| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import contextlib |
| 5 | import json |
| 6 | from abc import abstractmethod |
| 7 | from collections.abc import Sequence |
| 8 | from dataclasses import dataclass, field, replace |
| 9 | |
| 10 | from openai.types.responses import ToolChoiceFunction |
| 11 | from pydantic import TypeAdapter, ValidationError |
| 12 | from xgrammar import Grammar, StructuralTag |
| 13 | from xgrammar.structural_tag import ( |
| 14 | ConstStringFormat, |
| 15 | Format, |
| 16 | GrammarFormat, |
| 17 | JSONSchemaFormat, |
| 18 | OrFormat, |
| 19 | RegexFormat, |
| 20 | TagFormat, |
| 21 | TriggeredTagsFormat, |
| 22 | ) |
| 23 | |
| 24 | from vllm.entrypoints.chat_utils import ( |
| 25 | get_tool_call_id_type, |
| 26 | make_tool_call_id, |
| 27 | ) |
| 28 | from vllm.entrypoints.generate.base.protocol import ( |
| 29 | DeltaMessage, |
| 30 | ExtractedToolCallInformation, |
| 31 | FunctionCall, |
| 32 | FunctionDefinition, |
| 33 | ) |
| 34 | from vllm.entrypoints.openai.chat_completion.protocol import ( |
| 35 | ChatCompletionNamedToolChoiceParam, |
| 36 | ChatCompletionRequest, |
| 37 | ) |
| 38 | from vllm.entrypoints.openai.responses.protocol import ResponsesRequest |
| 39 | from vllm.exceptions import VLLMValidationError |
| 40 | from vllm.logger import init_logger |
| 41 | from vllm.parser.engine.adapters import ParserEngineToolAdapter |
| 42 | from vllm.parser.metrics import record_tool_parser_invocation |
| 43 | from vllm.parser.utils import count_history_tool_calls |
| 44 | from vllm.reasoning.abs_reasoning_parsers import ReasoningParser |
| 45 | from vllm.sampling_params import SamplingParams, StructuredOutputsParams |
| 46 | from vllm.tokenizers import TokenizerLike |
| 47 | from vllm.tool_parsers.abstract_tool_parser import Tool, ToolParser |
| 48 | from vllm.tool_parsers.streaming import ( |
| 49 | extract_named_tool_call_streaming, |
| 50 | extract_required_tool_call_streaming, |
| 51 | ) |
| 52 | from vllm.tool_parsers.structural_tag_registry import ( |
| 53 | limit_to_single_tool_call, |
| 54 | resolve_tool_strictness, |
| 55 | ) |
| 56 | from vllm.tool_parsers.tool_strict_level import ToolStrictLevel |
| 57 | |
| 58 | logger = init_logger(__name__) |
| 59 | |
| 60 | |
| 61 | @dataclass |
| 62 | class StreamState: |
| 63 | """Mutable state for ``Parser.parse_delta()``. One per stream.""" |
| 64 | |
| 65 | reasoning_ended: bool = False |
| 66 | tool_call_text_started: bool = False |
| 67 | prompt_reasoning_checked: bool = False |
| 68 | previous_text: str = "" |
| 69 | previous_token_ids: list[int] = field(default_factory=list) |
| 70 | history_tool_call_cnt: int = 0 |
| 71 | history_tool_call_cnt_initialized: bool = False |
| 72 | tool_call_id_type: str = "random" |
| 73 | # only used for "required" and "named tool" choices, |
| 74 | # tracks whether function name has been fully returned in the stream yet |
| 75 | function_name_returned: bool = False |
| 76 | engine_based: bool = False |
| 77 | |
| 78 | def advance( |
| 79 | self, |
| 80 | delta_text: str, |
| 81 | delta_token_ids: list[int], |
| 82 | ) -> tuple[str, list[int]]: |
| 83 | if self.engine_based: |
| 84 | return delta_text, delta_token_ids |
| 85 | return ( |
| 86 | self.previous_text + delta_text, |
| 87 | self.previous_token_ids + delta_token_ids, |
| 88 | ) |
| 89 | |
| 90 | def commit( |
| 91 | self, |
| 92 | current_text: str, |
| 93 | current_token_ids: list[int], |
| 94 | ) -> None: |
| 95 | if self.engine_based: |
| 96 | self.previous_text = "" |
| 97 | self.previous_token_ids = [] |
| 98 | else: |
| 99 | self.previous_text = current_text |
| 100 | self.previous_token_ids = current_token_ids |
| 101 | |
| 102 | |
| 103 | class Parser: |
| 104 | """Parse model output into reasoning, content and tool calls. |
| 105 | |
| 106 | The serving layer holds one ``Parser`` per request and calls only the |
| 107 | members defined here. ``ParserEngine`` implements them over a single |
| 108 | declarative engine; ``DelegatingParser`` composes a legacy |
| 109 | ``ReasoningParser`` / ``ToolParser`` pair. |
| 110 | """ |
| 111 | |
| 112 | # Class-level parser classes for compatibility with existing patterns |
| 113 | # Subclasses should override these if they use specific parser classes |
| 114 | reasoning_parser_cls: type[ReasoningParser] | None = None |
| 115 | tool_parser_cls: type[ToolParser] | None = None |
| 116 | # Server-side floor for tool-call structural tags (--tool-strict-level). |
| 117 | tool_strict_level: ToolStrictLevel = ToolStrictLevel.AUTO |
| 118 | always_adjust_request: bool = False |
| 119 | |
| 120 | def __init__( |
| 121 | self, |
| 122 | tokenizer: TokenizerLike, |
| 123 | tools: list[Tool] | None = None, |
| 124 | *args, |
| 125 | model_config=None, |
| 126 | **kwargs, |
| 127 | ): |
| 128 | self.model_tokenizer = tokenizer |
| 129 | self._reasoning_parser: ReasoningParser | None = None |
| 130 | self._tool_parser: ToolParser | None = None |
| 131 | if self.__class__.reasoning_parser_cls is not None: |
| 132 | self._reasoning_parser = self.__class__.reasoning_parser_cls( |
| 133 | tokenizer, *args, model_config=model_config, **kwargs |
| 134 | ) |
| 135 | if self.__class__.tool_parser_cls is not None: |
| 136 | # Engine-based adapters take the same construction kwargs as |
| 137 | # the reasoning parser (e.g. chat_template_kwargs for per-request |
| 138 | # thinking toggles); legacy ToolParser classes take none. |
| 139 | if issubclass(self.__class__.tool_parser_cls, ParserEngineToolAdapter): |
| 140 | self._tool_parser = self.__class__.tool_parser_cls( |
| 141 | tokenizer, tools, model_config=model_config, **kwargs |
| 142 | ) |
| 143 | else: |
| 144 | self._tool_parser = self.__class__.tool_parser_cls(tokenizer, tools) |
| 145 | |
| 146 | self._engine_based = ( |
| 147 | self._reasoning_parser is None |
| 148 | or self._reasoning_parser.engine_based_streaming |
| 149 | ) and (self._tool_parser is None or self._tool_parser.engine_based_streaming) |
| 150 | if ( |
| 151 | self._reasoning_parser is None |
| 152 | and self._tool_parser is not None |
| 153 | and hasattr(self._tool_parser, "skip_reasoning_parsing") |
| 154 | ): |
| 155 | # With no reasoning parser configured, reasoning markup is |
| 156 | # plain content: an engine-based tool parser should pass it |
| 157 | # through verbatim where its grammar allows, not consume or |
| 158 | # reclassify it. The engine ignores the flag for markers |
| 159 | # shared with non-reasoning structure. |
| 160 | self._tool_parser.skip_reasoning_parsing = True |
| 161 | self._stream_state = StreamState( |
| 162 | tool_call_id_type=( |
| 163 | get_tool_call_id_type(model_config) |
| 164 | if model_config is not None |
| 165 | else "random" |
| 166 | ), |
| 167 | engine_based=self._engine_based, |
| 168 | ) |
| 169 | |
| 170 | @property |
| 171 | def reasoning_parser(self) -> ReasoningParser | None: |
| 172 | """The underlying reasoning parser, if any.""" |
| 173 | return self._reasoning_parser |
| 174 | |
| 175 | @property |
| 176 | def tool_parser(self) -> ToolParser | None: |
| 177 | """The underlying tool parser, if any.""" |
| 178 | return self._tool_parser |
| 179 | |
| 180 | def _initialize_history_tool_call_cnt( |
| 181 | self, |
| 182 | request: ChatCompletionRequest | ResponsesRequest, |
| 183 | ) -> None: |
| 184 | state = self._stream_state |
| 185 | if state.history_tool_call_cnt_initialized: |
| 186 | return |
| 187 | if state.tool_call_id_type != "kimi_k2": |
| 188 | state.history_tool_call_cnt_initialized = True |
| 189 | return |
| 190 | state.history_tool_call_cnt = count_history_tool_calls(request) |
| 191 | state.history_tool_call_cnt_initialized = True |
| 192 | |
| 193 | def adjust_request( |
| 194 | self, request: ChatCompletionRequest | ResponsesRequest |
| 195 | ) -> ChatCompletionRequest | ResponsesRequest: |
| 196 | """Adjust the request parameters for tool calling. |
| 197 | |
| 198 | Can be overridden by subclasses to modify request parameters |
| 199 | (e.g., setting structured output schemas for tool calling). |
| 200 | |
| 201 | Args: |
| 202 | request: The original request. |
| 203 | |
| 204 | Returns: |
| 205 | The adjusted request. |
| 206 | |
| 207 | """ |
| 208 | return request |
| 209 | |
| 210 | def set_prompt_token_ids(self, prompt_token_ids: Sequence[int]) -> None: |
| 211 | """Provide the exact rendered prompt to parsers that need prefix state.""" |
| 212 | return |
| 213 | |
| 214 | @abstractmethod |
| 215 | def is_reasoning_end(self, input_ids: list[int]) -> bool: |
| 216 | """Check if the reasoning content ends in the input_ids. |
| 217 | |
| 218 | Called with the rendered prompt to decide whether generation starts |
| 219 | after reasoning. Must be a pure function of the input_ids. |
| 220 | |
| 221 | Args: |
| 222 | input_ids: The token IDs of the model output. |
| 223 | |
| 224 | Returns: |
| 225 | True if the reasoning content ends in the input_ids. |
| 226 | |
| 227 | """ |
| 228 | |
| 229 | @abstractmethod |
| 230 | def parse( |
| 231 | self, |
| 232 | model_output: str, |
| 233 | request: ChatCompletionRequest | ResponsesRequest, |
| 234 | enable_auto_tools: bool = False, |
| 235 | model_output_token_ids: Sequence[int] = (), |
| 236 | ) -> tuple[str | None, str | None, list[FunctionCall] | None]: |
| 237 | """Parse a complete model output, extracting reasoning and tool calls. |
| 238 | |
| 239 | Args: |
| 240 | model_output: The complete model-generated string. |
| 241 | request: The request object used to generate the output. |
| 242 | enable_auto_tools: Whether to enable automatic tool call parsing. |
| 243 | model_output_token_ids: The generated raw output token IDs. |
| 244 | |
| 245 | Returns: |
| 246 | A tuple of (reasoning, content, tool_calls). |
| 247 | |
| 248 | """ |
| 249 | |
| 250 | @abstractmethod |
| 251 | def parse_delta( |
| 252 | self, |
| 253 | delta_text: str, |
| 254 | delta_token_ids: list[int], |
| 255 | request: ChatCompletionRequest | ResponsesRequest, |
| 256 | prompt_token_ids: list[int] | None = None, |
| 257 | *, |
| 258 | finished: bool, |
| 259 | ) -> DeltaMessage | None: |
| 260 | """Parse a single streaming delta, orchestrating reasoning then |
| 261 | tool call extraction via internal stream state. |
| 262 | """ |
| 263 | |
| 264 | def count_reasoning_tokens(self, token_ids: Sequence[int]) -> int: |
| 265 | """Return the number of reasoning tokens in generated token IDs.""" |
| 266 | return 0 |
| 267 | |
| 268 | |
| 269 | def structured_outputs_to_format(params: StructuredOutputsParams) -> Format | None: |
| 270 | """Map StructuredOutputsParams in a XGrammar Format.""" |
| 271 | if params.json_object: |
| 272 | return JSONSchemaFormat(json_schema={"type": "object"}) |
| 273 | if params.json is not None: |
| 274 | schema = params.json |
| 275 | if isinstance(schema, str): |
| 276 | schema = json.loads(schema) |
| 277 | return JSONSchemaFormat(json_schema=schema) |
| 278 | if params.regex is not None: |
| 279 | return RegexFormat(pattern=params.regex) |
| 280 | if params.choice is not None: |
| 281 | return OrFormat( |
| 282 | elements=[ConstStringFormat(value=choice) for choice in params.choice] |
| 283 | ) |
| 284 | if params.grammar is not None: |
| 285 | from vllm.v1.structured_output.utils import grammar_is_likely_lark |
| 286 | |
| 287 | grammar = params.grammar |
| 288 | # IMPORTANT(arpera): |
| 289 | # GrammarFormat only accepts EBNF. |
| 290 | if grammar_is_likely_lark(grammar): |
| 291 | try: |
| 292 | grammar = str(Grammar.from_lark(grammar)) |
| 293 | except Exception as e: |
| 294 | raise VLLMValidationError("Invalid grammar specification.") from e |
| 295 | return GrammarFormat(grammar=grammar) |
| 296 | if params.structural_tag is not None: |
| 297 | s_tag = json.loads(params.structural_tag) |
| 298 | if "structures" in s_tag: |
| 299 | # LegacyStructuralTagResponseFormat |
| 300 | return TriggeredTagsFormat( |
| 301 | triggers=s_tag["triggers"], |
| 302 | tags=[ |
| 303 | TagFormat( |
| 304 | begin=structure["begin"], |
| 305 | content=JSONSchemaFormat(json_schema=structure["schema"]), |
| 306 | end=structure["end"], |
| 307 | ) |
| 308 | for structure in s_tag["structures"] |
| 309 | ], |
| 310 | ) |
| 311 | # StructuralTagResponseFormat |
| 312 | return StructuralTag.model_validate(s_tag).format |
| 313 | return None |
| 314 | |
| 315 | |
| 316 | def _xgrammar_supports(structured_outputs: StructuredOutputsParams) -> bool: |
| 317 | """Whether the constraint passes xgrammar validation on its own.""" |
| 318 | from vllm.v1.structured_output.backend_xgrammar import validate_xgrammar_grammar |
| 319 | |
| 320 | try: |
| 321 | # validate_xgrammar_grammar rewrites `choice` into `grammar` in place. |
| 322 | validate_xgrammar_grammar( |
| 323 | SamplingParams(structured_outputs=replace(structured_outputs)) |
| 324 | ) |
| 325 | except (TypeError, ValueError, VLLMValidationError): |
| 326 | return False |
| 327 | return True |
| 328 | |
| 329 | |
| 330 | class DelegatingParser(Parser): |
| 331 | """A Parser implementation that delegates to separate ReasoningParser and |
| 332 | ToolParser instances. |
| 333 | |
| 334 | This is the recommended base class for creating model-specific parsers |
| 335 | that combine existing reasoning and tool parser implementations. |
| 336 | Subclasses should set `self._reasoning_parser` and `self._tool_parser` |
| 337 | in their `__init__` method. |
| 338 | |
| 339 | If either parser is None, the corresponding methods will return default |
| 340 | values (no reasoning extraction, no tool calls). |
| 341 | """ |
| 342 | |
| 343 | def extract_reasoning( |
| 344 | self, |
| 345 | model_output: str, |
| 346 | request: ChatCompletionRequest | ResponsesRequest, |
| 347 | ) -> tuple[str | None, str | None]: |
| 348 | if self._reasoning_parser is None: |
| 349 | return None, model_output |
| 350 | return self._reasoning_parser.extract_reasoning(model_output, request) |
| 351 | |
| 352 | def _get_function_name( |
| 353 | self, request: ChatCompletionRequest | ResponsesRequest |
| 354 | ) -> str: |
| 355 | if request.tool_choice and isinstance(request.tool_choice, ToolChoiceFunction): |
| 356 | return request.tool_choice.name |
| 357 | if request.tool_choice and isinstance( |
| 358 | request.tool_choice, ChatCompletionNamedToolChoiceParam |
| 359 | ): |
| 360 | return request.tool_choice.function.name |
| 361 | raise ValueError("Invalid tool_choice for function name extraction.") |
| 362 | |
| 363 | def _make_tool_call_id(self, function_name: str) -> str | None: |
| 364 | state = self._stream_state |
| 365 | if state.tool_call_id_type != "kimi_k2": |
| 366 | return None |
| 367 | tool_call_id = make_tool_call_id( |
| 368 | id_type=state.tool_call_id_type, |
| 369 | func_name=function_name, |
| 370 | idx=state.history_tool_call_cnt, |
| 371 | ) |
| 372 | state.history_tool_call_cnt += 1 |
| 373 | return tool_call_id |
| 374 | |
| 375 | def _extract_tool_calls( |
| 376 | self, |
| 377 | content: str | None, |
| 378 | request: ChatCompletionRequest | ResponsesRequest, |
| 379 | enable_auto_tools: bool = False, |
| 380 | ) -> tuple[list[FunctionCall] | None, str | None]: |
| 381 | tool_parser = self._tool_parser |
| 382 | if tool_parser is None: |
| 383 | return [], content |
| 384 | |
| 385 | if request.tool_choice == "none": |
| 386 | if self._engine_based: |
| 387 | result = self.extract_tool_calls(content or "", request=request) |
| 388 | return [], result.content |
| 389 | return [], content |
| 390 | |
| 391 | supports_required_and_named = tool_parser.supports_required_and_named |
| 392 | is_named_tool_choice = request.tool_choice and isinstance( |
| 393 | request.tool_choice, |
| 394 | (ToolChoiceFunction, ChatCompletionNamedToolChoiceParam), |
| 395 | ) |
| 396 | is_required_tool_choice = request.tool_choice == "required" |
| 397 | is_auto_tool_choice = enable_auto_tools and ( |
| 398 | request.tool_choice == "auto" |
| 399 | or request.tool_choice is None |
| 400 | or ( |
| 401 | not supports_required_and_named |
| 402 | and (is_named_tool_choice or is_required_tool_choice) |
| 403 | ) |
| 404 | ) |
| 405 | |
| 406 | tool_calls = list[FunctionCall]() |
| 407 | if is_named_tool_choice and supports_required_and_named: |
| 408 | if content is None or (isinstance(content, str) and not content.strip()): |
| 409 | return [], None |
| 410 | function_name = self._get_function_name(request) |
| 411 | tool_calls.append( |
| 412 | FunctionCall( |
| 413 | id=self._make_tool_call_id(function_name), |
| 414 | name=function_name, |
| 415 | arguments=content, |
| 416 | ) |
| 417 | ) |
| 418 | content = None |
| 419 | elif is_required_tool_choice and supports_required_and_named: |
| 420 | # "required" with standard JSON-based parsing |
| 421 | parsed_calls = [] |
| 422 | with contextlib.suppress(ValidationError): |
| 423 | content = content or "" |
| 424 | parsed_calls = TypeAdapter(list[FunctionDefinition]).validate_json( |
| 425 | content |
| 426 | ) |
| 427 | for tc in parsed_calls: |
| 428 | tool_calls.append( |
| 429 | FunctionCall( |
| 430 | id=self._make_tool_call_id(tc.name), |
| 431 | name=tc.name, |
| 432 | arguments=json.dumps(tc.parameters, ensure_ascii=False), |
| 433 | ) |
| 434 | ) |
| 435 | content = None |
| 436 | elif is_auto_tool_choice: |
| 437 | # Automatic Tool Call Parsing (also used as fallback for |
| 438 | # required/named when supports_required_and_named=False) |
| 439 | tool_call_info = self.extract_tool_calls( |
| 440 | content if content is not None else "", |
| 441 | request=request, |
| 442 | ) |
| 443 | if tool_call_info is not None and tool_call_info.tools_called: |
| 444 | tool_calls.extend( |
| 445 | FunctionCall( |
| 446 | id=tc.id, |
| 447 | name=tc.function.name, |
| 448 | arguments=tc.function.arguments, |
| 449 | ) |
| 450 | for tc in tool_call_info.tool_calls |
| 451 | ) |
| 452 | content = tool_call_info.content |
| 453 | if content and content.strip() == "": |
| 454 | content = None |
| 455 | else: |
| 456 | # No tool calls. |
| 457 | # For required/named tool choice (when falling back to auto |
| 458 | # parsing), if content is empty or whitespace-only, return |
| 459 | # empty list with None content. |
| 460 | if (is_required_tool_choice or is_named_tool_choice) and ( |
| 461 | content is None |
| 462 | or (isinstance(content, str) and not content.strip()) |
| 463 | ): |
| 464 | return [], None |
| 465 | # No complete tool calls: for engine-based parsers, return |
| 466 | # the tool parser's content, which drops incomplete |
| 467 | # tool-call markup (e.g. a <tool_call> opener truncated by |
| 468 | # max_tokens or a stop string), so the non-streaming path |
| 469 | # matches streaming. Legacy parsers keep their existing |
| 470 | # behavior of returning the raw content. |
| 471 | if self._engine_based and tool_call_info is not None: |
| 472 | return None, tool_call_info.content or None |
| 473 | return None, content |
| 474 | |
| 475 | return tool_calls, content |
| 476 | |
| 477 | def adjust_request( |
| 478 | self, request: ChatCompletionRequest | ResponsesRequest |
| 479 | ) -> ChatCompletionRequest | ResponsesRequest: |
| 480 | if self._reasoning_parser is not None: |
| 481 | request = self._reasoning_parser.adjust_request(request) |
| 482 | if self._tool_parser is not None: |
| 483 | request = self._apply_structural_tag(request) |
| 484 | request = self._tool_parser.adjust_request(request) |
| 485 | return request |
| 486 | |
| 487 | def _apply_structural_tag( |
| 488 | self, request: ChatCompletionRequest | ResponsesRequest |
| 489 | ) -> ChatCompletionRequest | ResponsesRequest: |
| 490 | tool_parser = self._tool_parser |
| 491 | if tool_parser is None or not request.tools: |
| 492 | return request |
| 493 | |
| 494 | need_tool_calling = ( |
| 495 | request.tool_choice == "auto" |
| 496 | or request.tool_choice == "required" |
| 497 | or isinstance( |
| 498 | request.tool_choice, |
| 499 | (ChatCompletionNamedToolChoiceParam, ToolChoiceFunction), |
| 500 | ) |
| 501 | ) |
| 502 | if not need_tool_calling: |
| 503 | return request |
| 504 | |
| 505 | structured_outputs = request.extract_structured_outputs() |
| 506 | is_auto = request.tool_choice == "auto" |
| 507 | single_call = request.parallel_tool_calls is False |
| 508 | strict_level = self.tool_strict_level |
| 509 | if single_call and not (is_auto and structured_outputs is not None): |
| 510 | # Limiting the call count needs the call envelope in the grammar. |
| 511 | # Auto with structured outputs keeps its format-only grammar, so the |
| 512 | # flag never makes calls possible that are impossible without it. |
| 513 | strict_level = max(strict_level, ToolStrictLevel.FUNCTION) |
| 514 | |
| 515 | resolved_tools = None |
| 516 | if tool_parser.structural_tag_model is not None: |
| 517 | resolved_tools = resolve_tool_strictness( |
| 518 | request.tools, |
| 519 | request.tool_choice, |
| 520 | strict_level, |
| 521 | ) |
| 522 | |
| 523 | output_format = None |
| 524 | if resolved_tools and is_auto and structured_outputs: |
| 525 | if not _xgrammar_supports(structured_outputs): |
| 526 | logger.warning_once( |
| 527 | "Tool calls are not constrained for tool_choice=auto with " |
| 528 | "structured outputs because xgrammar does not support the " |
| 529 | "structured output constraint; the structured output " |
| 530 | "constraint applies.", |
| 531 | scope="local", |
| 532 | ) |
| 533 | return request |
| 534 | output_format = structured_outputs_to_format(structured_outputs) |
| 535 | |
| 536 | if resolved_tools is not None: |
| 537 | tag_request = request |
| 538 | if output_format is not None: |
| 539 | # IMPORTANT(arpera): |
| 540 | # The "auto" tag allows plain text in response. |
| 541 | # Use tag "required" here to ensure structured-output branch |
| 542 | # remains meaningful. |
| 543 | tag_request = request.model_copy(update={"tool_choice": "required"}) |
| 544 | tools_structural_tag = tool_parser.get_structural_tag( |
| 545 | tag_request, |
| 546 | reasoning=False, |
| 547 | strict_level=strict_level, |
| 548 | ) |
| 549 | else: |
| 550 | tools_structural_tag = None |
| 551 | |
| 552 | if tools_structural_tag is None: |
| 553 | if is_auto and structured_outputs is not None: |
| 554 | logger.warning_once( |
| 555 | "Tool calls are not constrained for tool_choice=auto with " |
| 556 | "structured outputs because structural tags are unavailable; " |
| 557 | "the structured output constraint applies.", |
| 558 | scope="local", |
| 559 | ) |
| 560 | return request |
| 561 | |
| 562 | if single_call: |
| 563 | limit_to_single_tool_call(tools_structural_tag) |
| 564 | structural_tag = tools_structural_tag |
| 565 | if output_format is not None: |
| 566 | structural_tag = StructuralTag( |
| 567 | format=OrFormat(elements=[tools_structural_tag.format, output_format]) |
| 568 | ) |
| 569 | elif structured_outputs is not None and not is_auto: |
| 570 | logger.warning_once( |
| 571 | "structured outputs are ignored because tool_choice forces tool call.", |
| 572 | scope="local", |
| 573 | ) |
| 574 | |
| 575 | request.structured_outputs = StructuredOutputsParams( |
| 576 | structural_tag=json.dumps(structural_tag.model_dump()), |
| 577 | ) |
| 578 | if isinstance(request, ResponsesRequest): |
| 579 | request.text = None |
| 580 | else: |
| 581 | request.response_format = None |
| 582 | return request |
| 583 | |
| 584 | def extract_reasoning_streaming( |
| 585 | self, |
| 586 | previous_text: str, |
| 587 | current_text: str, |
| 588 | delta_text: str, |
| 589 | previous_token_ids: Sequence[int], |
| 590 | current_token_ids: Sequence[int], |
| 591 | delta_token_ids: Sequence[int], |
| 592 | ) -> DeltaMessage | None: |
| 593 | if self._reasoning_parser is None: |
| 594 | return DeltaMessage(content=delta_text) |
| 595 | return self._reasoning_parser.extract_reasoning_streaming( |
| 596 | previous_text, |
| 597 | current_text, |
| 598 | delta_text, |
| 599 | previous_token_ids, |
| 600 | current_token_ids, |
| 601 | delta_token_ids, |
| 602 | ) |
| 603 | |
| 604 | def extract_tool_calls( |
| 605 | self, |
| 606 | model_output: str, |
| 607 | request: ChatCompletionRequest | ResponsesRequest, |
| 608 | ) -> ExtractedToolCallInformation: |
| 609 | if self._tool_parser is None: |
| 610 | return ExtractedToolCallInformation( |
| 611 | tools_called=False, tool_calls=[], content=model_output |
| 612 | ) |
| 613 | result = None |
| 614 | is_tool_called: bool | Exception = False |
| 615 | try: |
| 616 | result = self._tool_parser.extract_tool_calls( |
| 617 | model_output, |
| 618 | request=request, # type: ignore[arg-type] |
| 619 | ) |
| 620 | is_tool_called = bool(result.tools_called) |
| 621 | except Exception as e: |
| 622 | is_tool_called = e |
| 623 | raise |
| 624 | finally: |
| 625 | record_tool_parser_invocation( |
| 626 | is_tool_called=is_tool_called, |
| 627 | is_streaming=False, |
| 628 | request=request, |
| 629 | ) |
| 630 | return result |
| 631 | |
| 632 | def extract_tool_calls_streaming( |
| 633 | self, |
| 634 | previous_text: str, |
| 635 | current_text: str, |
| 636 | delta_text: str, |
| 637 | previous_token_ids: Sequence[int], |
| 638 | current_token_ids: Sequence[int], |
| 639 | delta_token_ids: Sequence[int], |
| 640 | request: ChatCompletionRequest | ResponsesRequest, |
| 641 | ) -> DeltaMessage | None: |
| 642 | if self._tool_parser is None: |
| 643 | return None |
| 644 | result = None |
| 645 | is_tool_called: bool | Exception = False |
| 646 | try: |
| 647 | result = self._tool_parser.extract_tool_calls_streaming( |
| 648 | previous_text, |
| 649 | current_text, |
| 650 | delta_text, |
| 651 | previous_token_ids, |
| 652 | current_token_ids, |
| 653 | delta_token_ids, |
| 654 | request, # type: ignore[arg-type] |
| 655 | ) |
| 656 | is_tool_called = bool(result and result.tool_calls) |
| 657 | except Exception as e: |
| 658 | is_tool_called = e |
| 659 | raise |
| 660 | finally: |
| 661 | record_tool_parser_invocation( |
| 662 | is_tool_called=is_tool_called, |
| 663 | is_streaming=True, |
| 664 | request=request, |
| 665 | ) |
| 666 | return result |
| 667 | |
| 668 | def _extract_tool_calls_streaming( |
| 669 | self, |
| 670 | previous_text: str, |
| 671 | current_text: str, |
| 672 | delta_text: str, |
| 673 | previous_token_ids: Sequence[int], |
| 674 | current_token_ids: Sequence[int], |
| 675 | delta_token_ids: Sequence[int], |
| 676 | request: ChatCompletionRequest | ResponsesRequest, |
| 677 | # The following parameters are used for "required" tool choice parsing and are |
| 678 | # tracked in StreamState for streaming parsing. |
| 679 | tool_call_idx: int | None = None, |
| 680 | tool_call_id_type: str = "random", |
| 681 | function_name_returned: bool = False, |
| 682 | ) -> tuple[DeltaMessage | None, bool]: |
| 683 | assert self._tool_parser is not None |
| 684 | supports_required_and_named = self._tool_parser.supports_required_and_named |
| 685 | |
| 686 | if request.tool_choice == "none": |
| 687 | if self._engine_based: |
| 688 | # Engine-backed parsers route content extraction through |
| 689 | # extract_tool_calls_streaming, so run the full pipeline |
| 690 | # and strip tool_calls after. |
| 691 | delta_message = self.extract_tool_calls_streaming( |
| 692 | previous_text, |
| 693 | current_text, |
| 694 | delta_text, |
| 695 | previous_token_ids, |
| 696 | current_token_ids, |
| 697 | delta_token_ids, |
| 698 | request, # type: ignore[arg-type] |
| 699 | ) |
| 700 | if delta_message: |
| 701 | delta_message.tool_calls = [] |
| 702 | return delta_message, False |
| 703 | return (DeltaMessage(content=delta_text) if delta_text else None), False |
| 704 | |
| 705 | if ( |
| 706 | supports_required_and_named |
| 707 | and request.tool_choice |
| 708 | and isinstance( |
| 709 | request.tool_choice, |
| 710 | (ToolChoiceFunction, ChatCompletionNamedToolChoiceParam), |
| 711 | ) |
| 712 | ): |
| 713 | delta_message, function_name_returned = extract_named_tool_call_streaming( |
| 714 | delta_text=delta_text, |
| 715 | function_name=self._get_function_name(request), |
| 716 | function_name_returned=function_name_returned, |
| 717 | tool_call_idx=tool_call_idx, |
| 718 | tool_call_id_type=tool_call_id_type, |
| 719 | tokenizer=self.model_tokenizer, |
| 720 | ) |
| 721 | return delta_message, function_name_returned |
| 722 | |
| 723 | if supports_required_and_named and request.tool_choice == "required": |
| 724 | delta_message, function_name_returned = ( |
| 725 | extract_required_tool_call_streaming( |
| 726 | previous_text=previous_text, |
| 727 | current_text=current_text, |
| 728 | delta_text=delta_text, |
| 729 | function_name_returned=function_name_returned, |
| 730 | tool_call_idx=tool_call_idx, |
| 731 | tool_call_id_type=tool_call_id_type, |
| 732 | ) |
| 733 | ) |
| 734 | return delta_message, function_name_returned |
| 735 | return self.extract_tool_calls_streaming( |
| 736 | previous_text, |
| 737 | current_text, |
| 738 | delta_text, |
| 739 | previous_token_ids, |
| 740 | current_token_ids, |
| 741 | delta_token_ids, |
| 742 | request, |
| 743 | ), False |
| 744 | |
| 745 | def is_reasoning_end(self, input_ids: list[int]) -> bool: |
| 746 | if self._reasoning_parser is None: |
| 747 | return False |
| 748 | return self._reasoning_parser.is_reasoning_end(input_ids) |
| 749 | |
| 750 | def _is_reasoning_end_streaming( |
| 751 | self, input_ids: list[int], delta_ids: list[int] |
| 752 | ) -> bool: |
| 753 | if self._reasoning_parser is None: |
| 754 | return False |
| 755 | return self._reasoning_parser.is_reasoning_end_streaming(input_ids, delta_ids) |
| 756 | |
| 757 | def _extract_content_ids(self, input_ids: list[int]) -> list[int]: |
| 758 | if self._reasoning_parser is None: |
| 759 | return input_ids |
| 760 | return self._reasoning_parser.extract_content_ids(input_ids) |
| 761 | |
| 762 | def _in_reasoning_phase(self, state: StreamState) -> bool: |
| 763 | if self._reasoning_parser is None: |
| 764 | return False |
| 765 | return not state.reasoning_ended |
| 766 | |
| 767 | def _in_tool_call_phase(self, state: StreamState) -> bool: |
| 768 | if self._tool_parser is None: |
| 769 | return False |
| 770 | return state.reasoning_ended |
| 771 | |
| 772 | def _append_unstreamed_tool_args( |
| 773 | self, |
| 774 | delta_message: DeltaMessage | None, |
| 775 | ) -> None: |
| 776 | """Append parsed-but-unstreamed tool-call arguments to *delta_message*.""" |
| 777 | if ( |
| 778 | self._tool_parser is not None |
| 779 | and delta_message |
| 780 | and delta_message.tool_calls |
| 781 | and (last_tc := delta_message.tool_calls[-1]).function |
| 782 | ): |
| 783 | last_tc.function.arguments = ( |
| 784 | last_tc.function.arguments or "" |
| 785 | ) + self._tool_parser.get_remaining_unstreamed_args() |
| 786 | |
| 787 | def finalize_generation( |
| 788 | self, |
| 789 | delta_message: DeltaMessage | None, |
| 790 | request: ChatCompletionRequest | ResponsesRequest, |
| 791 | state: StreamState, |
| 792 | ) -> DeltaMessage | None: |
| 793 | """Finalize generation for cases where generation was incomplete. |
| 794 | For example, if streaming terminated before reasoning ended |
| 795 | """ |
| 796 | fallback_fn = getattr( |
| 797 | self._reasoning_parser, "get_streaming_fallback_content", None |
| 798 | ) |
| 799 | if fallback_fn is not None and not state.reasoning_ended: |
| 800 | promoted = fallback_fn(state.previous_text, request) |
| 801 | if promoted: |
| 802 | if delta_message is None: |
| 803 | delta_message = DeltaMessage() |
| 804 | delta_message.content = (delta_message.content or "") + promoted |
| 805 | |
| 806 | self._append_unstreamed_tool_args(delta_message) |
| 807 | return delta_message |
| 808 | |
| 809 | def parse( |
| 810 | self, |
| 811 | model_output: str, |
| 812 | request: ChatCompletionRequest | ResponsesRequest, |
| 813 | enable_auto_tools: bool = False, |
| 814 | model_output_token_ids: Sequence[int] = (), |
| 815 | ) -> tuple[str | None, str | None, list[FunctionCall] | None]: |
| 816 | self._initialize_history_tool_call_cnt(request) |
| 817 | reasoning, content = self.extract_reasoning(model_output, request) |
| 818 | tool_calls, content = self._extract_tool_calls( |
| 819 | content=content, |
| 820 | request=request, |
| 821 | enable_auto_tools=enable_auto_tools, |
| 822 | ) |
| 823 | return reasoning, content, tool_calls |
| 824 | |
| 825 | def parse_delta( |
| 826 | self, |
| 827 | delta_text: str, |
| 828 | delta_token_ids: list[int], |
| 829 | request: ChatCompletionRequest | ResponsesRequest, |
| 830 | prompt_token_ids: list[int] | None = None, |
| 831 | *, |
| 832 | finished: bool, |
| 833 | ) -> DeltaMessage | None: |
| 834 | self._initialize_history_tool_call_cnt(request) |
| 835 | state = self._stream_state |
| 836 | |
| 837 | if not state.prompt_reasoning_checked and prompt_token_ids is not None: |
| 838 | state.prompt_reasoning_checked = True |
| 839 | if self._reasoning_parser is None or self.is_reasoning_end( |
| 840 | prompt_token_ids |
| 841 | ): |
| 842 | state.reasoning_ended = True |
| 843 | else: |
| 844 | # Reasoning is still open at the end of the prompt; let the |
| 845 | # reasoning parser adjust its initial parsing state so the |
| 846 | # first generated tokens are classified correctly. |
| 847 | self._reasoning_parser.adjust_initial_state_from_prompt( |
| 848 | prompt_token_ids |
| 849 | ) |
| 850 | |
| 851 | current_text, current_token_ids = state.advance(delta_text, delta_token_ids) |
| 852 | delta_message: DeltaMessage | None = None |
| 853 | reasoning_transitioned = False |
| 854 | |
| 855 | # Reasoning extraction |
| 856 | if self._in_reasoning_phase(state): |
| 857 | delta_message = self.extract_reasoning_streaming( |
| 858 | previous_text=state.previous_text, |
| 859 | current_text=current_text, |
| 860 | delta_text=delta_text, |
| 861 | previous_token_ids=state.previous_token_ids, |
| 862 | current_token_ids=current_token_ids, |
| 863 | delta_token_ids=delta_token_ids, |
| 864 | ) |
| 865 | reasoning_parser = self._reasoning_parser |
| 866 | if reasoning_parser is not None and reasoning_parser.engine_based_streaming: |
| 867 | should_transition = ( |
| 868 | reasoning_parser.has_engine_confirmed_reasoning_end() |
| 869 | ) |
| 870 | else: |
| 871 | should_transition = self._is_reasoning_end_streaming( |
| 872 | current_token_ids, delta_token_ids |
| 873 | ) |
| 874 | if should_transition: |
| 875 | state.reasoning_ended = True |
| 876 | reasoning_transitioned = True |
| 877 | current_token_ids = self._extract_content_ids(delta_token_ids) |
| 878 | # Flush whenever the reasoning parser is engine-based (not only |
| 879 | # when _engine_based is True): it buffers the post-marker text |
| 880 | # (e.g. the "<" of "<tool_call>"), surfaced via finish_streaming(). |
| 881 | flush_delta = ( |
| 882 | reasoning_parser.finish_streaming() # type: ignore[union-attr, attr-defined] |
| 883 | if reasoning_parser is not None |
| 884 | and reasoning_parser.engine_based_streaming |
| 885 | else None |
| 886 | ) |
| 887 | current_text = ( |
| 888 | (delta_message.content if delta_message else None) or "" |
| 889 | ) + ((flush_delta.content if flush_delta else None) or "") |
| 890 | if self._engine_based: |
| 891 | if delta_message and self._tool_parser is not None: |
| 892 | delta_message.content = None |
| 893 | else: |
| 894 | delta_text = current_text |
| 895 | |
| 896 | # Tool call extraction |
| 897 | if self._in_tool_call_phase(state): |
| 898 | if not state.tool_call_text_started: |
| 899 | state.tool_call_text_started = True |
| 900 | state.previous_text = "" |
| 901 | state.previous_token_ids = [] |
| 902 | delta_text = current_text |
| 903 | delta_token_ids = current_token_ids |
| 904 | |
| 905 | reasoning_from_this_batch = ( |
| 906 | delta_message.reasoning if delta_message else None |
| 907 | ) |
| 908 | |
| 909 | delta_message, state.function_name_returned = ( |
| 910 | self._extract_tool_calls_streaming( |
| 911 | previous_text=state.previous_text, |
| 912 | current_text=current_text, |
| 913 | delta_text=delta_text, |
| 914 | previous_token_ids=state.previous_token_ids, |
| 915 | current_token_ids=current_token_ids, |
| 916 | delta_token_ids=delta_token_ids, |
| 917 | request=request, # type: ignore[arg-type] |
| 918 | tool_call_idx=state.history_tool_call_cnt, |
| 919 | tool_call_id_type=state.tool_call_id_type, |
| 920 | function_name_returned=state.function_name_returned, |
| 921 | ) |
| 922 | ) |
| 923 | |
| 924 | if reasoning_from_this_batch: |
| 925 | if delta_message is None: |
| 926 | delta_message = DeltaMessage(reasoning=reasoning_from_this_batch) |
| 927 | elif not delta_message.reasoning: |
| 928 | delta_message.reasoning = reasoning_from_this_batch |
| 929 | |
| 930 | if ( |
| 931 | delta_message |
| 932 | and delta_message.tool_calls |
| 933 | and delta_message.tool_calls[0].id is not None |
| 934 | ): |
| 935 | state.history_tool_call_cnt += 1 |
| 936 | |
| 937 | # No phase active: pass through as content. |
| 938 | # Skip when reasoning just ended in this delta — the engine already |
| 939 | # consumed the end-of-reasoning marker (e.g. </think>) and |
| 940 | # delta_text still contains the raw marker text. |
| 941 | if ( |
| 942 | delta_message is None |
| 943 | and not reasoning_transitioned |
| 944 | and not self._in_reasoning_phase(state) |
| 945 | and not self._in_tool_call_phase(state) |
| 946 | ): |
| 947 | delta_message = DeltaMessage(content=delta_text) |
| 948 | |
| 949 | state.commit(current_text, current_token_ids) |
| 950 | |
| 951 | if finished: |
| 952 | delta_message = self.finalize_generation(delta_message, request, state) |
| 953 | delta_message = self._flush_engine_parsers(delta_message) |
| 954 | |
| 955 | # Suppress reasoning deltas if not requested |
| 956 | if delta_message and not request.include_reasoning: |
| 957 | delta_message.reasoning = None |
| 958 | |
| 959 | # If only reasoning was in the message (no content, no tool_calls) |
| 960 | # skip emitting entirely |
| 961 | if not delta_message.content and not delta_message.tool_calls: |
| 962 | delta_message = None |
| 963 | |
| 964 | return delta_message |
| 965 | |
| 966 | def count_reasoning_tokens(self, token_ids: Sequence[int]) -> int: |
| 967 | """Count reasoning tokens through the configured reasoning parser.""" |
| 968 | if self._reasoning_parser is None: |
| 969 | return 0 |
| 970 | return self._reasoning_parser.count_reasoning_tokens(token_ids) |
| 971 | |
| 972 | def _flush_engine_parsers( |
| 973 | self, delta_message: DeltaMessage | None |
| 974 | ) -> DeltaMessage | None: |
| 975 | """Flush buffered state from engine-based parsers at stream end.""" |
| 976 | reasoning_ended = self._stream_state.reasoning_ended |
| 977 | for parser in (self._reasoning_parser, self._tool_parser): |
| 978 | if not getattr(parser, "engine_based_streaming", False): |
| 979 | continue |
| 980 | # When reasoning has ended and we transitioned to the tool |
| 981 | # phase, the reasoning parser's engine may still have buffered |
| 982 | # characters from tool-call markup it saw with |
| 983 | # skip_tool_parsing=True. Flushing that would leak spurious |
| 984 | # content (e.g. a stray '"'), so skip it. |
| 985 | if parser is self._reasoning_parser and reasoning_ended: |
| 986 | continue |
| 987 | finish = getattr(parser, "finish_streaming", None) |
| 988 | if finish is None: |
| 989 | continue |
| 990 | flush_delta = finish() |
| 991 | if flush_delta is None: |
| 992 | continue |
| 993 | if delta_message is None: |
| 994 | delta_message = flush_delta |
| 995 | else: |
| 996 | if flush_delta.content: |
| 997 | delta_message.content = ( |
| 998 | delta_message.content or "" |
| 999 | ) + flush_delta.content |
| 1000 | if flush_delta.reasoning: |
| 1001 | delta_message.reasoning = ( |
| 1002 | delta_message.reasoning or "" |
| 1003 | ) + flush_delta.reasoning |
| 1004 | if flush_delta.tool_calls: |
| 1005 | delta_message.tool_calls = ( |
| 1006 | delta_message.tool_calls or [] |
| 1007 | ) + flush_delta.tool_calls |
| 1008 | return delta_message |