| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | """Unit tests for DeepSeekV32EngineToolParser. |
| 5 | |
| 6 | These tests use a minimal mock tokenizer so no real model weights are required. |
| 7 | """ |
| 8 | |
| 9 | import json |
| 10 | from unittest.mock import MagicMock |
| 11 | |
| 12 | import pytest |
| 13 | from openai.types.responses.function_tool import FunctionTool |
| 14 | |
| 15 | from tests.tool_parsers.utils import run_tool_extraction_streaming |
| 16 | from vllm.entrypoints.openai.chat_completion.protocol import ( |
| 17 | ChatCompletionToolsParam, |
| 18 | FunctionDefinition, |
| 19 | ) |
| 20 | from vllm.tool_parsers.deepseekv32_engine_tool_parser import ( |
| 21 | DeepSeekV32EngineToolParser, |
| 22 | ) |
| 23 | |
| 24 | pytestmark = pytest.mark.skip_global_cleanup |
| 25 | |
| 26 | # --------------------------------------------------------------------------- |
| 27 | # Helpers |
| 28 | # --------------------------------------------------------------------------- |
| 29 | |
| 30 | # Token IDs are not used by the V32 parser logic, so we only need the |
| 31 | # tokenizer object to be truthy (the parser checks `if not self.model_tokenizer`). |
| 32 | MOCK_TOKENIZER = MagicMock() |
| 33 | MOCK_TOKENIZER.get_vocab.return_value = {} |
| 34 | MOCK_TOKENIZER.tokenize.return_value = [] |
| 35 | |
| 36 | |
| 37 | def make_parser(tools=None) -> DeepSeekV32EngineToolParser: |
| 38 | return DeepSeekV32EngineToolParser(MOCK_TOKENIZER, tools=tools) |
| 39 | |
| 40 | |
| 41 | def make_tool_param(name: str, params: dict) -> MagicMock: |
| 42 | """Build a mock tool matching the ChatCompletionToolsParam shape.""" |
| 43 | tool = MagicMock() |
| 44 | tool.function.name = name |
| 45 | tool.function.parameters = params |
| 46 | return tool |
| 47 | |
| 48 | |
| 49 | def make_request(tools=None) -> MagicMock: |
| 50 | req = MagicMock() |
| 51 | req.tools = tools |
| 52 | return req |
| 53 | |
| 54 | |
| 55 | # Shorthand for the DSML tokens used throughout |
| 56 | FC_START = "<｜DSML｜function_calls>" |
| 57 | FC_END = "</｜DSML｜function_calls>" |
| 58 | INV_START = '<｜DSML｜invoke name="' |
| 59 | INV_END = "</｜DSML｜invoke>" |
| 60 | PARAM_START = '<｜DSML｜parameter name="' |
| 61 | PARAM_END = "</｜DSML｜parameter>" |
| 62 | |
| 63 | |
| 64 | def build_tool_call(func_name: str, params: dict[str, str]) -> str: |
| 65 | """Build a complete model-output tool call string.""" |
| 66 | param_strs = "".join( |
| 67 | f'{PARAM_START}{k}" string="true">{v}{PARAM_END}' for k, v in params.items() |
| 68 | ) |
| 69 | return f'{FC_START}\n{INV_START}{func_name}">\n{param_strs}\n{INV_END}\n{FC_END}' |
| 70 | |
| 71 | |
| 72 | # --------------------------------------------------------------------------- |
| 73 | # Tests: extract_tool_calls (non-streaming) |
| 74 | # --------------------------------------------------------------------------- |
| 75 | |
| 76 | |
| 77 | class TestExtractToolCalls: |
| 78 | @pytest.fixture |
| 79 | def parser(self): |
| 80 | return make_parser() |
| 81 | |
| 82 | def test_no_tool_call(self, parser): |
| 83 | result = parser.extract_tool_calls("just some text", None) |
| 84 | assert not result.tools_called |
| 85 | assert result.tool_calls == [] |
| 86 | assert result.content == "just some text" |
| 87 | |
| 88 | def test_single_tool_no_params(self, parser): |
| 89 | model_output = f'{FC_START}\n{INV_START}get_time">\n{INV_END}\n{FC_END}' |
| 90 | result = parser.extract_tool_calls(model_output, None) |
| 91 | assert result.tools_called |
| 92 | assert len(result.tool_calls) == 1 |
| 93 | assert result.tool_calls[0].function.name == "get_time" |
| 94 | assert json.loads(result.tool_calls[0].function.arguments) == {} |
| 95 | |
| 96 | def test_single_tool_with_params(self, parser): |
| 97 | model_output = build_tool_call( |
| 98 | "get_weather", {"location": "SF", "date": "2024-01-16"} |
| 99 | ) |
| 100 | result = parser.extract_tool_calls(model_output, None) |
| 101 | assert result.tools_called |
| 102 | assert len(result.tool_calls) == 1 |
| 103 | tc = result.tool_calls[0] |
| 104 | assert tc.function.name == "get_weather" |
| 105 | assert json.loads(tc.function.arguments) == { |
| 106 | "location": "SF", |
| 107 | "date": "2024-01-16", |
| 108 | } |
| 109 | |
| 110 | def test_content_before_tool_call(self, parser): |
| 111 | model_output = "Sure, let me check! " + build_tool_call( |
| 112 | "get_weather", {"location": "NYC"} |
| 113 | ) |
| 114 | result = parser.extract_tool_calls(model_output, None) |
| 115 | assert result.tools_called |
| 116 | assert result.content == "Sure, let me check! " |
| 117 | |
| 118 | def test_no_content_prefix_returns_none(self, parser): |
| 119 | model_output = build_tool_call("get_weather", {"location": "NYC"}) |
| 120 | result = parser.extract_tool_calls(model_output, None) |
| 121 | assert result.tools_called |
| 122 | assert result.content is None |
| 123 | |
| 124 | def test_multiple_tools(self, parser): |
| 125 | model_output = ( |
| 126 | f"{FC_START}\n" |
| 127 | f'{INV_START}get_weather">\n' |
| 128 | f'{PARAM_START}location" string="true">SF{PARAM_END}\n' |
| 129 | f"{INV_END}\n" |
| 130 | f'{INV_START}get_weather">\n' |
| 131 | f'{PARAM_START}location" string="true">NYC{PARAM_END}\n' |
| 132 | f"{INV_END}\n" |
| 133 | f"{FC_END}" |
| 134 | ) |
| 135 | result = parser.extract_tool_calls(model_output, None) |
| 136 | assert result.tools_called |
| 137 | assert len(result.tool_calls) == 2 |
| 138 | assert json.loads(result.tool_calls[0].function.arguments) == {"location": "SF"} |
| 139 | assert json.loads(result.tool_calls[1].function.arguments) == { |
| 140 | "location": "NYC" |
| 141 | } |
| 142 | |
| 143 | def test_type_conversion_in_non_streaming(self): |
| 144 | """Non-streaming extraction must convert params using the tool schema.""" |
| 145 | tool = ChatCompletionToolsParam( |
| 146 | function=FunctionDefinition( |
| 147 | name="toggle", |
| 148 | parameters={ |
| 149 | "type": "object", |
| 150 | "properties": { |
| 151 | "enabled": {"type": "boolean"}, |
| 152 | "count": {"type": "integer"}, |
| 153 | }, |
| 154 | }, |
| 155 | ), |
| 156 | ) |
| 157 | parser = make_parser(tools=[tool]) |
| 158 | model_output = ( |
| 159 | f"{FC_START}\n" |
| 160 | f'{INV_START}toggle">\n' |
| 161 | f'{PARAM_START}enabled" string="false">true{PARAM_END}\n' |
| 162 | f'{PARAM_START}count" string="false">42{PARAM_END}\n' |
| 163 | f"{INV_END}\n" |
| 164 | f"{FC_END}" |
| 165 | ) |
| 166 | result = parser.extract_tool_calls(model_output, None) |
| 167 | assert result.tools_called |
| 168 | assert len(result.tool_calls) == 1 |
| 169 | args = json.loads(result.tool_calls[0].function.arguments) |
| 170 | assert args == {"enabled": True, "count": 42} |
| 171 | assert isinstance(args["enabled"], bool) |
| 172 | assert isinstance(args["count"], int) |
| 173 | |
| 174 | def test_string_attr_true_coerced_by_schema(self): |
| 175 | """string="true" delivers a string, but the engine's schema-aware |
| 176 | type fixer coerces it to the schema type (integer).""" |
| 177 | tool = ChatCompletionToolsParam( |
| 178 | function=FunctionDefinition( |
| 179 | name="score", |
| 180 | parameters={ |
| 181 | "type": "object", |
| 182 | "properties": { |
| 183 | "value": {"type": "integer"}, |
| 184 | }, |
| 185 | }, |
| 186 | ), |
| 187 | ) |
| 188 | parser = make_parser(tools=[tool]) |
| 189 | model_output = ( |
| 190 | f"{FC_START}\n" |
| 191 | f'{INV_START}score">\n' |
| 192 | f'{PARAM_START}value" string="true">42{PARAM_END}\n' |
| 193 | f"{INV_END}\n" |
| 194 | f"{FC_END}" |
| 195 | ) |
| 196 | result = parser.extract_tool_calls(model_output, None) |
| 197 | assert result.tools_called |
| 198 | args = json.loads(result.tool_calls[0].function.arguments) |
| 199 | assert args == {"value": 42} |
| 200 | assert isinstance(args["value"], int) |
| 201 | |
| 202 | def test_string_attr_false_allows_schema_conversion(self): |
| 203 | """string="false" allows the parser to convert via the tool schema.""" |
| 204 | tool = ChatCompletionToolsParam( |
| 205 | function=FunctionDefinition( |
| 206 | name="score", |
| 207 | parameters={ |
| 208 | "type": "object", |
| 209 | "properties": { |
| 210 | "value": {"type": "integer"}, |
| 211 | }, |
| 212 | }, |
| 213 | ), |
| 214 | ) |
| 215 | parser = make_parser(tools=[tool]) |
| 216 | model_output = ( |
| 217 | f"{FC_START}\n" |
| 218 | f'{INV_START}score">\n' |
| 219 | f'{PARAM_START}value" string="false">42{PARAM_END}\n' |
| 220 | f"{INV_END}\n" |
| 221 | f"{FC_END}" |
| 222 | ) |
| 223 | result = parser.extract_tool_calls(model_output, None) |
| 224 | assert result.tools_called |
| 225 | args = json.loads(result.tool_calls[0].function.arguments) |
| 226 | assert args == {"value": 42} |
| 227 | assert isinstance(args["value"], int) |
| 228 | |
| 229 | def test_composed_schema_converts_object_and_array_params(self): |
| 230 | """Composed JSON Schema types must still drive DSML type coercion.""" |
| 231 | tool = ChatCompletionToolsParam( |
| 232 | function=FunctionDefinition( |
| 233 | name="set_timer", |
| 234 | parameters={ |
| 235 | "type": "object", |
| 236 | "properties": { |
| 237 | "wait": { |
| 238 | "anyOf": [ |
| 239 | { |
| 240 | "type": "object", |
| 241 | "properties": { |
| 242 | "type": {"const": "until"}, |
| 243 | "date": {"type": "string"}, |
| 244 | }, |
| 245 | }, |
| 246 | { |
| 247 | "type": "object", |
| 248 | "properties": { |
| 249 | "type": {"const": "for"}, |
| 250 | "minutes": {"type": "number"}, |
| 251 | }, |
| 252 | }, |
| 253 | ], |
| 254 | }, |
| 255 | "patches": { |
| 256 | "oneOf": [ |
| 257 | {"type": "array", "items": {"type": "object"}}, |
| 258 | {"type": "null"}, |
| 259 | ], |
| 260 | }, |
| 261 | }, |
| 262 | }, |
| 263 | ), |
| 264 | ) |
| 265 | parser = make_parser(tools=[tool]) |
| 266 | model_output = ( |
| 267 | f"{FC_START}\n" |
| 268 | f'{INV_START}set_timer">\n' |
| 269 | f'{PARAM_START}wait" string="false">' |
| 270 | f'{{"type":"for","minutes":2880}}' |
| 271 | f"{PARAM_END}\n" |
| 272 | f'{PARAM_START}patches" string="false">' |
| 273 | f'[{{"op":"replace","path":"/schedule","value":"quiet"}}]' |
| 274 | f"{PARAM_END}\n" |
| 275 | f"{INV_END}\n" |
| 276 | f"{FC_END}" |
| 277 | ) |
| 278 | result = parser.extract_tool_calls(model_output, None) |
| 279 | assert result.tools_called |
| 280 | args = json.loads(result.tool_calls[0].function.arguments) |
| 281 | assert args == { |
| 282 | "wait": {"type": "for", "minutes": 2880}, |
| 283 | "patches": [{"op": "replace", "path": "/schedule", "value": "quiet"}], |
| 284 | } |
| 285 | assert isinstance(args["wait"], dict) |
| 286 | assert isinstance(args["patches"], list) |
| 287 | |
| 288 | def test_string_attr_true_coerced_by_composed_schema(self): |
| 289 | """string="true" delivers a JSON string, but the engine's schema-aware |
| 290 | type fixer coerces it to the composed schema type (object).""" |
| 291 | tool = ChatCompletionToolsParam( |
| 292 | function=FunctionDefinition( |
| 293 | name="set_timer", |
| 294 | parameters={ |
| 295 | "type": "object", |
| 296 | "properties": { |
| 297 | "wait": { |
| 298 | "anyOf": [ |
| 299 | {"type": "object"}, |
| 300 | {"type": "null"}, |
| 301 | ], |
| 302 | }, |
| 303 | }, |
| 304 | }, |
| 305 | ), |
| 306 | ) |
| 307 | parser = make_parser(tools=[tool]) |
| 308 | model_output = ( |
| 309 | f"{FC_START}\n" |
| 310 | f'{INV_START}set_timer">\n' |
| 311 | f'{PARAM_START}wait" string="true">' |
| 312 | f'{{"type":"for","minutes":2880}}' |
| 313 | f"{PARAM_END}\n" |
| 314 | f"{INV_END}\n" |
| 315 | f"{FC_END}" |
| 316 | ) |
| 317 | result = parser.extract_tool_calls(model_output, None) |
| 318 | assert result.tools_called |
| 319 | args = json.loads(result.tool_calls[0].function.arguments) |
| 320 | assert args == {"wait": {"type": "for", "minutes": 2880}} |
| 321 | |
| 322 | def test_arguments_wrapper_repaired(self): |
| 323 | """A single 'arguments' wrapper parameter must be unwrapped when it |
| 324 | is not part of the tool schema and the inner object matches schema fields.""" |
| 325 | tool = ChatCompletionToolsParam( |
| 326 | function=FunctionDefinition( |
| 327 | name="get_weather", |
| 328 | parameters={ |
| 329 | "type": "object", |
| 330 | "properties": { |
| 331 | "location": {"type": "string"}, |
| 332 | }, |
| 333 | }, |
| 334 | ), |
| 335 | ) |
| 336 | parser = make_parser(tools=[tool]) |
| 337 | model_output = ( |
| 338 | f"{FC_START}\n" |
| 339 | f'{INV_START}get_weather">\n' |
| 340 | f'{PARAM_START}arguments" string="false">' |
| 341 | f'{{"location":"Beijing"}}' |
| 342 | f"{PARAM_END}\n" |
| 343 | f"{INV_END}\n" |
| 344 | f"{FC_END}" |
| 345 | ) |
| 346 | result = parser.extract_tool_calls(model_output, None) |
| 347 | assert result.tools_called |
| 348 | args = json.loads(result.tool_calls[0].function.arguments) |
| 349 | assert args == {"location": "Beijing"} |
| 350 | |
| 351 | def test_input_wrapper_repaired(self): |
| 352 | """A single 'input' wrapper parameter must be unwrapped similarly.""" |
| 353 | tool = ChatCompletionToolsParam( |
| 354 | function=FunctionDefinition( |
| 355 | name="get_weather", |
| 356 | parameters={ |
| 357 | "type": "object", |
| 358 | "properties": { |
| 359 | "location": {"type": "string"}, |
| 360 | }, |
| 361 | }, |
| 362 | ), |
| 363 | ) |
| 364 | parser = make_parser(tools=[tool]) |
| 365 | model_output = ( |
| 366 | f"{FC_START}\n" |
| 367 | f'{INV_START}get_weather">\n' |
| 368 | f'{PARAM_START}input" string="true">' |
| 369 | f'{{"location":"Beijing"}}' |
| 370 | f"{PARAM_END}\n" |
| 371 | f"{INV_END}\n" |
| 372 | f"{FC_END}" |
| 373 | ) |
| 374 | result = parser.extract_tool_calls(model_output, None) |
| 375 | assert result.tools_called |
| 376 | args = json.loads(result.tool_calls[0].function.arguments) |
| 377 | assert args == {"location": "Beijing"} |
| 378 | |
| 379 | def test_object_and_array_params(self): |
| 380 | """Object and array schema types must be JSON-parsed, not left as strings.""" |
| 381 | tool = ChatCompletionToolsParam( |
| 382 | function=FunctionDefinition( |
| 383 | name="update", |
| 384 | parameters={ |
| 385 | "type": "object", |
| 386 | "properties": { |
| 387 | "tags": {"type": "array"}, |
| 388 | "meta": {"type": "object"}, |
| 389 | }, |
| 390 | }, |
| 391 | ), |
| 392 | ) |
| 393 | parser = make_parser(tools=[tool]) |
| 394 | model_output = ( |
| 395 | f"{FC_START}\n" |
| 396 | f'{INV_START}update">\n' |
| 397 | f'{PARAM_START}tags" string="false">["a", "b"]{PARAM_END}\n' |
| 398 | f'{PARAM_START}meta" string="false">{{"k": 1}}{PARAM_END}\n' |
| 399 | f"{INV_END}\n" |
| 400 | f"{FC_END}" |
| 401 | ) |
| 402 | result = parser.extract_tool_calls(model_output, None) |
| 403 | assert result.tools_called |
| 404 | args = json.loads(result.tool_calls[0].function.arguments) |
| 405 | assert args["tags"] == ["a", "b"] |
| 406 | assert isinstance(args["tags"], list) |
| 407 | assert args["meta"] == {"k": 1} |
| 408 | assert isinstance(args["meta"], dict) |
| 409 | |
| 410 | def test_number_param(self): |
| 411 | """Number (float) schema type must be converted correctly.""" |
| 412 | tool = ChatCompletionToolsParam( |
| 413 | function=FunctionDefinition( |
| 414 | name="measure", |
| 415 | parameters={ |
| 416 | "type": "object", |
| 417 | "properties": { |
| 418 | "ratio": {"type": "number"}, |
| 419 | "whole": {"type": "number"}, |
| 420 | }, |
| 421 | }, |
| 422 | ), |
| 423 | ) |
| 424 | parser = make_parser(tools=[tool]) |
| 425 | model_output = ( |
| 426 | f"{FC_START}\n" |
| 427 | f'{INV_START}measure">\n' |
| 428 | f'{PARAM_START}ratio" string="false">3.14{PARAM_END}\n' |
| 429 | f'{PARAM_START}whole" string="false">5.0{PARAM_END}\n' |
| 430 | f"{INV_END}\n" |
| 431 | f"{FC_END}" |
| 432 | ) |
| 433 | result = parser.extract_tool_calls(model_output, None) |
| 434 | args = json.loads(result.tool_calls[0].function.arguments) |
| 435 | assert args["ratio"] == pytest.approx(3.14) |
| 436 | assert args["whole"] == 5 |
| 437 | assert isinstance(args["whole"], int) |
| 438 | |
| 439 | def test_multi_typed_schema(self): |
| 440 | """Schema with type: ["integer", "null"] must handle both cases.""" |
| 441 | tool = ChatCompletionToolsParam( |
| 442 | function=FunctionDefinition( |
| 443 | name="set_val", |
| 444 | parameters={ |
| 445 | "type": "object", |
| 446 | "properties": { |
| 447 | "count": {"type": ["integer", "null"]}, |
| 448 | "label": {"type": ["string", "null"]}, |
| 449 | }, |
| 450 | }, |
| 451 | ), |
| 452 | ) |
| 453 | parser = make_parser(tools=[tool]) |
| 454 | model_output = ( |
| 455 | f"{FC_START}\n" |
| 456 | f'{INV_START}set_val">\n' |
| 457 | f'{PARAM_START}count" string="false">42{PARAM_END}\n' |
| 458 | f'{PARAM_START}label" string="false">hello{PARAM_END}\n' |
| 459 | f"{INV_END}\n" |
| 460 | f"{FC_END}" |
| 461 | ) |
| 462 | result = parser.extract_tool_calls(model_output, None) |
| 463 | args = json.loads(result.tool_calls[0].function.arguments) |
| 464 | assert args["count"] == 42 |
| 465 | assert isinstance(args["count"], int) |
| 466 | assert args["label"] == "hello" |
| 467 | |
| 468 | def test_multi_typed_null_value(self): |
| 469 | """Literal 'null' must become None when the schema includes 'null'.""" |
| 470 | tool = ChatCompletionToolsParam( |
| 471 | function=FunctionDefinition( |
| 472 | name="clear", |
| 473 | parameters={ |
| 474 | "type": "object", |
| 475 | "properties": { |
| 476 | "value": {"type": ["integer", "null"]}, |
| 477 | }, |
| 478 | }, |
| 479 | ), |
| 480 | ) |
| 481 | parser = make_parser(tools=[tool]) |
| 482 | model_output = ( |
| 483 | f"{FC_START}\n" |
| 484 | f'{INV_START}clear">\n' |
| 485 | f'{PARAM_START}value" string="false">null{PARAM_END}\n' |
| 486 | f"{INV_END}\n" |
| 487 | f"{FC_END}" |
| 488 | ) |
| 489 | result = parser.extract_tool_calls(model_output, None) |
| 490 | args = json.loads(result.tool_calls[0].function.arguments) |
| 491 | assert args["value"] is None |
| 492 | |
| 493 | def test_null_coerced_back_to_string_by_schema(self): |
| 494 | """string="false" with 'null' is json-parsed to None, but the |
| 495 | engine's schema fixer coerces it back to "null" for string schemas.""" |
| 496 | tool = ChatCompletionToolsParam( |
| 497 | function=FunctionDefinition( |
| 498 | name="echo", |
| 499 | parameters={ |
| 500 | "type": "object", |
| 501 | "properties": { |
| 502 | "text": {"type": "string"}, |
| 503 | }, |
| 504 | }, |
| 505 | ), |
| 506 | ) |
| 507 | parser = make_parser(tools=[tool]) |
| 508 | model_output = ( |
| 509 | f"{FC_START}\n" |
| 510 | f'{INV_START}echo">\n' |
| 511 | f'{PARAM_START}text" string="false">null{PARAM_END}\n' |
| 512 | f"{INV_END}\n" |
| 513 | f"{FC_END}" |
| 514 | ) |
| 515 | result = parser.extract_tool_calls(model_output, None) |
| 516 | args = json.loads(result.tool_calls[0].function.arguments) |
| 517 | assert args["text"] == "null" |
| 518 | assert isinstance(args["text"], str) |
| 519 | |
| 520 | def test_no_schema_parses_json(self): |
| 521 | """Without a tool schema, string='false' params are JSON-parsed.""" |
| 522 | parser = make_parser(tools=None) |
| 523 | model_output = ( |
| 524 | f"{FC_START}\n" |
| 525 | f'{INV_START}unknown_fn">\n' |
| 526 | f'{PARAM_START}count" string="false">42{PARAM_END}\n' |
| 527 | f'{PARAM_START}flag" string="false">true{PARAM_END}\n' |
| 528 | f"{INV_END}\n" |
| 529 | f"{FC_END}" |
| 530 | ) |
| 531 | result = parser.extract_tool_calls(model_output, None) |
| 532 | assert result.tools_called |
| 533 | args = json.loads(result.tool_calls[0].function.arguments) |
| 534 | assert args["count"] == 42 |
| 535 | assert args["flag"] is True |
| 536 | |
| 537 | |
| 538 | # --------------------------------------------------------------------------- |
| 539 | # Tests: extract_tool_calls_streaming |
| 540 | # --------------------------------------------------------------------------- |
| 541 | |
| 542 | |
| 543 | class TestExtractToolCallsStreaming: |
| 544 | """Simulate character-by-character streaming and verify reconstructed args.""" |
| 545 | |
| 546 | @pytest.fixture |
| 547 | def parser(self): |
| 548 | return make_parser() |
| 549 | |
| 550 | def _stream(self, parser, full_text: str, request=None): |
| 551 | """Drive the parser line-by-line and collect non-None deltas. |
| 552 | |
| 553 | Real tokenizers emit multi-character chunks, not individual characters. |
| 554 | Streaming character-by-character would never deliver the full sentinel |
| 555 | token (e.g. '｜DSML｜') in a single delta, so we split on newlines to |
| 556 | ensure each sentinel always lands in one chunk. |
| 557 | """ |
| 558 | if request is None: |
| 559 | request = make_request() |
| 560 | # Split into lines, preserving the trailing newline in each chunk. |
| 561 | chunks: list[str] = [] |
| 562 | remaining = full_text |
| 563 | while remaining: |
| 564 | nl = remaining.find("\n") |
| 565 | if nl == -1: |
| 566 | chunks.append(remaining) |
| 567 | break |
| 568 | chunks.append(remaining[: nl + 1]) |
| 569 | remaining = remaining[nl + 1 :] |
| 570 | |
| 571 | deltas = [] |
| 572 | prev = "" |
| 573 | for chunk in chunks: |
| 574 | curr = prev + chunk |
| 575 | result = parser.extract_tool_calls_streaming( |
| 576 | previous_text=prev, |
| 577 | current_text=curr, |
| 578 | delta_text=chunk, |
| 579 | previous_token_ids=[], |
| 580 | current_token_ids=[], |
| 581 | delta_token_ids=[1], |
| 582 | request=request, |
| 583 | ) |
| 584 | prev = curr |
| 585 | if result is not None: |
| 586 | deltas.append(result) |
| 587 | return deltas |
| 588 | |
| 589 | def _reconstruct_args(self, deltas, tool_index=0) -> str: |
| 590 | """Concatenate all argument fragments for a given tool index.""" |
| 591 | fragments = [] |
| 592 | for d in deltas: |
| 593 | if d.tool_calls: |
| 594 | for tc in d.tool_calls: |
| 595 | if tc.index == tool_index and tc.function and tc.function.arguments: |
| 596 | fragments.append(tc.function.arguments) |
| 597 | return "".join(fragments) |
| 598 | |
| 599 | def test_plain_content_no_tool(self, parser): |
| 600 | full_text = "Hello, world!" |
| 601 | deltas = self._stream(parser, full_text) |
| 602 | content = "".join(d.content for d in deltas if d.content is not None) |
| 603 | assert "Hello, world!" in content |
| 604 | assert all(not d.tool_calls for d in deltas) |
| 605 | |
| 606 | def test_single_tool_streaming(self, parser): |
| 607 | full_text = build_tool_call("get_weather", {"location": "SF"}) |
| 608 | deltas = self._stream(parser, full_text) |
| 609 | args_str = self._reconstruct_args(deltas) |
| 610 | assert json.loads(args_str) == {"location": "SF"} |
| 611 | |
| 612 | def test_tool_name_emitted(self, parser): |
| 613 | full_text = build_tool_call("my_func", {"x": "1"}) |
| 614 | deltas = self._stream(parser, full_text) |
| 615 | func_names = [ |
| 616 | tc.function.name |
| 617 | for d in deltas |
| 618 | if d.tool_calls |
| 619 | for tc in d.tool_calls |
| 620 | if tc.function and tc.function.name |
| 621 | ] |
| 622 | assert any("my_func" in n for n in func_names) |
| 623 | |
| 624 | def test_content_before_tool_call_streaming(self, parser): |
| 625 | full_text = "Thinking... " + build_tool_call("fn", {"a": "b"}) |
| 626 | deltas = self._stream(parser, full_text) |
| 627 | content = "".join(d.content for d in deltas if d.content is not None) |
| 628 | assert "Thinking" in content |
| 629 | |
| 630 | def test_type_conversion_in_streaming(self): |
| 631 | tool = ChatCompletionToolsParam( |
| 632 | function=FunctionDefinition( |
| 633 | name="add", |
| 634 | parameters={ |
| 635 | "type": "object", |
| 636 | "properties": { |
| 637 | "x": {"type": "integer"}, |
| 638 | "y": {"type": "integer"}, |
| 639 | }, |
| 640 | }, |
| 641 | ), |
| 642 | ) |
| 643 | parser = make_parser(tools=[tool]) |
| 644 | full_text = ( |
| 645 | f"{FC_START}\n" |
| 646 | f'{INV_START}add">\n' |
| 647 | f'{PARAM_START}x" string="false">3{PARAM_END}\n' |
| 648 | f'{PARAM_START}y" string="false">4{PARAM_END}\n' |
| 649 | f"{INV_END}\n" |
| 650 | f"{FC_END}" |
| 651 | ) |
| 652 | deltas = self._stream(parser, full_text) |
| 653 | args_str = self._reconstruct_args(deltas) |
| 654 | assert json.loads(args_str) == {"x": 3, "y": 4} |
| 655 | |
| 656 | def test_string_attr_true_coerced_by_schema_streaming(self): |
| 657 | """Streaming: string='true' delivers a string but the engine's |
| 658 | schema fixer coerces it to the schema type (integer).""" |
| 659 | tool = ChatCompletionToolsParam( |
| 660 | function=FunctionDefinition( |
| 661 | name="score", |
| 662 | parameters={ |
| 663 | "type": "object", |
| 664 | "properties": { |
| 665 | "value": {"type": "integer"}, |
| 666 | }, |
| 667 | }, |
| 668 | ), |
| 669 | ) |
| 670 | parser = make_parser(tools=[tool]) |
| 671 | full_text = ( |
| 672 | f"{FC_START}\n" |
| 673 | f'{INV_START}score">\n' |
| 674 | f'{PARAM_START}value" string="true">42{PARAM_END}\n' |
| 675 | f"{INV_END}\n" |
| 676 | f"{FC_END}" |
| 677 | ) |
| 678 | deltas = self._stream(parser, full_text) |
| 679 | args_str = self._reconstruct_args(deltas) |
| 680 | args = json.loads(args_str) |
| 681 | assert args == {"value": 42} |
| 682 | assert isinstance(args["value"], int) |
| 683 | |
| 684 | def test_composed_schema_conversion_in_streaming(self): |
| 685 | tool = ChatCompletionToolsParam( |
| 686 | function=FunctionDefinition( |
| 687 | name="set_timer", |
| 688 | parameters={ |
| 689 | "type": "object", |
| 690 | "properties": { |
| 691 | "wait": { |
| 692 | "anyOf": [ |
| 693 | {"type": "object"}, |
| 694 | {"type": "null"}, |
| 695 | ], |
| 696 | }, |
| 697 | "patches": { |
| 698 | "oneOf": [ |
| 699 | {"type": "array", "items": {"type": "object"}}, |
| 700 | {"type": "null"}, |
| 701 | ], |
| 702 | }, |
| 703 | }, |
| 704 | }, |
| 705 | ), |
| 706 | ) |
| 707 | parser = make_parser(tools=[tool]) |
| 708 | full_text = ( |
| 709 | f"{FC_START}\n" |
| 710 | f'{INV_START}set_timer">\n' |
| 711 | f'{PARAM_START}wait" string="false">' |
| 712 | f'{{"type":"for","minutes":2880}}' |
| 713 | f"{PARAM_END}\n" |
| 714 | f'{PARAM_START}patches" string="false">' |
| 715 | f'[{{"op":"replace","path":"/schedule","value":"quiet"}}]' |
| 716 | f"{PARAM_END}\n" |
| 717 | f"{INV_END}\n" |
| 718 | f"{FC_END}" |
| 719 | ) |
| 720 | deltas = self._stream(parser, full_text) |
| 721 | args = json.loads(self._reconstruct_args(deltas)) |
| 722 | assert args == { |
| 723 | "wait": {"type": "for", "minutes": 2880}, |
| 724 | "patches": [{"op": "replace", "path": "/schedule", "value": "quiet"}], |
| 725 | } |
| 726 | |
| 727 | def test_responses_function_tool_schema_in_streaming(self): |
| 728 | """Responses API FunctionTool schemas must drive streaming conversion.""" |
| 729 | tool = FunctionTool( |
| 730 | type="function", |
| 731 | name="toggle", |
| 732 | parameters={ |
| 733 | "type": "object", |
| 734 | "properties": { |
| 735 | "enabled": {"type": "boolean"}, |
| 736 | "count": {"type": "integer"}, |
| 737 | }, |
| 738 | }, |
| 739 | ) |
| 740 | parser = make_parser(tools=[tool]) |
| 741 | full_text = ( |
| 742 | f"{FC_START}\n" |
| 743 | f'{INV_START}toggle">\n' |
| 744 | f'{PARAM_START}enabled" string="false">true{PARAM_END}\n' |
| 745 | f'{PARAM_START}count" string="false">42{PARAM_END}\n' |
| 746 | f"{INV_END}\n" |
| 747 | f"{FC_END}" |
| 748 | ) |
| 749 | |
| 750 | deltas = self._stream(parser, full_text) |
| 751 | args = json.loads(self._reconstruct_args(deltas)) |
| 752 | |
| 753 | assert args == {"enabled": True, "count": 42} |
| 754 | assert isinstance(args["enabled"], bool) |
| 755 | assert isinstance(args["count"], int) |
| 756 | |
| 757 | def test_streaming_matches_non_streaming_conversion_fallbacks(self): |
| 758 | """Streaming must reuse conversion fallback semantics for string=false.""" |
| 759 | tool = ChatCompletionToolsParam( |
| 760 | function=FunctionDefinition( |
| 761 | name="coerce", |
| 762 | parameters={ |
| 763 | "type": "object", |
| 764 | "properties": { |
| 765 | "union_value": {"type": ["null", "string"]}, |
| 766 | "bad_int": {"type": "integer"}, |
| 767 | "nullable_string": {"type": ["null", "string"]}, |
| 768 | "null_string": {"type": "string"}, |
| 769 | "whole_number": {"type": "number"}, |
| 770 | }, |
| 771 | }, |
| 772 | ), |
| 773 | ) |
| 774 | parser = make_parser(tools=[tool]) |
| 775 | full_text = ( |
| 776 | f"{FC_START}\n" |
| 777 | f'{INV_START}coerce">\n' |
| 778 | f'{PARAM_START}union_value" string="false">hello{PARAM_END}\n' |
| 779 | f'{PARAM_START}bad_int" string="false">abc{PARAM_END}\n' |
| 780 | f'{PARAM_START}nullable_string" string="false">null{PARAM_END}\n' |
| 781 | f'{PARAM_START}null_string" string="false">null{PARAM_END}\n' |
| 782 | f'{PARAM_START}whole_number" string="false">3.0{PARAM_END}\n' |
| 783 | f"{INV_END}\n" |
| 784 | f"{FC_END}" |
| 785 | ) |
| 786 | |
| 787 | non_stream = parser.extract_tool_calls(full_text, None) |
| 788 | non_stream_args = json.loads(non_stream.tool_calls[0].function.arguments) |
| 789 | |
| 790 | deltas = self._stream(parser, full_text) |
| 791 | stream_args = json.loads(self._reconstruct_args(deltas)) |
| 792 | |
| 793 | assert stream_args == non_stream_args |
| 794 | assert stream_args == { |
| 795 | "union_value": "hello", |
| 796 | "bad_int": "abc", |
| 797 | "nullable_string": None, |
| 798 | "null_string": "null", |
| 799 | "whole_number": 3, |
| 800 | } |
| 801 | |
| 802 | def test_multiple_tools_streaming(self, parser): |
| 803 | full_text = ( |
| 804 | f"{FC_START}\n" |
| 805 | f'{INV_START}func_a">\n' |
| 806 | f'{PARAM_START}p" string="true">v1{PARAM_END}\n' |
| 807 | f"{INV_END}\n" |
| 808 | f'{INV_START}func_b">\n' |
| 809 | f'{PARAM_START}q" string="true">v2{PARAM_END}\n' |
| 810 | f"{INV_END}\n" |
| 811 | f"{FC_END}" |
| 812 | ) |
| 813 | deltas = self._stream(parser, full_text) |
| 814 | |
| 815 | # Collect function names by index |
| 816 | names_by_index: dict[int, str] = {} |
| 817 | for d in deltas: |
| 818 | if d.tool_calls: |
| 819 | for tc in d.tool_calls: |
| 820 | if tc.function and tc.function.name: |
| 821 | names_by_index[tc.index] = tc.function.name |
| 822 | |
| 823 | assert names_by_index.get(0) == "func_a" |
| 824 | assert names_by_index.get(1) == "func_b" |
| 825 | |
| 826 | assert json.loads(self._reconstruct_args(deltas, tool_index=0)) == {"p": "v1"} |
| 827 | assert json.loads(self._reconstruct_args(deltas, tool_index=1)) == {"q": "v2"} |
| 828 | |
| 829 | def test_state_reset_on_new_stream(self): |
| 830 | """A fresh parser instance must produce identical results.""" |
| 831 | full_text = build_tool_call("fn", {"k": "v"}) |
| 832 | # First stream |
| 833 | self._stream(make_parser(), full_text) |
| 834 | # Second stream with fresh parser |
| 835 | deltas2 = self._stream(make_parser(), full_text) |
| 836 | assert json.loads(self._reconstruct_args(deltas2)) == {"k": "v"} |
| 837 | |
| 838 | def test_empty_arguments_streaming(self, parser): |
| 839 | """Invoke block with zero parameters should produce empty JSON.""" |
| 840 | full_text = f'{FC_START}\n{INV_START}get_time">\n{INV_END}\n{FC_END}' |
| 841 | deltas = self._stream(parser, full_text) |
| 842 | args_str = self._reconstruct_args(deltas) |
| 843 | assert json.loads(args_str) == {} |
| 844 | |
| 845 | def test_unique_tool_call_ids(self, parser): |
| 846 | """Each tool call in a parallel stream must get a distinct id.""" |
| 847 | full_text = ( |
| 848 | f"{FC_START}\n" |
| 849 | f'{INV_START}fn_a">\n' |
| 850 | f'{PARAM_START}x" string="true">1{PARAM_END}\n' |
| 851 | f"{INV_END}\n" |
| 852 | f'{INV_START}fn_b">\n' |
| 853 | f'{PARAM_START}y" string="true">2{PARAM_END}\n' |
| 854 | f"{INV_END}\n" |
| 855 | f"{FC_END}" |
| 856 | ) |
| 857 | deltas = self._stream(parser, full_text) |
| 858 | ids = [ |
| 859 | tc.id |
| 860 | for d in deltas |
| 861 | if d.tool_calls |
| 862 | for tc in d.tool_calls |
| 863 | if tc.id is not None |
| 864 | ] |
| 865 | assert len(ids) == 2 |
| 866 | assert ids[0] != ids[1] |
| 867 | |
| 868 | def test_streaming_matches_non_streaming(self, parser): |
| 869 | """Streaming and non-streaming must produce the same result.""" |
| 870 | full_text = build_tool_call( |
| 871 | "get_weather", {"location": "SF", "date": "2024-01-16"} |
| 872 | ) |
| 873 | # Non-streaming |
| 874 | non_stream = parser.extract_tool_calls(full_text, None) |
| 875 | assert non_stream.tools_called |
| 876 | ns_name = non_stream.tool_calls[0].function.name |
| 877 | ns_args = json.loads(non_stream.tool_calls[0].function.arguments) |
| 878 | # Streaming |
| 879 | deltas = self._stream(parser, full_text) |
| 880 | s_names = [ |
| 881 | tc.function.name |
| 882 | for d in deltas |
| 883 | if d.tool_calls |
| 884 | for tc in d.tool_calls |
| 885 | if tc.function and tc.function.name |
| 886 | ] |
| 887 | s_args = json.loads(self._reconstruct_args(deltas)) |
| 888 | assert s_names[0] == ns_name |
| 889 | assert s_args == ns_args |
| 890 | |
| 891 | def _stream_chunked(self, parser, full_text: str, chunk_size: int, request=None): |
| 892 | """Drive the parser with fixed-size chunks (simulates stream interval). |
| 893 | |
| 894 | Unlike ``_stream`` which splits on newlines, this splits the text |
| 895 | into ``chunk_size``-character pieces so the start token can be |
| 896 | split across chunks — exactly what happens with stream interval > 1. |
| 897 | """ |
| 898 | if request is None: |
| 899 | request = make_request() |
| 900 | chunks = [ |
| 901 | full_text[i : i + chunk_size] for i in range(0, len(full_text), chunk_size) |
| 902 | ] |
| 903 | deltas = [] |
| 904 | prev = "" |
| 905 | for chunk in chunks: |
| 906 | curr = prev + chunk |
| 907 | result = parser.extract_tool_calls_streaming( |
| 908 | previous_text=prev, |
| 909 | current_text=curr, |
| 910 | delta_text=chunk, |
| 911 | previous_token_ids=[], |
| 912 | current_token_ids=[], |
| 913 | delta_token_ids=[1], |
| 914 | request=request, |
| 915 | ) |
| 916 | prev = curr |
| 917 | if result is not None: |
| 918 | deltas.append(result) |
| 919 | return deltas |
| 920 | |
| 921 | def test_single_tool_chunked_stream_interval(self, parser): |
| 922 | """Start token split across chunks (stream interval > 1).""" |
| 923 | full_text = build_tool_call("get_weather", {"location": "SF"}) |
| 924 | # Use a chunk size that splits the start token |
| 925 | deltas = self._stream_chunked(parser, full_text, chunk_size=5) |
| 926 | args_str = self._reconstruct_args(deltas) |
| 927 | assert json.loads(args_str) == {"location": "SF"} |
| 928 | |
| 929 | def test_content_before_tool_chunked(self, parser): |
| 930 | """Content before tool call with chunked streaming.""" |
| 931 | full_text = "Thinking... " + build_tool_call("fn", {"a": "b"}) |
| 932 | deltas = self._stream_chunked(parser, full_text, chunk_size=7) |
| 933 | content = "".join(d.content for d in deltas if d.content is not None) |
| 934 | assert "Thinking" in content |
| 935 | args_str = self._reconstruct_args(deltas) |
| 936 | assert json.loads(args_str) == {"a": "b"} |
| 937 | |
| 938 | def test_multiple_tools_chunked(self, parser): |
| 939 | """Multiple tools with chunked streaming.""" |
| 940 | full_text = ( |
| 941 | f"{FC_START}\n" |
| 942 | f'{INV_START}func_a">\n' |
| 943 | f'{PARAM_START}p" string="true">v1{PARAM_END}\n' |
| 944 | f"{INV_END}\n" |
| 945 | f'{INV_START}func_b">\n' |
| 946 | f'{PARAM_START}q" string="true">v2{PARAM_END}\n' |
| 947 | f"{INV_END}\n" |
| 948 | f"{FC_END}" |
| 949 | ) |
| 950 | deltas = self._stream_chunked(parser, full_text, chunk_size=10) |
| 951 | assert json.loads(self._reconstruct_args(deltas, tool_index=0)) == {"p": "v1"} |
| 952 | assert json.loads(self._reconstruct_args(deltas, tool_index=1)) == {"q": "v2"} |
| 953 | |
| 954 | def test_emits_arguments_before_invoke_completes(self, parser): |
| 955 | """Argument deltas should stream before the invoke block closes.""" |
| 956 | partial_text = ( |
| 957 | f"{FC_START}\n" |
| 958 | f'{INV_START}fn">\n' |
| 959 | f'{PARAM_START}k" string="true">val{PARAM_END}\n' |
| 960 | ) |
| 961 | deltas = self._stream(parser, partial_text) |
| 962 | arg_chunks = [ |
| 963 | tc.function.arguments |
| 964 | for delta in deltas |
| 965 | for tc in delta.tool_calls or [] |
| 966 | if tc.function and tc.function.arguments is not None |
| 967 | ] |
| 968 | combined = "".join(arg_chunks) |
| 969 | assert combined # some partial args emitted |
| 970 | assert combined.startswith('{"k"') |
| 971 | |
| 972 | def test_no_marker_leak_chunked(self, parser): |
| 973 | """Chunked streaming must NOT leak DSML start-marker fragments |
| 974 | as content (GitHub #40801).""" |
| 975 | full_text = build_tool_call("fn", {"k": "v"}) |
| 976 | deltas = self._stream_chunked(parser, full_text, chunk_size=5) |
| 977 | content = "".join(d.content for d in deltas if d.content is not None) |
| 978 | assert content == "" |
| 979 | args_str = self._reconstruct_args(deltas) |
| 980 | assert json.loads(args_str) == {"k": "v"} |
| 981 | |
| 982 | def test_no_marker_leak_with_prefix_chunked(self, parser): |
| 983 | """Content before a tool call must not include start-marker |
| 984 | fragments when chunked (GitHub #40801).""" |
| 985 | full_text = "Hello!" + build_tool_call("fn", {"a": "b"}) |
| 986 | deltas = self._stream_chunked(parser, full_text, chunk_size=5) |
| 987 | content = "".join(d.content for d in deltas if d.content is not None) |
| 988 | assert content == "Hello!" |
| 989 | assert "DSML" not in content |
| 990 | assert "<｜" not in content |
| 991 | args_str = self._reconstruct_args(deltas) |
| 992 | assert json.loads(args_str) == {"a": "b"} |
| 993 | |
| 994 | def test_no_marker_leak_char_by_char(self, parser): |
| 995 | """Character-by-character streaming must not leak marker |
| 996 | fragments (GitHub #40801).""" |
| 997 | full_text = build_tool_call("fn", {"k": "v"}) |
| 998 | deltas = self._stream_chunked(parser, full_text, chunk_size=1) |
| 999 | content = "".join(d.content for d in deltas if d.content is not None) |
| 1000 | assert content == "" |
| 1001 | args_str = self._reconstruct_args(deltas) |
| 1002 | assert json.loads(args_str) == {"k": "v"} |
| 1003 | |
| 1004 | def test_no_marker_leak_all_split_points(self, parser): |
| 1005 | """Start token split at every possible boundary must not |
| 1006 | leak (GitHub #40801).""" |
| 1007 | for chunk_size in range(1, len(FC_START) + 2): |
| 1008 | p = make_parser() |
| 1009 | full_text = build_tool_call("fn", {"k": "v"}) |
| 1010 | deltas = self._stream_chunked(p, full_text, chunk_size=chunk_size) |
| 1011 | content = "".join(d.content for d in deltas if d.content is not None) |
| 1012 | assert content == "", ( |
| 1013 | f"Leaked content {content!r} at chunk_size={chunk_size}" |
| 1014 | ) |
| 1015 | |
| 1016 | def test_false_partial_marker_emitted(self, parser): |
| 1017 | """Text ending with a prefix of the start token that turns out |
| 1018 | NOT to be a marker must still be emitted as content.""" |
| 1019 | full_text = "<｜DSM some regular text" |
| 1020 | deltas = self._stream_chunked(parser, full_text, chunk_size=3) |
| 1021 | content = "".join(d.content for d in deltas if d.content is not None) |
| 1022 | assert content == full_text |
| 1023 | |
| 1024 | def test_object_and_array_params_streaming(self): |
| 1025 | """Streaming: object/array params must be JSON-parsed.""" |
| 1026 | tool = ChatCompletionToolsParam( |
| 1027 | function=FunctionDefinition( |
| 1028 | name="update", |
| 1029 | parameters={ |
| 1030 | "type": "object", |
| 1031 | "properties": { |
| 1032 | "tags": {"type": "array"}, |
| 1033 | "meta": {"type": "object"}, |
| 1034 | }, |
| 1035 | }, |
| 1036 | ), |
| 1037 | ) |
| 1038 | parser = make_parser(tools=[tool]) |
| 1039 | full_text = ( |
| 1040 | f"{FC_START}\n" |
| 1041 | f'{INV_START}update">\n' |
| 1042 | f'{PARAM_START}tags" string="false">["a", "b"]{PARAM_END}\n' |
| 1043 | f'{PARAM_START}meta" string="false">{{"k": 1}}{PARAM_END}\n' |
| 1044 | f"{INV_END}\n" |
| 1045 | f"{FC_END}" |
| 1046 | ) |
| 1047 | deltas = self._stream(parser, full_text) |
| 1048 | args = json.loads(self._reconstruct_args(deltas)) |
| 1049 | assert args["tags"] == ["a", "b"] |
| 1050 | assert args["meta"] == {"k": 1} |
| 1051 | |
| 1052 | def test_multi_typed_schema_streaming(self): |
| 1053 | """Streaming: type: ["integer", "null"] must coerce correctly.""" |
| 1054 | tool = ChatCompletionToolsParam( |
| 1055 | function=FunctionDefinition( |
| 1056 | name="set_val", |
| 1057 | parameters={ |
| 1058 | "type": "object", |
| 1059 | "properties": { |
| 1060 | "count": {"type": ["integer", "null"]}, |
| 1061 | }, |
| 1062 | }, |
| 1063 | ), |
| 1064 | ) |
| 1065 | parser = make_parser(tools=[tool]) |
| 1066 | full_text = ( |
| 1067 | f"{FC_START}\n" |
| 1068 | f'{INV_START}set_val">\n' |
| 1069 | f'{PARAM_START}count" string="false">42{PARAM_END}\n' |
| 1070 | f"{INV_END}\n" |
| 1071 | f"{FC_END}" |
| 1072 | ) |
| 1073 | deltas = self._stream(parser, full_text) |
| 1074 | args = json.loads(self._reconstruct_args(deltas)) |
| 1075 | assert args["count"] == 42 |
| 1076 | assert isinstance(args["count"], int) |
| 1077 | |
| 1078 | def test_multi_typed_null_streaming(self): |
| 1079 | """Streaming: 'null' with ["integer", "null"] schema must become None.""" |
| 1080 | tool = ChatCompletionToolsParam( |
| 1081 | function=FunctionDefinition( |
| 1082 | name="clear", |
| 1083 | parameters={ |
| 1084 | "type": "object", |
| 1085 | "properties": { |
| 1086 | "value": {"type": ["integer", "null"]}, |
| 1087 | }, |
| 1088 | }, |
| 1089 | ), |
| 1090 | ) |
| 1091 | parser = make_parser(tools=[tool]) |
| 1092 | full_text = ( |
| 1093 | f"{FC_START}\n" |
| 1094 | f'{INV_START}clear">\n' |
| 1095 | f'{PARAM_START}value" string="false">null{PARAM_END}\n' |
| 1096 | f"{INV_END}\n" |
| 1097 | f"{FC_END}" |
| 1098 | ) |
| 1099 | deltas = self._stream(parser, full_text) |
| 1100 | args = json.loads(self._reconstruct_args(deltas)) |
| 1101 | assert args["value"] is None |
| 1102 | |
| 1103 | def test_number_param_streaming(self): |
| 1104 | """Streaming: number type must be converted.""" |
| 1105 | tool = ChatCompletionToolsParam( |
| 1106 | function=FunctionDefinition( |
| 1107 | name="measure", |
| 1108 | parameters={ |
| 1109 | "type": "object", |
| 1110 | "properties": { |
| 1111 | "ratio": {"type": "number"}, |
| 1112 | }, |
| 1113 | }, |
| 1114 | ), |
| 1115 | ) |
| 1116 | parser = make_parser(tools=[tool]) |
| 1117 | full_text = ( |
| 1118 | f"{FC_START}\n" |
| 1119 | f'{INV_START}measure">\n' |
| 1120 | f'{PARAM_START}ratio" string="false">3.14{PARAM_END}\n' |
| 1121 | f"{INV_END}\n" |
| 1122 | f"{FC_END}" |
| 1123 | ) |
| 1124 | deltas = self._stream(parser, full_text) |
| 1125 | args = json.loads(self._reconstruct_args(deltas)) |
| 1126 | assert args["ratio"] == pytest.approx(3.14) |
| 1127 | |
| 1128 | |
| 1129 | class TestDelimiterPreservation: |
| 1130 | """Regression: fast detokenization skipping DSML delimiters (PR #33964).""" |
| 1131 | |
| 1132 | @pytest.fixture |
| 1133 | def parser(self): |
| 1134 | return make_parser() |
| 1135 | |
| 1136 | def test_delimiter_preserved_fast_detokenization(self, parser): |
| 1137 | """DSML delimiters as literal text must still be detected.""" |
| 1138 | # Delimiters appear as regular text (fast detokenization scenario). |
| 1139 | model_output = ( |
| 1140 | f"{FC_START}\n" |
| 1141 | f'{INV_START}get_weather">\n' |
| 1142 | f'{PARAM_START}location" string="true">Tokyo{PARAM_END}\n' |
| 1143 | f"{INV_END}\n" |
| 1144 | f"{FC_END}" |
| 1145 | ) |
| 1146 | |
| 1147 | # Non-streaming: parser must detect the tool call |
| 1148 | result = parser.extract_tool_calls(model_output, None) |
| 1149 | assert result.tools_called |
| 1150 | assert len(result.tool_calls) == 1 |
| 1151 | assert result.tool_calls[0].function.name == "get_weather" |
| 1152 | assert json.loads(result.tool_calls[0].function.arguments) == { |
| 1153 | "location": "Tokyo" |
| 1154 | } |
| 1155 | |
| 1156 | assert result.content is None |
| 1157 | |
| 1158 | # With content prefix |
| 1159 | prefixed_output = "Here is the weather: " + model_output |
| 1160 | result2 = parser.extract_tool_calls(prefixed_output, None) |
| 1161 | assert result2.tools_called |
| 1162 | assert result2.content == "Here is the weather: " |
| 1163 | |
| 1164 | def test_tool_detection_skip_special_tokens_false(self, parser): |
| 1165 | """Regression: skip_special_tokens must be False when tools are enabled.""" |
| 1166 | # adjust_request must set skip_special_tokens=False |
| 1167 | tool = make_tool_param( |
| 1168 | "search", |
| 1169 | { |
| 1170 | "type": "object", |
| 1171 | "properties": { |
| 1172 | "query": {"type": "string"}, |
| 1173 | }, |
| 1174 | }, |
| 1175 | ) |
| 1176 | request = make_request(tools=[tool]) |
| 1177 | request.tool_choice = "auto" |
| 1178 | adjusted = parser.adjust_request(request) |
| 1179 | assert adjusted.skip_special_tokens is False |
| 1180 | |
| 1181 | full_text = build_tool_call("search", {"query": "vllm documentation"}) |
| 1182 | |
| 1183 | # Non-streaming extraction |
| 1184 | non_stream_result = parser.extract_tool_calls(full_text, request) |
| 1185 | assert non_stream_result.tools_called |
| 1186 | assert len(non_stream_result.tool_calls) == 1 |
| 1187 | assert non_stream_result.tool_calls[0].function.name == "search" |
| 1188 | ns_args = json.loads(non_stream_result.tool_calls[0].function.arguments) |
| 1189 | assert ns_args == {"query": "vllm documentation"} |
| 1190 | |
| 1191 | # Streaming extraction: drive the parser line-by-line |
| 1192 | chunks: list[str] = [] |
| 1193 | remaining = full_text |
| 1194 | while remaining: |
| 1195 | nl = remaining.find("\n") |
| 1196 | if nl == -1: |
| 1197 | chunks.append(remaining) |
| 1198 | break |
| 1199 | chunks.append(remaining[: nl + 1]) |
| 1200 | remaining = remaining[nl + 1 :] |
| 1201 | |
| 1202 | reconstructor = run_tool_extraction_streaming( |
| 1203 | parser, chunks, request, assert_one_tool_per_delta=False |
| 1204 | ) |
| 1205 | assert len(reconstructor.tool_calls) == 1 |
| 1206 | assert reconstructor.tool_calls[0].function.name == "search" |
| 1207 | streamed_args = json.loads(reconstructor.tool_calls[0].function.arguments) |
| 1208 | assert streamed_args == ns_args |