == "get_weather" |
| 213 | args = json.loads(tool_call.arguments) |
| 214 | assert args["latitude"] is not None |
| 215 | assert args["longitude"] is not None |
| 216 | # call the tool |
| 217 | result = get_weather(args["latitude"], args["longitude"]) |
| 218 | input_messages.append(tool_call) # append model's function call message |
| 219 | input_messages.append( |
| 220 | { # append result message |
| 221 | "type": "function_call_output", |
| 222 | "call_id": tool_call.call_id, |
| 223 | "output": str(result), |
| 224 | } |
| 225 | ) |
| 226 | # create a new response with the tool call result |
| 227 | response_2 = await client.responses.create(model=MODEL_NAME, input=input_messages) |
| 228 | # check the output |
| 229 | assert len(response_2.output_text) > 0 |
| 230 | |
| 231 | |
| 232 | @pytest.mark.asyncio |
| 233 | @pytest.mark.parametrize("model_name", [MODEL_NAME]) |
| 234 | async def test_function_calling_with_streaming_expected_arguments( |
| 235 | client: openai.AsyncOpenAI, model_name: str |
| 236 | ): |
| 237 | tools = [ |
| 238 | { |
| 239 | "type": "function", |
| 240 | "name": "get_weather", |
| 241 | "description": "Get current temperature for provided location in celsius.", |
| 242 | "parameters": { |
| 243 | "type": "object", |
| 244 | "properties": { |
| 245 | "location": {"type": "string"}, |
| 246 | }, |
| 247 | "required": ["location"], |
| 248 | "additionalProperties": False, |
| 249 | }, |
| 250 | "strict": True, |
| 251 | }, |
| 252 | { |
| 253 | "type": "function", |
| 254 | "name": "get_time", |
| 255 | "description": "Get current local time for provided location.", |
| 256 | "parameters": { |
| 257 | "type": "object", |
| 258 | "properties": { |
| 259 | "location": {"type": "string"}, |
| 260 | }, |
| 261 | "required": ["location"], |
| 262 | "additionalProperties": False, |
| 263 | }, |
| 264 | "strict": True, |
| 265 | }, |
| 266 | ] |
| 267 | |
| 268 | stream_response = await client.responses.create( |
| 269 | model=model_name, |
| 270 | input=( |
| 271 | "Use tools only. Call get_weather for Berlin and get_time for Tokyo. " |
| 272 | "Do not answer directly." |
| 273 | ), |
| 274 | tools=tools, |
| 275 | stream=True, |
| 276 | ) |
| 277 | |
| 278 | tool_call_items = {} |
| 279 | arguments_done_events = {} |
| 280 | completed_events = {} |
| 281 | async for event in stream_response: |
| 282 | if ( |
| 283 | event.type == "response.output_item.added" |
| 284 | and event.item.type == "function_call" |
| 285 | ): |
| 286 | tool_call_items[event.output_index] = event.item |
| 287 | elif event.type == "response.function_call_arguments.delta": |
| 288 | tool_call_item = tool_call_items[event.output_index] |
| 289 | tool_call_item.arguments += event.delta |
| 290 | elif event.type == "response.function_call_arguments.done": |
| 291 | arguments_done_events[event.output_index] = event |
| 292 | elif ( |
| 293 | event.type == "response.output_item.done" |
| 294 | and event.item.type == "function_call" |
| 295 | ): |
| 296 | completed_events[event.output_index] = event |
| 297 | assert len(tool_call_items) >= 2 |
| 298 | assert len(arguments_done_events) >= 2 |
| 299 | assert len(completed_events) >= 2 |
| 300 | |
| 301 | tool_calls_by_name = { |
| 302 | event.item.name: ( |
| 303 | tool_call_items[output_index], |
| 304 | arguments_done_events[output_index], |
| 305 | event.item, |
| 306 | ) |
| 307 | for output_index, event in completed_events.items() |
| 308 | } |
| 309 | assert {"get_weather", "get_time"}.issubset(tool_calls_by_name) |
| 310 | for added_item, arguments_done_event, completed_item in tool_calls_by_name.values(): |
| 311 | assert added_item.type == "function_call" |
| 312 | assert added_item.arguments == arguments_done_event.arguments |
| 313 | assert added_item.arguments == completed_item.arguments |
| 314 | assert added_item.name == arguments_done_event.name |
| 315 | assert added_item.name == completed_item.name |
| 316 | args = json.loads(added_item.arguments) |
| 317 | assert "location" in args |
| 318 | assert args["location"] is not None |
| 319 | |
| 320 | |
| 321 | @pytest.mark.asyncio |
| 322 | @pytest.mark.parametrize("model_name", [MODEL_NAME]) |
| 323 | @pytest.mark.parametrize( |
| 324 | "tool_choice", |
| 325 | ["auto", "required", {"type": "function", "name": "get_current_weather"}], |
| 326 | ) |
| 327 | @pytest.mark.parametrize( |
| 328 | "enable_thinking", |
| 329 | [True, False], |
| 330 | ) |
| 331 | async def test_function_calling_with_streaming_types( |
| 332 | client: openai.AsyncOpenAI, model_name: str, tool_choice, enable_thinking: bool |
| 333 | ): |
| 334 | # this links the "done" type with the "start" type |
| 335 | # so every "done" type should have a corresponding "start" type |
| 336 | # and every open block should be closed by the end of the stream |
| 337 | # |
| 338 | # stream of events for a response with function call could look like this: |
| 339 | # option1: reasoning -> content(option) -> function_call |
| 340 | # response.created |
| 341 | # -> response.in_progress |
| 342 | # -> response.output_item.added |
| 343 | # -> response.content_part.added |
| 344 | # -> response.reasoning_text.delta |
| 345 | # .... |
| 346 | # -> response.reasoning_text.delta |
| 347 | # -> response.reasoning_text.done |
| 348 | # -> response.content_part.done |
| 349 | # -> response.output_item.done |
| 350 | # -> response.output_item.added |
| 351 | # -> response.content_part.added |
| 352 | # -> response.output_text.delta |
| 353 | # ... |
| 354 | # -> response.output_text.delta |
| 355 | # -> response.output_text.done |
| 356 | # -> response.content_part.done |
| 357 | # -> response.output_item.done |
| 358 | # -> response.output_item.added |
| 359 | # -> response.function_call_arguments.delta |
| 360 | # ... |
| 361 | # -> response.function_call_arguments.delta |
| 362 | # -> response.function_call_arguments.done |
| 363 | # -> response.output_item.done |
| 364 | # -> response.completed |
| 365 | # |
| 366 | # |
| 367 | # option2: reasoning -> content |
| 368 | # response.created |
| 369 | # -> response.in_progress |
| 370 | # -> response.output_item.added |
| 371 | # -> response.content_part.added |
| 372 | # -> response.reasoning_text.delta |
| 373 | # .... |
| 374 | # -> response.reasoning_text.delta |
| 375 | # -> response.reasoning_text.done |
| 376 | # -> response.content_part.done |
| 377 | # -> response.output_item.done |
| 378 | # -> response.output_item.added |
| 379 | # -> response.content_part.added |
| 380 | # -> response.output_text.delta |
| 381 | # .. |
| 382 | # -> response.output_text.delta |
| 383 | # -> response.output_text.done |
| 384 | # -> response.content_part.done |
| 385 | # -> response.output_item.done |
| 386 | # -> response.completed |
| 387 | # |
| 388 | # option3: content |
| 389 | # |
| 390 | # response.created |
| 391 | # -> response.in_progress |
| 392 | # -> response.output_item.added |
| 393 | # -> response.content_part.added |
| 394 | # -> response.output_text.delta |
| 395 | # ... |
| 396 | # -> response.output_text.delta |
| 397 | # -> response.output_text.done |
| 398 | # -> response.content_part.done |
| 399 | # -> response.output_item.done |
| 400 | # -> response.completed |
| 401 | # |
| 402 | # option4: content -> function_call |
| 403 | # response.created |
| 404 | # -> response.in_progress |
| 405 | # -> response.output_item.added |
| 406 | # -> response.content_part.added |
| 407 | # -> response.output_text.delta |
| 408 | # ... |
| 409 | # -> response.output_text.delta |
| 410 | # -> response.output_text.done |
| 411 | # -> response.content_part.done |
| 412 | # -> response.output_item.done |
| 413 | # -> response.output_item.added |
| 414 | # -> response.function_call_arguments.delta |
| 415 | # ... |
| 416 | # -> response.function_call_arguments.delta |
| 417 | # -> response.function_call_arguments.done |
| 418 | # -> response.output_item.done |
| 419 | # -> response.completed |
| 420 | |
| 421 | pairs_of_event_types = { |
| 422 | "response.completed": "response.created", |
| 423 | "response.output_item.done": "response.output_item.added", |
| 424 | "response.output_text.done": "response.output_text.delta", |
| 425 | "response.content_part.done": "response.content_part.added", |
| 426 | "response.reasoning_text.done": "response.reasoning_text.delta", |
| 427 | "response.function_call_arguments.done": "response.function_call_arguments.delta", # noqa |
| 428 | } |
| 429 | |
| 430 | input_list = [ |
| 431 | { |
| 432 | "role": "user", |
| 433 | "content": "Can you tell me what the current weather is in Berlin?", |
| 434 | } |
| 435 | ] |
| 436 | stream_response = await client.responses.create( |
| 437 | model=model_name, |
| 438 | input=input_list, |
| 439 | tools=tools, |
| 440 | tool_choice=tool_choice, |
| 441 | extra_body={"chat_template_kwargs": {"enable_thinking": enable_thinking}}, |
| 442 | stream=True, |
| 443 | ) |
| 444 | |
| 445 | stack_of_event_types = [] |
| 446 | async for event in stream_response: |
| 447 | if event.type == "response.created": |
| 448 | stack_of_event_types.append(event.type) |
| 449 | elif event.type == "response.completed": |
| 450 | assert stack_of_event_types[-1] == pairs_of_event_types[event.type] |
| 451 | stack_of_event_types.pop() |
| 452 | if event.type.endswith("added"): |
| 453 | stack_of_event_types.append(event.type) |
| 454 | elif event.type.endswith("delta"): |
| 455 | if stack_of_event_types[-1] == event.type: |
| 456 | continue |
| 457 | stack_of_event_types.append(event.type) |
| 458 | elif event.type.endswith("done"): |
| 459 | assert stack_of_event_types[-1] == pairs_of_event_types[event.type] |
| 460 | stack_of_event_types.pop() |
| 461 | assert len(stack_of_event_types) == 0 |
| 462 | |
| 463 | |
| 464 | @pytest.mark.asyncio |
| 465 | @pytest.mark.parametrize("model_name", [MODEL_NAME]) |
| 466 | @pytest.mark.parametrize( |
| 467 | "tool_choice", |
| 468 | ["required", "auto", {"type": "function", "name": "get_weather"}], |
| 469 | ) |
| 470 | async def test_function_calling_with_streaming_forced_tool_choice( |
| 471 | client: openai.AsyncOpenAI, model_name: str, tool_choice: str |
| 472 | ): |
| 473 | tools = [ |
| 474 | { |
| 475 | "type": "function", |
| 476 | "name": "get_weather", |
| 477 | "description": "Get current temperature for provided location in celsius.", |
| 478 | "parameters": { |
| 479 | "type": "object", |
| 480 | "properties": { |
| 481 | "location": {"type": "string"}, |
| 482 | }, |
| 483 | "required": ["location"], |
| 484 | "additionalProperties": False, |
| 485 | }, |
| 486 | "strict": True, |
| 487 | } |
| 488 | ] |
| 489 | |
| 490 | stream_response = await client.responses.create( |
| 491 | model=model_name, |
| 492 | input="Call the get_weather function for Berlin and do not answer directly.", |
| 493 | tools=tools, |
| 494 | tool_choice=tool_choice, |
| 495 | stream=True, |
| 496 | ) |
| 497 | |
| 498 | tool_call_item = None |
| 499 | completed_event = None |
| 500 | text_deltas = [] |
| 501 | async for event in stream_response: |
| 502 | if ( |
| 503 | event.type == "response.output_item.added" |
| 504 | and event.item.type == "function_call" |
| 505 | ): |
| 506 | tool_call_item = event.item |
| 507 | elif event.type == "response.output_text.delta": |
| 508 | text_deltas.append(event.delta) |
| 509 | elif event.type == "response.function_call_arguments.delta" and tool_call_item: |
| 510 | tool_call_item.arguments += event.delta |
| 511 | elif ( |
| 512 | event.type == "response.output_item.done" |
| 513 | and event.item.type == "function_call" |
| 514 | ): |
| 515 | completed_event = event |
| 516 | |
| 517 | assert tool_call_item is not None |
| 518 | assert tool_call_item.type == "function_call" |
| 519 | assert tool_call_item.name == "get_weather" |
| 520 | assert completed_event is not None |
| 521 | assert tool_call_item.arguments == completed_event.item.arguments |
| 522 | assert tool_call_item.name == completed_event.item.name |
| 523 | args = json.loads(tool_call_item.arguments) |
| 524 | assert "location" in args |
| 525 | assert args["location"] is not None |
| 526 | # Forced tool choice should not leak tool-call JSON via output_text delta. |
| 527 | assert "".join(text_deltas).strip() == "" |