| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import json |
| 5 | from typing import Any |
| 6 | from unittest.mock import MagicMock |
| 7 | |
| 8 | import pytest |
| 9 | |
| 10 | from vllm.entrypoints.openai.chat_completion.protocol import ChatCompletionRequest |
| 11 | from vllm.parser.gemma4 import ( |
| 12 | TOOL_CALL_END, |
| 13 | TOOL_CALL_START, |
| 14 | _parse_gemma4_args, |
| 15 | _parse_gemma4_array, |
| 16 | ) |
| 17 | from vllm.tool_parsers.gemma4_engine_tool_parser import Gemma4EngineToolParser |
| 18 | |
| 19 | # --------------------------------------------------------------------------- |
| 20 | # Fixtures |
| 21 | # --------------------------------------------------------------------------- |
| 22 | |
| 23 | |
| 24 | TOOL_CALL_START_ID = 48 |
| 25 | TOOL_CALL_END_ID = 49 |
| 26 | CHANNEL_START = "<|channel>" |
| 27 | CHANNEL_END = "<channel|>" |
| 28 | CHANNEL_START_ID = 50 |
| 29 | CHANNEL_END_ID = 51 |
| 30 | |
| 31 | |
| 32 | def _make_tool(name, properties): |
| 33 | from vllm.entrypoints.openai.chat_completion.protocol import ( |
| 34 | ChatCompletionToolsParam, |
| 35 | ) |
| 36 | |
| 37 | return ChatCompletionToolsParam( |
| 38 | type="function", |
| 39 | function={ |
| 40 | "name": name, |
| 41 | "parameters": {"type": "object", "properties": properties}, |
| 42 | }, |
| 43 | ) |
| 44 | |
| 45 | |
| 46 | _TOOLS = [ |
| 47 | _make_tool( |
| 48 | "set_status", |
| 49 | { |
| 50 | "is_active": {"type": "boolean"}, |
| 51 | "count": {"type": "integer"}, |
| 52 | "score": {"type": "number"}, |
| 53 | }, |
| 54 | ), |
| 55 | _make_tool( |
| 56 | "set_config", |
| 57 | { |
| 58 | "count": {"type": "integer"}, |
| 59 | "active": {"type": "boolean"}, |
| 60 | }, |
| 61 | ), |
| 62 | _make_tool( |
| 63 | "search", |
| 64 | { |
| 65 | "input": { |
| 66 | "type": "object", |
| 67 | "properties": {"all": {"type": "boolean"}}, |
| 68 | }, |
| 69 | }, |
| 70 | ), |
| 71 | _make_tool( |
| 72 | "set", |
| 73 | { |
| 74 | "flag": {"type": "boolean"}, |
| 75 | "count": {"type": "integer"}, |
| 76 | }, |
| 77 | ), |
| 78 | _make_tool( |
| 79 | "Edit", |
| 80 | { |
| 81 | "file_path": {"type": "string"}, |
| 82 | "old_string": {"type": "string"}, |
| 83 | "new_string": {"type": "string"}, |
| 84 | "replace_all": {"type": "boolean"}, |
| 85 | }, |
| 86 | ), |
| 87 | ] |
| 88 | |
| 89 | |
| 90 | @pytest.fixture |
| 91 | def mock_tokenizer(): |
| 92 | vocab = { |
| 93 | TOOL_CALL_START: TOOL_CALL_START_ID, |
| 94 | TOOL_CALL_END: TOOL_CALL_END_ID, |
| 95 | CHANNEL_START: CHANNEL_START_ID, |
| 96 | CHANNEL_END: CHANNEL_END_ID, |
| 97 | } |
| 98 | decode_map = {v: k for k, v in vocab.items()} |
| 99 | |
| 100 | tokenizer = MagicMock() |
| 101 | tokenizer.encode.return_value = [1, 2, 3] |
| 102 | tokenizer.get_vocab.return_value = vocab |
| 103 | tokenizer.decode.side_effect = lambda ids: decode_map.get(ids[0], f"tok{ids[0]}") |
| 104 | return tokenizer |
| 105 | |
| 106 | |
| 107 | @pytest.fixture |
| 108 | def parser(mock_tokenizer): |
| 109 | return Gemma4EngineToolParser(mock_tokenizer, tools=_TOOLS) |
| 110 | |
| 111 | |
| 112 | @pytest.fixture |
| 113 | def mock_request(): |
| 114 | request = MagicMock(spec=ChatCompletionRequest) |
| 115 | request.tools = [] |
| 116 | request.tool_choice = "auto" |
| 117 | return request |
| 118 | |
| 119 | |
| 120 | # --------------------------------------------------------------------------- |
| 121 | # Unit tests for _parse_gemma4_args (shared parser logic) |
| 122 | # --------------------------------------------------------------------------- |
| 123 | |
| 124 | |
| 125 | class TestParseGemma4Args: |
| 126 | """Values are returned as strings; type coercion to proper JSON types |
| 127 | happens at the engine layer.""" |
| 128 | |
| 129 | def test_empty_string(self): |
| 130 | assert _parse_gemma4_args("") == {} |
| 131 | |
| 132 | def test_whitespace_only(self): |
| 133 | assert _parse_gemma4_args(" ") == {} |
| 134 | |
| 135 | def test_single_string_value(self): |
| 136 | result = _parse_gemma4_args('location:<|"|>Paris<|"|>') |
| 137 | assert result == {"location": "Paris"} |
| 138 | |
| 139 | def test_string_value_with_comma(self): |
| 140 | result = _parse_gemma4_args('location:<|"|>Paris, France<|"|>') |
| 141 | assert result == {"location": "Paris, France"} |
| 142 | |
| 143 | def test_multiple_string_values(self): |
| 144 | result = _parse_gemma4_args( |
| 145 | 'location:<|"|>San Francisco<|"|>,unit:<|"|>celsius<|"|>' |
| 146 | ) |
| 147 | assert result == {"location": "San Francisco", "unit": "celsius"} |
| 148 | |
| 149 | def test_integer_value(self): |
| 150 | result = _parse_gemma4_args("count:42") |
| 151 | assert result == {"count": "42"} |
| 152 | |
| 153 | def test_float_value(self): |
| 154 | result = _parse_gemma4_args("score:3.14") |
| 155 | assert result == {"score": "3.14"} |
| 156 | |
| 157 | def test_boolean_true(self): |
| 158 | result = _parse_gemma4_args("flag:true") |
| 159 | assert result == {"flag": "true"} |
| 160 | |
| 161 | def test_boolean_false(self): |
| 162 | result = _parse_gemma4_args("flag:false") |
| 163 | assert result == {"flag": "false"} |
| 164 | |
| 165 | def test_null_value(self): |
| 166 | result = _parse_gemma4_args("param:null") |
| 167 | assert result == {"param": "null"} |
| 168 | |
| 169 | def test_mixed_types(self): |
| 170 | result = _parse_gemma4_args( |
| 171 | 'name:<|"|>test<|"|>,count:42,active:true,score:3.14' |
| 172 | ) |
| 173 | assert result == { |
| 174 | "name": "test", |
| 175 | "count": "42", |
| 176 | "active": "true", |
| 177 | "score": "3.14", |
| 178 | } |
| 179 | |
| 180 | def test_nested_object(self): |
| 181 | result = _parse_gemma4_args('nested:{inner:<|"|>value<|"|>}') |
| 182 | assert result == {"nested": {"inner": "value"}} |
| 183 | |
| 184 | def test_array_of_strings(self): |
| 185 | result = _parse_gemma4_args('items:[<|"|>a<|"|>,<|"|>b<|"|>]') |
| 186 | assert result == {"items": ["a", "b"]} |
| 187 | |
| 188 | def test_delimited_keys_stripped(self): |
| 189 | """Keys wrapped in <|"|> delimiters are stripped.""" |
| 190 | result = _parse_gemma4_args('<|"|>location<|"|>:<|"|>Paris<|"|>') |
| 191 | assert result == {"location": "Paris"} |
| 192 | |
| 193 | result = _parse_gemma4_args('outer:{<|"|>inner<|"|>:<|"|>val<|"|>}') |
| 194 | assert result == {"outer": {"inner": "val"}} |
| 195 | |
| 196 | result = _parse_gemma4_args('<|"|>name<|"|>:<|"|>Alice<|"|>,count:42') |
| 197 | assert result == {"name": "Alice", "count": "42"} |
| 198 | |
| 199 | def test_unterminated_string(self): |
| 200 | """Unterminated strings should take everything after the delimiter.""" |
| 201 | result = _parse_gemma4_args('key:<|"|>unterminated') |
| 202 | assert result == {"key": "unterminated"} |
| 203 | |
| 204 | def test_empty_value(self): |
| 205 | """Key with no value after colon.""" |
| 206 | result = _parse_gemma4_args("key:") |
| 207 | assert result == {"key": ""} |
| 208 | |
| 209 | def test_empty_value_partial_withheld(self): |
| 210 | """Key with no value is withheld in partial mode to avoid premature emission.""" |
| 211 | result = _parse_gemma4_args("key:", partial=True) |
| 212 | assert result == {} |
| 213 | # also with a space after the colon |
| 214 | result = _parse_gemma4_args("key: ", partial=True) |
| 215 | assert result == {} |
| 216 | |
| 217 | def test_empty_value_after_other_keys_partial_withheld(self): |
| 218 | """Trailing key with no value is withheld; earlier keys are kept.""" |
| 219 | result = _parse_gemma4_args('name:<|"|>test<|"|>,flag:', partial=True) |
| 220 | assert result == {"name": "test"} |
| 221 | |
| 222 | def test_trailing_dot_float_partial_withheld(self): |
| 223 | """Bare float ending with '.' is withheld in partial mode. |
| 224 | |
| 225 | Regression test for #42047: float("108.") → 108.0 causes |
| 226 | streaming diff corruption (108.0 → 108.2 becomes 108.02). |
| 227 | """ |
| 228 | # Single key with trailing dot — withheld entirely |
| 229 | result = _parse_gemma4_args("left:108.,right:22.8", partial=True) |
| 230 | assert result == {} |
| 231 | |
| 232 | # Stable key before trailing-dot key — stable key is kept |
| 233 | result = _parse_gemma4_args( |
| 234 | 'name:<|"|>test<|"|>,score:3.,count:1', partial=True |
| 235 | ) |
| 236 | assert result == {"name": "test"} |
| 237 | |
| 238 | # Non-partial mode parses trailing dot normally |
| 239 | result = _parse_gemma4_args("left:108.,right:22.8", partial=False) |
| 240 | assert result == {"left": "108.", "right": "22.8"} |
| 241 | |
| 242 | @pytest.mark.timeout(5) |
| 243 | def test_malformed_partial_array(self): |
| 244 | result = _parse_gemma4_args(":[t:[]") |
| 245 | assert isinstance(result, dict) |
| 246 | |
| 247 | |
| 248 | class TestParseGemma4Array: |
| 249 | def test_string_array(self): |
| 250 | result = _parse_gemma4_array('<|"|>a<|"|>,<|"|>b<|"|>') |
| 251 | assert result == ["a", "b"] |
| 252 | |
| 253 | def test_empty_array(self): |
| 254 | result = _parse_gemma4_array("") |
| 255 | assert result == [] |
| 256 | |
| 257 | def test_bare_values(self): |
| 258 | result = _parse_gemma4_array("42,true,3.14") |
| 259 | assert result == ["42", "true", "3.14"] |
| 260 | |
| 261 | @pytest.mark.timeout(5) |
| 262 | def test_string_element_with_closing_bracket(self): |
| 263 | result = _parse_gemma4_array('[<|"|>a]b<|"|>,<|"|>c<|"|>],<|"|>tail<|"|>') |
| 264 | assert result == [["a]b", "c"], "tail"] |
| 265 | |
| 266 | @pytest.mark.timeout(5) |
| 267 | def test_stray_closing_bracket(self): |
| 268 | result = _parse_gemma4_array("42,]trailing") |
| 269 | assert result == ["42"] |
| 270 | |
| 271 | def test_trailing_dot_float_partial_withheld(self): |
| 272 | """Array elements with trailing dot withheld in partial mode.""" |
| 273 | result = _parse_gemma4_array("108.,22.8", partial=True) |
| 274 | assert result == [] |
| 275 | |
| 276 | # Stable elements before trailing-dot element are kept |
| 277 | result = _parse_gemma4_array("42,108.,3", partial=True) |
| 278 | assert result == ["42"] |
| 279 | |
| 280 | |
| 281 | # --------------------------------------------------------------------------- |
| 282 | # Non-streaming extraction tests |
| 283 | # --------------------------------------------------------------------------- |
| 284 | |
| 285 | |
| 286 | class TestExtractToolCalls: |
| 287 | def test_no_tool_calls(self, parser, mock_request): |
| 288 | model_output = "Hello, how can I help you today?" |
| 289 | result = parser.extract_tool_calls(model_output, mock_request) |
| 290 | |
| 291 | assert result.tools_called is False |
| 292 | assert result.tool_calls == [] |
| 293 | assert result.content == model_output |
| 294 | |
| 295 | def test_single_tool_call(self, parser, mock_request): |
| 296 | model_output = ( |
| 297 | '<|tool_call>call:get_weather{location:<|"|>London<|"|>}<tool_call|>' |
| 298 | ) |
| 299 | result = parser.extract_tool_calls(model_output, mock_request) |
| 300 | |
| 301 | assert result.tools_called is True |
| 302 | assert len(result.tool_calls) == 1 |
| 303 | assert result.tool_calls[0].function.name == "get_weather" |
| 304 | args = json.loads(result.tool_calls[0].function.arguments) |
| 305 | assert args == {"location": "London"} |
| 306 | |
| 307 | def test_multiple_arguments(self, parser, mock_request): |
| 308 | model_output = ( |
| 309 | "<|tool_call>call:get_weather{" |
| 310 | 'location:<|"|>San Francisco<|"|>,' |
| 311 | 'unit:<|"|>celsius<|"|>}' |
| 312 | "<tool_call|>" |
| 313 | ) |
| 314 | result = parser.extract_tool_calls(model_output, mock_request) |
| 315 | |
| 316 | assert result.tools_called is True |
| 317 | assert len(result.tool_calls) == 1 |
| 318 | assert result.tool_calls[0].function.name == "get_weather" |
| 319 | args = json.loads(result.tool_calls[0].function.arguments) |
| 320 | assert args == {"location": "San Francisco", "unit": "celsius"} |
| 321 | |
| 322 | def test_text_before_tool_call(self, parser, mock_request): |
| 323 | model_output = ( |
| 324 | "Let me check the weather for you. " |
| 325 | '<|tool_call>call:get_weather{location:<|"|>Paris<|"|>}' |
| 326 | "<tool_call|>" |
| 327 | ) |
| 328 | result = parser.extract_tool_calls(model_output, mock_request) |
| 329 | |
| 330 | assert result.tools_called is True |
| 331 | assert result.content == "Let me check the weather for you." |
| 332 | assert len(result.tool_calls) == 1 |
| 333 | assert result.tool_calls[0].function.name == "get_weather" |
| 334 | |
| 335 | def test_multiple_tool_calls(self, parser, mock_request): |
| 336 | model_output = ( |
| 337 | '<|tool_call>call:get_weather{location:<|"|>London<|"|>}' |
| 338 | "<tool_call|>" |
| 339 | '<|tool_call>call:get_time{location:<|"|>London<|"|>}' |
| 340 | "<tool_call|>" |
| 341 | ) |
| 342 | result = parser.extract_tool_calls(model_output, mock_request) |
| 343 | |
| 344 | assert result.tools_called is True |
| 345 | assert len(result.tool_calls) == 2 |
| 346 | assert result.tool_calls[0].function.name == "get_weather" |
| 347 | assert result.tool_calls[1].function.name == "get_time" |
| 348 | |
| 349 | def test_nested_arguments(self, parser, mock_request): |
| 350 | model_output = ( |
| 351 | "<|tool_call>call:complex_function{" |
| 352 | 'nested:{inner:<|"|>value<|"|>},' |
| 353 | 'list:[<|"|>a<|"|>,<|"|>b<|"|>]}' |
| 354 | "<tool_call|>" |
| 355 | ) |
| 356 | result = parser.extract_tool_calls(model_output, mock_request) |
| 357 | |
| 358 | assert result.tools_called is True |
| 359 | assert len(result.tool_calls) == 1 |
| 360 | assert result.tool_calls[0].function.name == "complex_function" |
| 361 | args = json.loads(result.tool_calls[0].function.arguments) |
| 362 | assert args == {"nested": {"inner": "value"}, "list": ["a", "b"]} |
| 363 | |
| 364 | def test_tool_call_with_number_and_boolean(self, parser, mock_request): |
| 365 | model_output = ( |
| 366 | "<|tool_call>call:set_status{" |
| 367 | "is_active:true," |
| 368 | "count:42," |
| 369 | "score:3.14}" |
| 370 | "<tool_call|>" |
| 371 | ) |
| 372 | result = parser.extract_tool_calls(model_output, mock_request) |
| 373 | |
| 374 | assert result.tools_called is True |
| 375 | assert len(result.tool_calls) == 1 |
| 376 | assert result.tool_calls[0].function.name == "set_status" |
| 377 | args = json.loads(result.tool_calls[0].function.arguments) |
| 378 | assert args == {"is_active": True, "count": 42, "score": 3.14} |
| 379 | |
| 380 | def test_incomplete_tool_call(self, parser, mock_request): |
| 381 | model_output = '<|tool_call>call:get_weather{location:<|"|>London' |
| 382 | result = parser.extract_tool_calls(model_output, mock_request) |
| 383 | |
| 384 | assert result.tools_called is True |
| 385 | assert len(result.tool_calls) == 1 |
| 386 | assert result.tool_calls[0].function.name == "get_weather" |
| 387 | args = json.loads(result.tool_calls[0].function.arguments) |
| 388 | assert args == {"location": "London"} |
| 389 | |
| 390 | def test_hyphenated_function_name(self, parser, mock_request): |
| 391 | """Ensure function names with hyphens are parsed correctly.""" |
| 392 | model_output = ( |
| 393 | '<|tool_call>call:get-weather{location:<|"|>London<|"|>}<tool_call|>' |
| 394 | ) |
| 395 | result = parser.extract_tool_calls(model_output, mock_request) |
| 396 | |
| 397 | assert result.tools_called is True |
| 398 | assert result.tool_calls[0].function.name == "get-weather" |
| 399 | |
| 400 | def test_dotted_function_name(self, parser, mock_request): |
| 401 | """Ensure function names with dots are parsed correctly.""" |
| 402 | model_output = ( |
| 403 | '<|tool_call>call:weather.get{location:<|"|>London<|"|>}<tool_call|>' |
| 404 | ) |
| 405 | result = parser.extract_tool_calls(model_output, mock_request) |
| 406 | |
| 407 | assert result.tools_called is True |
| 408 | assert result.tool_calls[0].function.name == "weather.get" |
| 409 | |
| 410 | def test_no_arguments(self, parser, mock_request): |
| 411 | """Tool calls with empty arguments.""" |
| 412 | model_output = "<|tool_call>call:get_status{}<tool_call|>" |
| 413 | result = parser.extract_tool_calls(model_output, mock_request) |
| 414 | |
| 415 | assert result.tools_called is True |
| 416 | assert result.tool_calls[0].function.name == "get_status" |
| 417 | args = json.loads(result.tool_calls[0].function.arguments) |
| 418 | assert args == {} |
| 419 | |
| 420 | |
| 421 | # --------------------------------------------------------------------------- |
| 422 | # Streaming extraction tests |
| 423 | # --------------------------------------------------------------------------- |
| 424 | |
| 425 | |
| 426 | class TestStreamingExtraction: |
| 427 | """Tests for the streaming tool call extraction. |
| 428 | |
| 429 | These simulate the token-by-token streaming that vLLM performs, |
| 430 | feeding incremental text to extract_tool_calls_streaming() and |
| 431 | verifying that the accumulated argument deltas form valid JSON. |
| 432 | """ |
| 433 | |
| 434 | _SPECIAL_TOKEN_IDS = { |
| 435 | TOOL_CALL_START: TOOL_CALL_START_ID, |
| 436 | TOOL_CALL_END: TOOL_CALL_END_ID, |
| 437 | CHANNEL_START: CHANNEL_START_ID, |
| 438 | CHANNEL_END: CHANNEL_END_ID, |
| 439 | } |
| 440 | |
| 441 | def _simulate_streaming( |
| 442 | self, parser: Any, mock_request: Any, chunks: list[str] |
| 443 | ) -> list[tuple[Any, str]]: |
| 444 | """Feed chunks through the streaming parser and collect results. |
| 445 | |
| 446 | Returns a list of (delta_message, accumulated_text) tuples. |
| 447 | """ |
| 448 | results: list[tuple[Any, str]] = [] |
| 449 | previous_text: str = "" |
| 450 | previous_token_ids: list[int] = [] |
| 451 | |
| 452 | for chunk in chunks: |
| 453 | current_text = previous_text + chunk |
| 454 | found: list[tuple[int, int]] = [] |
| 455 | for token, tid in self._SPECIAL_TOKEN_IDS.items(): |
| 456 | pos = 0 |
| 457 | while True: |
| 458 | idx = chunk.find(token, pos) |
| 459 | if idx < 0: |
| 460 | break |
| 461 | found.append((idx, tid)) |
| 462 | pos = idx + len(token) |
| 463 | found.sort() |
| 464 | delta_token_ids: list[int] = [tid for _, tid in found] if found else [0] |
| 465 | |
| 466 | current_token_ids = previous_token_ids + delta_token_ids |
| 467 | |
| 468 | delta = parser.extract_tool_calls_streaming( |
| 469 | previous_text=previous_text, |
| 470 | current_text=current_text, |
| 471 | delta_text=chunk, |
| 472 | previous_token_ids=tuple(previous_token_ids), |
| 473 | current_token_ids=tuple(current_token_ids), |
| 474 | delta_token_ids=tuple(delta_token_ids), |
| 475 | request=mock_request, |
| 476 | ) |
| 477 | results.append((delta, current_text)) |
| 478 | previous_text = current_text |
| 479 | previous_token_ids = list(current_token_ids) |
| 480 | |
| 481 | return results |
| 482 | |
| 483 | def _collect_arguments(self, results): |
| 484 | """Collect all argument deltas from streaming results into one string.""" |
| 485 | args_text = "" |
| 486 | for delta, _ in results: |
| 487 | if delta and delta.tool_calls: |
| 488 | for tc in delta.tool_calls: |
| 489 | func = tc.function if isinstance(tc.function, dict) else tc.function |
| 490 | if isinstance(func, dict): |
| 491 | arg = func.get("arguments", "") |
| 492 | else: |
| 493 | arg = getattr(func, "arguments", "") or "" |
| 494 | if arg: |
| 495 | args_text += arg |
| 496 | return args_text |
| 497 | |
| 498 | def _collect_function_name(self, results): |
| 499 | """Extract the function name from streaming results.""" |
| 500 | for delta, _ in results: |
| 501 | if delta and delta.tool_calls: |
| 502 | for tc in delta.tool_calls: |
| 503 | func = tc.function if isinstance(tc.function, dict) else tc.function |
| 504 | if isinstance(func, dict): |
| 505 | name = func.get("name") |
| 506 | else: |
| 507 | name = getattr(func, "name", None) |
| 508 | if name: |
| 509 | return name |
| 510 | return None |
| 511 | |
| 512 | def test_basic_streaming_single_tool(self, parser, mock_request): |
| 513 | """Simulate the exact streaming scenario from the bug report. |
| 514 | |
| 515 | Model generates: |
| 516 | <|tool_call>call:get_weather{location:<|"|>Paris, France<|"|>}<tool_call|> |
| 517 | |
| 518 | Expected: arguments should be valid JSON {"location": "Paris, France"} |
| 519 | """ |
| 520 | chunks = [ |
| 521 | "<|tool_call>", |
| 522 | "call:get_weather{", |
| 523 | 'location:<|"|>Paris', |
| 524 | ", France", |
| 525 | '<|"|>}', |
| 526 | "<tool_call|>", |
| 527 | ] |
| 528 | |
| 529 | results = self._simulate_streaming(parser, mock_request, chunks) |
| 530 | |
| 531 | # Verify function name |
| 532 | name = self._collect_function_name(results) |
| 533 | assert name == "get_weather", f"Expected 'get_weather', got '{name}'" |
| 534 | |
| 535 | # Verify arguments form valid JSON |
| 536 | args_text = self._collect_arguments(results) |
| 537 | assert args_text, "No arguments were streamed" |
| 538 | parsed_args = json.loads(args_text) |
| 539 | assert parsed_args == {"location": "Paris, France"} |
| 540 | |
| 541 | def test_streaming_multi_arg(self, parser, mock_request): |
| 542 | """Streaming with multiple arguments.""" |
| 543 | chunks = [ |
| 544 | "<|tool_call>", |
| 545 | "call:get_weather{", |
| 546 | 'location:<|"|>Tokyo<|"|>,', |
| 547 | 'unit:<|"|>celsius<|"|>}', |
| 548 | "<tool_call|>", |
| 549 | ] |
| 550 | |
| 551 | results = self._simulate_streaming(parser, mock_request, chunks) |
| 552 | |
| 553 | name = self._collect_function_name(results) |
| 554 | assert name == "get_weather" |
| 555 | |
| 556 | args_text = self._collect_arguments(results) |
| 557 | assert args_text |
| 558 | parsed_args = json.loads(args_text) |
| 559 | assert parsed_args == {"location": "Tokyo", "unit": "celsius"} |
| 560 | |
| 561 | def test_streaming_no_extra_brace(self, parser, mock_request): |
| 562 | """Verify the closing } is NOT leaked into arguments (Bug #2).""" |
| 563 | chunks = [ |
| 564 | "<|tool_call>", |
| 565 | "call:get_weather{", |
| 566 | 'location:<|"|>London<|"|>}', |
| 567 | "<tool_call|>", |
| 568 | ] |
| 569 | |
| 570 | results = self._simulate_streaming(parser, mock_request, chunks) |
| 571 | args_text = self._collect_arguments(results) |
| 572 | assert args_text |
| 573 | |
| 574 | # The args text must be valid JSON (no extra }) |
| 575 | parsed = json.loads(args_text) |
| 576 | assert parsed == {"location": "London"} |
| 577 | |
| 578 | # Specifically assert no double-brace |
| 579 | assert args_text.count("}") <= 1, ( |
| 580 | f"Arguments contain extra closing brace: {args_text!r}" |
| 581 | ) |
| 582 | |
| 583 | def test_streaming_no_unquoted_keys(self, parser, mock_request): |
| 584 | """Verify keys are properly quoted in JSON (Bug #1).""" |
| 585 | chunks = [ |
| 586 | "<|tool_call>", |
| 587 | "call:get_weather{", |
| 588 | 'location:<|"|>Paris<|"|>}', |
| 589 | "<tool_call|>", |
| 590 | ] |
| 591 | |
| 592 | results = self._simulate_streaming(parser, mock_request, chunks) |
| 593 | args_text = self._collect_arguments(results) |
| 594 | |
| 595 | # Must start with { and contain quoted key |
| 596 | assert args_text.lstrip().startswith("{"), ( |
| 597 | f"Arguments don't start with '{{': {args_text!r}" |
| 598 | ) |
| 599 | assert '"location"' in args_text, ( |
| 600 | f"Key 'location' not properly quoted: {args_text!r}" |
| 601 | ) |
| 602 | |
| 603 | def test_streaming_name_no_call_prefix(self, parser, mock_request): |
| 604 | """Verify function name has no 'call:' prefix.""" |
| 605 | chunks = [ |
| 606 | "<|tool_call>", |
| 607 | "call:get_weather{", |
| 608 | 'location:<|"|>Paris<|"|>}', |
| 609 | "<tool_call|>", |
| 610 | ] |
| 611 | |
| 612 | results = self._simulate_streaming(parser, mock_request, chunks) |
| 613 | name = self._collect_function_name(results) |
| 614 | assert name == "get_weather" |
| 615 | assert not name.startswith("call:"), f"Name has 'call:' prefix: {name!r}" |
| 616 | |
| 617 | def test_streaming_text_before_tool_call(self, parser, mock_request): |
| 618 | """Text before tool call should be emitted as content.""" |
| 619 | chunks = [ |
| 620 | "Let me check ", |
| 621 | "the weather. ", |
| 622 | "<|tool_call>", |
| 623 | "call:get_weather{", |
| 624 | 'location:<|"|>London<|"|>}', |
| 625 | "<tool_call|>", |
| 626 | ] |
| 627 | |
| 628 | results = self._simulate_streaming(parser, mock_request, chunks) |
| 629 | |
| 630 | # First chunks should be content |
| 631 | content_parts = [] |
| 632 | for delta, _ in results: |
| 633 | if delta and delta.content: |
| 634 | content_parts.append(delta.content) |
| 635 | |
| 636 | assert "".join(content_parts).strip().startswith("Let me check") |
| 637 | |
| 638 | def test_streaming_numeric_args(self, parser, mock_request): |
| 639 | """Streaming with numeric and boolean argument values.""" |
| 640 | chunks = [ |
| 641 | "<|tool_call>", |
| 642 | "call:set_config{", |
| 643 | "count:42,", |
| 644 | "active:true}", |
| 645 | "<tool_call|>", |
| 646 | ] |
| 647 | |
| 648 | results = self._simulate_streaming(parser, mock_request, chunks) |
| 649 | args_text = self._collect_arguments(results) |
| 650 | assert args_text is not None |
| 651 | parsed_args = json.loads(args_text) |
| 652 | assert parsed_args["count"] == 42 |
| 653 | assert parsed_args["active"] is True |
| 654 | |
| 655 | def test_streaming_boolean_split_across_chunks(self, parser, mock_request): |
| 656 | """Boolean value split across token boundaries must not corrupt JSON.""" |
| 657 | chunks = [ |
| 658 | "<|tool_call>", |
| 659 | "call:search{input:{all:" + "true"[:3], |
| 660 | "e}}", |
| 661 | "<tool_call|>", |
| 662 | ] |
| 663 | |
| 664 | results = self._simulate_streaming(parser, mock_request, chunks) |
| 665 | args_text = self._collect_arguments(results) |
| 666 | assert args_text, "No arguments were streamed" |
| 667 | parsed_args = json.loads(args_text) |
| 668 | assert parsed_args["input"]["all"] is True |
| 669 | |
| 670 | def test_streaming_false_split_across_chunks(self, parser, mock_request): |
| 671 | """Boolean false split across chunks.""" |
| 672 | chunks = [ |
| 673 | "<|tool_call>", |
| 674 | "call:set{flag:" + "false"[:4], |
| 675 | "e}", |
| 676 | "<tool_call|>", |
| 677 | ] |
| 678 | |
| 679 | results = self._simulate_streaming(parser, mock_request, chunks) |
| 680 | args_text = self._collect_arguments(results) |
| 681 | assert args_text, "No arguments were streamed" |
| 682 | parsed_args = json.loads(args_text) |
| 683 | assert parsed_args["flag"] is False |
| 684 | |
| 685 | def test_streaming_number_split_across_chunks(self, parser, mock_request): |
| 686 | """Number split across chunks must not change type.""" |
| 687 | chunks = [ |
| 688 | "<|tool_call>", |
| 689 | "call:set{count:4", |
| 690 | "2}", |
| 691 | "<tool_call|>", |
| 692 | ] |
| 693 | |
| 694 | results = self._simulate_streaming(parser, mock_request, chunks) |
| 695 | args_text = self._collect_arguments(results) |
| 696 | assert args_text, "No arguments were streamed" |
| 697 | parsed_args = json.loads(args_text) |
| 698 | assert parsed_args["count"] == 42 |
| 699 | |
| 700 | def test_streaming_empty_args(self, parser, mock_request): |
| 701 | """Tool call with no arguments.""" |
| 702 | chunks = [ |
| 703 | "<|tool_call>", |
| 704 | "call:get_status{}", |
| 705 | "<tool_call|>", |
| 706 | ] |
| 707 | |
| 708 | results = self._simulate_streaming(parser, mock_request, chunks) |
| 709 | name = self._collect_function_name(results) |
| 710 | assert name == "get_status" |
| 711 | |
| 712 | def test_streaming_split_delimiter_no_invalid_json(self, parser, mock_request): |
| 713 | """Partial <|"|> delimiter chars must not leak into streamed JSON. |
| 714 | |
| 715 | Reproduces the bug from https://github.com/vllm-project/vllm/issues/38946 |
| 716 | where a token boundary splits the string delimiter, leaving fragments |
| 717 | like '<|' at the end of a parsed value which then corrupt the JSON. |
| 718 | """ |
| 719 | chunks = [ |
| 720 | "<|tool_call>", |
| 721 | "call:todowrite{", |
| 722 | 'content:<|"|>Buy milk<|', |
| 723 | '"|>}', |
| 724 | "<tool_call|>", |
| 725 | ] |
| 726 | |
| 727 | results = self._simulate_streaming(parser, mock_request, chunks) |
| 728 | |
| 729 | args_text = self._collect_arguments(results) |
| 730 | assert args_text, "No arguments were streamed" |
| 731 | |
| 732 | # Must be valid JSON — the original bug caused a JSON parse error |
| 733 | parsed_args = json.loads(args_text) |
| 734 | assert parsed_args["content"] == "Buy milk" |
| 735 | |
| 736 | # Ensure no raw delimiter fragments leaked into the JSON |
| 737 | assert "<|" not in args_text, ( |
| 738 | f"Partial delimiter leaked into JSON: {args_text!r}" |
| 739 | ) |
| 740 | |
| 741 | def test_streaming_does_not_duplicate_plain_text_after_tool_call( |
| 742 | self, parser, mock_request |
| 743 | ): |
| 744 | """Buffered plain text after a tool call must not corrupt content.""" |
| 745 | chunks = [ |
| 746 | "<|tool_call>", |
| 747 | "call:get_weather{", |
| 748 | 'location:<|"|>Paris<|"|>}', |
| 749 | "<tool_call|>", |
| 750 | "<", |
| 751 | "div>", |
| 752 | ] |
| 753 | |
| 754 | results = self._simulate_streaming(parser, mock_request, chunks) |
| 755 | content_parts = [ |
| 756 | delta.content for delta, _ in results if delta is not None and delta.content |
| 757 | ] |
| 758 | assert "".join(content_parts) == "<div>" |
| 759 | assert "<<div>" not in "".join(content_parts) |
| 760 | |
| 761 | def test_streaming_html_argument_does_not_duplicate_tag_prefixes( |
| 762 | self, parser, mock_request |
| 763 | ): |
| 764 | """HTML content inside tool arguments must not be duplicated.""" |
| 765 | chunks = [ |
| 766 | "<|tool_call>", |
| 767 | "call:write_file{", |
| 768 | 'path:<|"|>index.html<|"|>,', |
| 769 | 'content:<|"|><!DOCTYPE html>\n<', |
| 770 | 'html lang="zh-CN">\n<', |
| 771 | "head>\n <", |
| 772 | 'meta charset="UTF-8">\n <', |
| 773 | 'meta name="viewport" content="width=device-width">\n', |
| 774 | '<|"|>}', |
| 775 | "<tool_call|>", |
| 776 | ] |
| 777 | |
| 778 | results = self._simulate_streaming(parser, mock_request, chunks) |
| 779 | args_text = self._collect_arguments(results) |
| 780 | assert args_text |
| 781 | |
| 782 | parsed_args = json.loads(args_text) |
| 783 | assert parsed_args["path"] == "index.html" |
| 784 | assert ( |
| 785 | parsed_args["content"] == "<!DOCTYPE html>\n" |
| 786 | '<html lang="zh-CN">\n' |
| 787 | "<head>\n" |
| 788 | ' <meta charset="UTF-8">\n' |
| 789 | ' <meta name="viewport" content="width=device-width">\n' |
| 790 | ) |
| 791 | |
| 792 | def _collect_tool_calls_by_index(self, results): |
| 793 | """Group streamed tool-call fragments by their ``index``. |
| 794 | |
| 795 | Returns ``{index: {"name": str | None, "arguments": str}}`` where |
| 796 | ``arguments`` is the concatenation of every streamed argument |
| 797 | fragment for that index (which should form valid JSON once complete). |
| 798 | """ |
| 799 | by_index: dict[int, dict[str, Any]] = {} |
| 800 | for delta, _ in results: |
| 801 | if not (delta and delta.tool_calls): |
| 802 | continue |
| 803 | for tc in delta.tool_calls: |
| 804 | entry = by_index.setdefault(tc.index, {"name": None, "arguments": ""}) |
| 805 | func = tc.function |
| 806 | if isinstance(func, dict): |
| 807 | name = func.get("name") |
| 808 | arg = func.get("arguments", "") |
| 809 | else: |
| 810 | name = getattr(func, "name", None) |
| 811 | arg = getattr(func, "arguments", "") or "" |
| 812 | if name: |
| 813 | entry["name"] = name |
| 814 | if arg: |
| 815 | entry["arguments"] += arg |
| 816 | return by_index |
| 817 | |
| 818 | def test_streaming_single_chunk_complete_tool_call(self, parser, mock_request): |
| 819 | """A backend may deliver a whole tool call in one streaming delta. |
| 820 | |
| 821 | The start token, ``call:name{...}`` payload and the end token all |
| 822 | arrive in a single chunk. The parser must still emit one |
| 823 | ``DeltaToolCall`` with the correct name + complete arguments JSON |
| 824 | (rather than swallowing it and finishing with finish_reason="stop"). |
| 825 | """ |
| 826 | chunks = [ |
| 827 | '<|tool_call>call:name_a_color{color_hex:<|"|>00ff11<|"|>}<tool_call|>', |
| 828 | ] |
| 829 | |
| 830 | results = self._simulate_streaming(parser, mock_request, chunks) |
| 831 | |
| 832 | # Exactly one delta should carry tool_calls, and it must not be |
| 833 | # emitted as plain content (which would yield finish_reason="stop"). |
| 834 | tool_call_deltas = [ |
| 835 | delta for delta, _ in results if delta is not None and delta.tool_calls |
| 836 | ] |
| 837 | assert len(tool_call_deltas) == 1, ( |
| 838 | "Expected exactly one delta carrying the batched tool call" |
| 839 | ) |
| 840 | assert all( |
| 841 | delta.content is None for delta, _ in results if delta is not None |
| 842 | ), "Complete tool call must not leak as content" |
| 843 | |
| 844 | by_index = self._collect_tool_calls_by_index(results) |
| 845 | assert set(by_index) == {0} |
| 846 | assert by_index[0]["name"] == "name_a_color" |
| 847 | assert json.loads(by_index[0]["arguments"]) == {"color_hex": "00ff11"} |
| 848 | |
| 849 | def test_streaming_multi_chunk_batched_tool_calls(self, parser, mock_request): |
| 850 | """A single delta may batch MULTIPLE complete tool calls. |
| 851 | |
| 852 | ``<|tool_call>...<tool_call|><|tool_call>...<tool_call|>`` arriving in |
| 853 | one chunk must emit BOTH calls (one DeltaToolCall each, with distinct |
| 854 | indices), not just the first. |
| 855 | """ |
| 856 | chunks = [ |
| 857 | '<|tool_call>call:get_weather{location:<|"|>London<|"|>}<tool_call|>' |
| 858 | '<|tool_call>call:get_time{timezone:<|"|>GMT<|"|>}<tool_call|>', |
| 859 | ] |
| 860 | |
| 861 | results = self._simulate_streaming(parser, mock_request, chunks) |
| 862 | |
| 863 | by_index = self._collect_tool_calls_by_index(results) |
| 864 | assert set(by_index) == {0, 1}, ( |
| 865 | f"Expected two tool calls (indices 0 and 1), got {sorted(by_index)}" |
| 866 | ) |
| 867 | |
| 868 | assert by_index[0]["name"] == "get_weather" |
| 869 | assert json.loads(by_index[0]["arguments"]) == {"location": "London"} |
| 870 | |
| 871 | assert by_index[1]["name"] == "get_time" |
| 872 | assert json.loads(by_index[1]["arguments"]) == {"timezone": "GMT"} |
| 873 | |
| 874 | def test_streaming_trailing_bare_bool_not_duplicated(self, parser, mock_request): |
| 875 | """Trailing bare boolean must not be streamed twice.""" |
| 876 | chunks = [ |
| 877 | "<|tool_call>", |
| 878 | "call:Edit{", |
| 879 | 'file_path:<|"|>src/env.py<|"|>,', |
| 880 | 'old_string:<|"|>old_val<|"|>,', |
| 881 | 'new_string:<|"|>new_val<|"|>,', |
| 882 | "replace_all:", |
| 883 | "false}", |
| 884 | "<tool_call|>", |
| 885 | ] |
| 886 | |
| 887 | results = self._simulate_streaming(parser, mock_request, chunks) |
| 888 | args_text = self._collect_arguments(results) |
| 889 | assert args_text, "No arguments were streamed" |
| 890 | |
| 891 | parsed_args = json.loads(args_text) |
| 892 | assert parsed_args == { |
| 893 | "file_path": "src/env.py", |
| 894 | "old_string": "old_val", |
| 895 | "new_string": "new_val", |
| 896 | "replace_all": False, |
| 897 | } |
| 898 | |
| 899 | assert args_text.count("replace_all") == 1 |
| 900 | |
| 901 | |
| 902 | class TestBareAndMalformedOpeners: |
| 903 | """Tests for bare and malformed tool call openers. |
| 904 | |
| 905 | See https://github.com/vllm-project/vllm/issues/53431. |
| 906 | """ |
| 907 | |
| 908 | def test_bare_opener_non_streaming(self, parser, mock_request): |
| 909 | model_output = '<|tool_call>:get_weather{location:<|"|>London<|"|>}<tool_call|>' |
| 910 | result = parser.extract_tool_calls(model_output, mock_request) |
| 911 | |
| 912 | assert result.tools_called is True |
| 913 | assert len(result.tool_calls) == 1 |
| 914 | assert result.tool_calls[0].function.name == "get_weather" |
| 915 | args = json.loads(result.tool_calls[0].function.arguments) |
| 916 | assert args == {"location": "London"} |
| 917 | |
| 918 | def test_documented_form_still_works(self, parser, mock_request): |
| 919 | model_output = ( |
| 920 | '<|tool_call>call:get_weather{location:<|"|>London<|"|>}<tool_call|>' |
| 921 | ) |
| 922 | result = parser.extract_tool_calls(model_output, mock_request) |
| 923 | |
| 924 | assert result.tools_called is True |
| 925 | assert len(result.tool_calls) == 1 |
| 926 | assert result.tool_calls[0].function.name == "get_weather" |
| 927 | |
| 928 | def test_bare_opener_multiple_args(self, parser, mock_request): |
| 929 | model_output = ( |
| 930 | "<|tool_call>:get_weather{" |
| 931 | 'location:<|"|>Tokyo<|"|>,unit:<|"|>celsius<|"|>}' |
| 932 | "<tool_call|>" |
| 933 | ) |
| 934 | result = parser.extract_tool_calls(model_output, mock_request) |
| 935 | |
| 936 | assert result.tools_called is True |
| 937 | assert result.tool_calls[0].function.name == "get_weather" |
| 938 | args = json.loads(result.tool_calls[0].function.arguments) |
| 939 | assert args == {"location": "Tokyo", "unit": "celsius"} |
| 940 | |
| 941 | def test_malformed_no_brace_bounded(self, parser, mock_request): |
| 942 | model_output = "<|tool_call>call:bad_func no brace<tool_call|>" |
| 943 | result = parser.extract_tool_calls(model_output, mock_request) |
| 944 | |
| 945 | assert result.tools_called is True |
| 946 | assert len(result.tool_calls) == 1 |
| 947 | name = result.tool_calls[0].function.name |
| 948 | assert "<tool_call|>" not in name |
| 949 | |
| 950 | def test_malformed_then_well_formed(self, parser, mock_request): |
| 951 | model_output = ( |
| 952 | "<|tool_call>call:bad_func no brace<tool_call|>" |
| 953 | '<|tool_call>call:get_weather{location:<|"|>London<|"|>}' |
| 954 | "<tool_call|>" |
| 955 | ) |
| 956 | result = parser.extract_tool_calls(model_output, mock_request) |
| 957 | |
| 958 | assert result.tools_called is True |
| 959 | assert len(result.tool_calls) == 2 |
| 960 | assert result.tool_calls[0].function.name == "bad_func no brace" |
| 961 | assert result.tool_calls[1].function.name == "get_weather" |
| 962 | assert "London" in result.tool_calls[1].function.arguments |
| 963 | |
| 964 | def test_bare_malformed_then_well_formed(self, parser, mock_request): |
| 965 | model_output = ( |
| 966 | "<|tool_call>:bad_func no brace<tool_call|>" |
| 967 | '<|tool_call>call:get_weather{location:<|"|>London<|"|>}' |
| 968 | "<tool_call|>" |
| 969 | ) |
| 970 | result = parser.extract_tool_calls(model_output, mock_request) |
| 971 | |
| 972 | assert result.tools_called is True |
| 973 | assert len(result.tool_calls) == 2 |
| 974 | assert result.tool_calls[0].function.name == "bad_func no brace" |
| 975 | assert result.tool_calls[1].function.name == "get_weather" |
| 976 | |
| 977 | def test_three_mixed_calls(self, parser, mock_request): |
| 978 | model_output = ( |
| 979 | "<|tool_call>:bad no brace<tool_call|>" |
| 980 | '<|tool_call>call:search{input:<|"|>hello<|"|>}<tool_call|>' |
| 981 | "<|tool_call>:set{flag:true,count:5}<tool_call|>" |
| 982 | ) |
| 983 | result = parser.extract_tool_calls(model_output, mock_request) |
| 984 | |
| 985 | assert result.tools_called is True |
| 986 | assert len(result.tool_calls) == 3 |
| 987 | assert result.tool_calls[0].function.name == "bad no brace" |
| 988 | assert result.tool_calls[1].function.name == "search" |
| 989 | assert result.tool_calls[2].function.name == "set" |
| 990 | |
| 991 | def test_colon_in_content_not_broken(self, parser, mock_request): |
| 992 | model_output = ( |
| 993 | "Hello: world" |
| 994 | '<|tool_call>:get_weather{location:<|"|>London<|"|>}<tool_call|>' |
| 995 | ) |
| 996 | result = parser.extract_tool_calls(model_output, mock_request) |
| 997 | |
| 998 | assert result.tools_called is True |
| 999 | assert result.content is not None |
| 1000 | assert "Hello" in result.content |
| 1001 | |
| 1002 | def test_bare_opener_empty_args(self, parser, mock_request): |
| 1003 | model_output = "<|tool_call>:set_status{}<tool_call|>" |
| 1004 | result = parser.extract_tool_calls(model_output, mock_request) |
| 1005 | |
| 1006 | assert result.tools_called is True |
| 1007 | assert result.tool_calls[0].function.name == "set_status" |