| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Gemma4 parser. |
| 4 | |
| 5 | Handles channel-based reasoning plus custom tool call format in a single |
| 6 | state machine:: |
| 7 | |
| 8 | <|channel>thought |
| 9 | ...reasoning...<channel|> |
| 10 | <|tool_call>call:func_name{key:<|"|>value<|"|>,num:42}<tool_call|> |
| 11 | """ |
| 12 | |
| 13 | from __future__ import annotations |
| 14 | |
| 15 | import functools |
| 16 | import json |
| 17 | from collections.abc import Sequence |
| 18 | from typing import TYPE_CHECKING |
| 19 | |
| 20 | from vllm.entrypoints.generate.base.protocol import DeltaMessage |
| 21 | from vllm.logger import init_logger |
| 22 | from vllm.parser.engine.events import EventType, SemanticEvent |
| 23 | from vllm.parser.engine.parser_engine import ParserEngine |
| 24 | from vllm.parser.engine.parser_engine_config import ( |
| 25 | ParserEngineConfig, |
| 26 | ParserState, |
| 27 | Transition, |
| 28 | ) |
| 29 | |
| 30 | if TYPE_CHECKING: |
| 31 | from vllm.entrypoints.openai.chat_completion.protocol import ( |
| 32 | ChatCompletionRequest, |
| 33 | ) |
| 34 | from vllm.entrypoints.openai.responses.protocol import ResponsesRequest |
| 35 | from vllm.tokenizers import TokenizerLike |
| 36 | from vllm.tool_parsers.abstract_tool_parser import Tool |
| 37 | |
| 38 | CHANNEL_START = "<|channel>" |
| 39 | CHANNEL_END = "<channel|>" |
| 40 | TOOL_CALL_START = "<|tool_call>" |
| 41 | TOOL_CALL_END = "<tool_call|>" |
| 42 | STRING_DELIM = '<|"|>' |
| 43 | _DELIM_LEN = len(STRING_DELIM) |
| 44 | |
| 45 | logger = init_logger(__name__) |
| 46 | |
| 47 | |
| 48 | # --------------------------------------------------------------------------- |
| 49 | # Gemma4 argument parser |
| 50 | # --------------------------------------------------------------------------- |
| 51 | |
| 52 | _PARTIAL_DELIM_SUFFIXES = tuple( |
| 53 | STRING_DELIM[:k] for k in range(len(STRING_DELIM), 0, -1) |
| 54 | ) |
| 55 | |
| 56 | |
| 57 | def _strip_partial_delim(value: str) -> str: |
| 58 | """Strip a trailing partial ``STRING_DELIM`` prefix from *value*. |
| 59 | |
| 60 | Prevents partial delimiters from leaking into the streamed JSON diff. |
| 61 | """ |
| 62 | for suffix in _PARTIAL_DELIM_SUFFIXES: |
| 63 | if value.endswith(suffix): |
| 64 | return value[: -len(suffix)] |
| 65 | return value |
| 66 | |
| 67 | |
| 68 | def _parse_gemma4_args(args_str: str, *, partial: bool = False) -> dict: |
| 69 | """Parse Gemma4's custom key:value format into a Python dict. |
| 70 | |
| 71 | Format examples:: |
| 72 | |
| 73 | location:<|"|>Tokyo<|"|> |
| 74 | location:<|"|>San Francisco<|"|>,unit:<|"|>celsius<|"|> |
| 75 | count:42,flag:true |
| 76 | nested:{inner_key:<|"|>val<|"|>} |
| 77 | items:[<|"|>a<|"|>,<|"|>b<|"|>] |
| 78 | |
| 79 | Args: |
| 80 | args_str: The raw Gemma4 argument string. |
| 81 | partial: When True (streaming), bare values at end of string are |
| 82 | omitted because they may be incomplete and type-unstable |
| 83 | (e.g. partial boolean parsed as bare string). |
| 84 | |
| 85 | Returns a dict ready for ``json.dumps()``. |
| 86 | |
| 87 | """ |
| 88 | if not args_str or not args_str.strip(): |
| 89 | return {} |
| 90 | |
| 91 | result: dict = {} |
| 92 | i = 0 |
| 93 | n = len(args_str) |
| 94 | |
| 95 | while i < n: |
| 96 | while i < n and args_str[i] in (" ", ",", "\n", "\t"): |
| 97 | i += 1 |
| 98 | if i >= n: |
| 99 | break |
| 100 | |
| 101 | key_start = i |
| 102 | while i < n and args_str[i] != ":": |
| 103 | i += 1 |
| 104 | if i >= n: |
| 105 | break |
| 106 | key = args_str[key_start:i].strip() |
| 107 | if key.startswith(STRING_DELIM) and key.endswith(STRING_DELIM): |
| 108 | key = key[_DELIM_LEN:-_DELIM_LEN] |
| 109 | i += 1 |
| 110 | |
| 111 | if i >= n: |
| 112 | if not partial: |
| 113 | result[key] = "" |
| 114 | break |
| 115 | |
| 116 | while i < n and args_str[i] in (" ", "\n", "\t"): |
| 117 | i += 1 |
| 118 | if i >= n: |
| 119 | if not partial: |
| 120 | result[key] = "" |
| 121 | break |
| 122 | |
| 123 | if args_str.startswith(STRING_DELIM, i): |
| 124 | i += _DELIM_LEN |
| 125 | val_start = i |
| 126 | end_pos = args_str.find(STRING_DELIM, i) |
| 127 | if end_pos == -1: |
| 128 | # Unterminated string — take rest, strip partial delimiter. |
| 129 | value = args_str[val_start:] |
| 130 | if partial: |
| 131 | value = _strip_partial_delim(value) |
| 132 | result[key] = value |
| 133 | break |
| 134 | result[key] = args_str[val_start:end_pos] |
| 135 | i = end_pos + _DELIM_LEN |
| 136 | |
| 137 | elif args_str[i] == "{": |
| 138 | depth = 1 |
| 139 | obj_start = i + 1 |
| 140 | i += 1 |
| 141 | while i < n and depth > 0: |
| 142 | if args_str.startswith(STRING_DELIM, i): |
| 143 | # Skip over string contents to avoid counting { inside strings |
| 144 | i += _DELIM_LEN |
| 145 | next_delim = args_str.find(STRING_DELIM, i) |
| 146 | i = n if next_delim == -1 else next_delim + _DELIM_LEN |
| 147 | continue |
| 148 | if args_str[i] == "{": |
| 149 | depth += 1 |
| 150 | elif args_str[i] == "}": |
| 151 | depth -= 1 |
| 152 | i += 1 |
| 153 | if depth > 0: |
| 154 | # Incomplete nested object — use i (not i-1) to avoid |
| 155 | # dropping the last char, and recurse as partial. |
| 156 | result[key] = _parse_gemma4_args(args_str[obj_start:i], partial=True) |
| 157 | else: |
| 158 | result[key] = _parse_gemma4_args(args_str[obj_start : i - 1]) |
| 159 | |
| 160 | elif args_str[i] == "[": |
| 161 | depth = 1 |
| 162 | arr_start = i + 1 |
| 163 | i += 1 |
| 164 | while i < n and depth > 0: |
| 165 | if args_str.startswith(STRING_DELIM, i): |
| 166 | i += _DELIM_LEN |
| 167 | next_delim = args_str.find(STRING_DELIM, i) |
| 168 | i = n if next_delim == -1 else next_delim + _DELIM_LEN |
| 169 | continue |
| 170 | if args_str[i] == "[": |
| 171 | depth += 1 |
| 172 | elif args_str[i] == "]": |
| 173 | depth -= 1 |
| 174 | i += 1 |
| 175 | if depth > 0: |
| 176 | result[key] = _parse_gemma4_array(args_str[arr_start:i], partial=True) |
| 177 | else: |
| 178 | result[key] = _parse_gemma4_array(args_str[arr_start : i - 1]) |
| 179 | |
| 180 | else: |
| 181 | val_start = i |
| 182 | while i < n and args_str[i] not in (",", "}", "]"): |
| 183 | i += 1 |
| 184 | if partial and i >= n: |
| 185 | # Value may be incomplete (e.g. partial boolean) — |
| 186 | # withhold to avoid type instability during streaming. |
| 187 | break |
| 188 | if i == val_start: |
| 189 | logger.warning( |
| 190 | "Gemma4 args parser made no progress at position %d; " |
| 191 | "aborting on malformed input.", |
| 192 | i, |
| 193 | ) |
| 194 | break |
| 195 | raw_val = args_str[val_start:i].strip() |
| 196 | if partial and raw_val.endswith("."): |
| 197 | # Digits may still arrive (e.g. "108." -> "108.2"); |
| 198 | # withhold to avoid corrupting the streaming diff. |
| 199 | break |
| 200 | result[key] = raw_val |
| 201 | |
| 202 | return result |
| 203 | |
| 204 | |
| 205 | def _parse_gemma4_array(arr_str: str, *, partial: bool = False) -> list: |
| 206 | items: list = [] |
| 207 | i = 0 |
| 208 | n = len(arr_str) |
| 209 | |
| 210 | while i < n: |
| 211 | while i < n and arr_str[i] in (" ", ",", "\n", "\t"): |
| 212 | i += 1 |
| 213 | if i >= n: |
| 214 | break |
| 215 | |
| 216 | if arr_str.startswith(STRING_DELIM, i): |
| 217 | i += _DELIM_LEN |
| 218 | end_pos = arr_str.find(STRING_DELIM, i) |
| 219 | if end_pos == -1: |
| 220 | items.append(arr_str[i:]) |
| 221 | break |
| 222 | items.append(arr_str[i:end_pos]) |
| 223 | i = end_pos + _DELIM_LEN |
| 224 | |
| 225 | elif arr_str[i] == "{": |
| 226 | depth = 1 |
| 227 | obj_start = i + 1 |
| 228 | i += 1 |
| 229 | while i < n and depth > 0: |
| 230 | if arr_str.startswith(STRING_DELIM, i): |
| 231 | i += _DELIM_LEN |
| 232 | nd = arr_str.find(STRING_DELIM, i) |
| 233 | i = nd + _DELIM_LEN if nd != -1 else n |
| 234 | continue |
| 235 | if arr_str[i] == "{": |
| 236 | depth += 1 |
| 237 | elif arr_str[i] == "}": |
| 238 | depth -= 1 |
| 239 | i += 1 |
| 240 | if depth > 0: |
| 241 | items.append(_parse_gemma4_args(arr_str[obj_start:i], partial=True)) |
| 242 | else: |
| 243 | items.append(_parse_gemma4_args(arr_str[obj_start : i - 1])) |
| 244 | |
| 245 | elif arr_str[i] == "[": |
| 246 | depth = 1 |
| 247 | sub_start = i + 1 |
| 248 | i += 1 |
| 249 | while i < n and depth > 0: |
| 250 | if arr_str.startswith(STRING_DELIM, i): |
| 251 | i += _DELIM_LEN |
| 252 | nd = arr_str.find(STRING_DELIM, i) |
| 253 | i = nd + _DELIM_LEN if nd != -1 else n |
| 254 | continue |
| 255 | if arr_str[i] == "[": |
| 256 | depth += 1 |
| 257 | elif arr_str[i] == "]": |
| 258 | depth -= 1 |
| 259 | i += 1 |
| 260 | if depth > 0: |
| 261 | items.append(_parse_gemma4_array(arr_str[sub_start:i], partial=True)) |
| 262 | else: |
| 263 | items.append(_parse_gemma4_array(arr_str[sub_start : i - 1])) |
| 264 | |
| 265 | else: |
| 266 | val_start = i |
| 267 | while i < n and arr_str[i] not in (",", "]"): |
| 268 | i += 1 |
| 269 | if partial and i >= n: |
| 270 | break |
| 271 | if i == val_start: |
| 272 | logger.warning( |
| 273 | "Gemma4 array parser made no progress at position %d; " |
| 274 | "aborting on malformed input.", |
| 275 | i, |
| 276 | ) |
| 277 | break |
| 278 | raw_val = arr_str[val_start:i].strip() |
| 279 | if partial and raw_val.endswith("."): |
| 280 | break |
| 281 | items.append(raw_val) |
| 282 | |
| 283 | return items |
| 284 | |
| 285 | |
| 286 | def _gemma4_arg_converter(raw_args: str, partial: bool) -> str: |
| 287 | """Convert Gemma4 custom arg format to a JSON string.""" |
| 288 | text = raw_args.strip() |
| 289 | if text.endswith("}") or text.endswith(")") and text.count("(") < text.count(")"): |
| 290 | text = text[:-1] |
| 291 | |
| 292 | parsed = _parse_gemma4_args(text, partial=partial) |
| 293 | return json.dumps(parsed, ensure_ascii=False) |
| 294 | |
| 295 | |
| 296 | @functools.cache |
| 297 | def gemma4_config(thinking: bool = False) -> ParserEngineConfig: |
| 298 | return ParserEngineConfig( |
| 299 | name="gemma4", |
| 300 | initial_state=ParserState.CONTENT, |
| 301 | # Decides a silent tail: the E-series HF templates write no marker |
| 302 | # with thinking off, and no template does after a tool response. |
| 303 | wait_for_reasoning=thinking, |
| 304 | turn_boundary_tokens=frozenset({"<|turn>", "<|tool_response>"}), |
| 305 | terminals={ |
| 306 | "THINK_START": CHANNEL_START, |
| 307 | "THINK_END": CHANNEL_END, |
| 308 | "TOOL_START": TOOL_CALL_START, |
| 309 | "TOOL_END": TOOL_CALL_END, |
| 310 | "CALL_PREFIX": "call:", |
| 311 | "COLON": ":", |
| 312 | "OPEN_BRACE": "{", |
| 313 | "OPEN_PAREN": "(", |
| 314 | }, |
| 315 | token_id_terminals={ |
| 316 | "THINK_START": CHANNEL_START, |
| 317 | "THINK_END": CHANNEL_END, |
| 318 | "TOOL_START": TOOL_CALL_START, |
| 319 | "TOOL_END": TOOL_CALL_END, |
| 320 | }, |
| 321 | transitions={ |
| 322 | # -- Reasoning transitions -- |
| 323 | (ParserState.CONTENT, "THINK_START"): Transition( |
| 324 | ParserState.REASONING, |
| 325 | (EventType.REASONING_START,), |
| 326 | ), |
| 327 | # No-op: if we pre-initialised the engine to REASONING from the |
| 328 | # prompt (see ``adjust_initial_state_from_prompt``) but the model |
| 329 | # still emits its own ``<|channel>`` opener, swallow it instead |
| 330 | # of leaking it as TEXT_CHUNK. |
| 331 | (ParserState.REASONING, "THINK_START"): Transition( |
| 332 | ParserState.REASONING, |
| 333 | (), |
| 334 | ), |
| 335 | (ParserState.REASONING, "THINK_END"): Transition( |
| 336 | ParserState.CONTENT, |
| 337 | (EventType.REASONING_END,), |
| 338 | ), |
| 339 | # Tool call directly from reasoning (no explicit <channel|>) |
| 340 | (ParserState.REASONING, "TOOL_START"): Transition( |
| 341 | ParserState.TOOL_PREAMBLE, |
| 342 | (EventType.REASONING_END, EventType.TOOL_CALL_START), |
| 343 | ), |
| 344 | # -- Tool call transitions -- |
| 345 | (ParserState.CONTENT, "TOOL_START"): Transition( |
| 346 | ParserState.TOOL_PREAMBLE, |
| 347 | (EventType.REASONING_END, EventType.TOOL_CALL_START), |
| 348 | ), |
| 349 | (ParserState.TOOL_PREAMBLE, "TOOL_END"): Transition( |
| 350 | ParserState.CONTENT, |
| 351 | (EventType.TOOL_CALL_END,), |
| 352 | ), |
| 353 | (ParserState.TOOL_PREAMBLE, "CALL_PREFIX"): Transition( |
| 354 | ParserState.TOOL_NAME, |
| 355 | (), |
| 356 | ), |
| 357 | # Bare opener: some checkpoints emit "<|tool_call>:name{...}" |
| 358 | # without the "call" prefix. Match the colon directly so the |
| 359 | # tool call is not silently dropped. |
| 360 | (ParserState.TOOL_PREAMBLE, "COLON"): Transition( |
| 361 | ParserState.TOOL_NAME, |
| 362 | (), |
| 363 | ), |
| 364 | (ParserState.TOOL_NAME, "OPEN_BRACE"): Transition( |
| 365 | ParserState.TOOL_ARGS, |
| 366 | (), |
| 367 | ), |
| 368 | (ParserState.TOOL_NAME, "OPEN_PAREN"): Transition( |
| 369 | ParserState.TOOL_ARGS, |
| 370 | (), |
| 371 | ), |
| 372 | (ParserState.TOOL_NAME, "TOOL_END"): Transition( |
| 373 | ParserState.CONTENT, |
| 374 | (EventType.TOOL_CALL_END,), |
| 375 | ), |
| 376 | (ParserState.TOOL_ARGS, "TOOL_END"): Transition( |
| 377 | ParserState.CONTENT, |
| 378 | (EventType.TOOL_CALL_END,), |
| 379 | ), |
| 380 | # Back-to-back tool calls |
| 381 | (ParserState.CONTENT, "TOOL_END"): Transition( |
| 382 | ParserState.CONTENT, |
| 383 | (), |
| 384 | ), |
| 385 | # Absorb a bare <channel|> that arrives after we already |
| 386 | # returned to CONTENT; prevents leaking it as TEXT_CHUNK. |
| 387 | (ParserState.CONTENT, "THINK_END"): Transition( |
| 388 | ParserState.CONTENT, |
| 389 | (), |
| 390 | ), |
| 391 | }, |
| 392 | content_events={ |
| 393 | ParserState.CONTENT: EventType.TEXT_CHUNK, |
| 394 | ParserState.REASONING: EventType.REASONING_CHUNK, |
| 395 | ParserState.TOOL_NAME: EventType.TOOL_NAME, |
| 396 | ParserState.TOOL_ARGS: EventType.ARG_VALUE_CHUNK, |
| 397 | }, |
| 398 | arg_converter=_gemma4_arg_converter, |
| 399 | tool_args_json=False, |
| 400 | arg_structural_chars=frozenset(",:{}[]<"), |
| 401 | preserve_tokens=frozenset({STRING_DELIM}), |
| 402 | ) |
| 403 | |
| 404 | |
| 405 | _GEMMA4_THOUGHT_PREFIX = "thought\n" |
| 406 | _GEMMA4_THOUGHT_TOKEN = "thought" |
| 407 | |
| 408 | |
| 409 | class Gemma4Parser(ParserEngine): |
| 410 | """Gemma4 parser: ``<|channel>`` reasoning + ``<|tool_call>`` |
| 411 | tool calls in a single engine. |
| 412 | |
| 413 | - Strips the ``thought\\n`` prefix from reasoning content |
| 414 | - Sets ``skip_special_tokens=False`` so boundary tokens are visible |
| 415 | - Detects ``<|tool_call>`` token as implicit reasoning end |
| 416 | """ |
| 417 | |
| 418 | def __init__( |
| 419 | self, |
| 420 | tokenizer: TokenizerLike, |
| 421 | tools: list[Tool] | None = None, |
| 422 | **kwargs, |
| 423 | ) -> None: |
| 424 | chat_kwargs = kwargs.get("chat_template_kwargs", {}) or {} |
| 425 | thinking = chat_kwargs.get("enable_thinking", False) |
| 426 | super().__init__( |
| 427 | tokenizer, |
| 428 | tools, |
| 429 | parser_engine_config=gemma4_config(thinking=thinking), |
| 430 | **kwargs, |
| 431 | ) |
| 432 | vocab = self.vocab |
| 433 | self._tool_call_token_id: int | None = vocab.get("<|tool_call>") |
| 434 | self._new_turn_token_id: int | None = vocab.get("<|turn>") |
| 435 | self._tool_response_token_id: int | None = vocab.get("<|tool_response>") |
| 436 | self._reasoning_text: str = "" |
| 437 | self._prefix_stripped: bool = False |
| 438 | self._is_first_feed: bool = True |
| 439 | |
| 440 | def _reset(self, initial_state=None) -> None: |
| 441 | super()._reset(initial_state=initial_state) |
| 442 | self._reasoning_text = "" |
| 443 | self._prefix_stripped = False |
| 444 | self._is_first_feed = True |
| 445 | |
| 446 | def _preprocess_feed( |
| 447 | self, |
| 448 | delta_text: str, |
| 449 | delta_token_ids: Sequence[int], |
| 450 | ) -> tuple[str, Sequence[int]]: |
| 451 | if not self._is_first_feed: |
| 452 | return delta_text, delta_token_ids |
| 453 | self._is_first_feed = False |
| 454 | |
| 455 | if ( |
| 456 | not delta_text |
| 457 | or self._engine.state != ParserState.CONTENT |
| 458 | or self._reasoning_start_token_id is None |
| 459 | or self._reasoning_end_token_id is None |
| 460 | ): |
| 461 | return delta_text, delta_token_ids |
| 462 | |
| 463 | if CHANNEL_START in delta_text: |
| 464 | return delta_text, delta_token_ids |
| 465 | |
| 466 | needs_injection = ( |
| 467 | CHANNEL_END in delta_text |
| 468 | or delta_text.startswith(_GEMMA4_THOUGHT_PREFIX) |
| 469 | or delta_text == _GEMMA4_THOUGHT_TOKEN |
| 470 | ) |
| 471 | if not needs_injection: |
| 472 | return delta_text, delta_token_ids |
| 473 | |
| 474 | delta_text = CHANNEL_START + delta_text |
| 475 | if delta_token_ids: |
| 476 | delta_token_ids = [self._reasoning_start_token_id, *delta_token_ids] |
| 477 | |
| 478 | return delta_text, delta_token_ids |
| 479 | |
| 480 | def _prompt_ends_in_open_reasoning(self, prompt_token_ids: Sequence[int]) -> bool: |
| 481 | """Whether the prompt tail is inside an open ``<|channel>`` block. |
| 482 | |
| 483 | Scans backwards: a ``<|channel>`` start token seen before any |
| 484 | closing or turn-boundary token means the block is still open. |
| 485 | """ |
| 486 | start_id = self._reasoning_start_token_id |
| 487 | if start_id is None: |
| 488 | return False |
| 489 | boundary_ids = { |
| 490 | tid |
| 491 | for tid in ( |
| 492 | self._reasoning_end_token_id, |
| 493 | self._tool_call_token_id, |
| 494 | self._new_turn_token_id, |
| 495 | self._tool_response_token_id, |
| 496 | ) |
| 497 | if tid is not None |
| 498 | } |
| 499 | for tid in reversed(prompt_token_ids): |
| 500 | if tid == start_id: |
| 501 | return True |
| 502 | if tid in boundary_ids: |
| 503 | return False |
| 504 | return False |
| 505 | |
| 506 | def adjust_initial_state_from_prompt(self, prompt_token_ids: Sequence[int]) -> None: |
| 507 | """Pre-initialise the engine to ``REASONING`` when the prompt ends |
| 508 | inside an open ``<|channel>`` block. |
| 509 | |
| 510 | This covers the post-tool-response continuation case where the chat |
| 511 | template leaves the prompt ending with ``<|channel>thought\n`` |
| 512 | (issue #45834). A prompt that merely starts a new model turn must |
| 513 | not pre-initialise reasoning: the model may answer directly without |
| 514 | emitting any channel markers, and the non-streaming path classifies |
| 515 | such output as content (issue #48217). When the model does open its |
| 516 | own ``<|channel>``, the ``(CONTENT, THINK_START)`` transition |
| 517 | handles it, and the ``thought\n`` prefix in the first reasoning |
| 518 | chunk is stripped by ``_events_to_delta`` as in the default flow. |
| 519 | """ |
| 520 | if not self._prompt_ends_in_open_reasoning(prompt_token_ids): |
| 521 | return |
| 522 | self._engine.reset(initial_state=ParserState.REASONING) |
| 523 | # Prevent a later default ``initialize_streaming()`` (e.g. from |
| 524 | # ``ParserEngineReasoningAdapter.extract_reasoning_streaming``) from |
| 525 | # clobbering this with ``CONTENT``. |
| 526 | self._streaming_initialized = True |
| 527 | |
| 528 | def _events_to_delta( |
| 529 | self, |
| 530 | events: list[SemanticEvent], |
| 531 | finished: bool = False, |
| 532 | ) -> DeltaMessage | None: |
| 533 | delta = super()._events_to_delta(events, finished=finished) |
| 534 | if delta is None or delta.reasoning is None: |
| 535 | return delta |
| 536 | |
| 537 | if self._prefix_stripped: |
| 538 | return delta |
| 539 | self._reasoning_text += delta.reasoning |
| 540 | |
| 541 | if self._reasoning_text.startswith(_GEMMA4_THOUGHT_PREFIX): |
| 542 | prefix_len = len(_GEMMA4_THOUGHT_PREFIX) |
| 543 | prev_reasoning_len = len(self._reasoning_text) - len(delta.reasoning) |
| 544 | if prev_reasoning_len >= prefix_len: |
| 545 | self._prefix_stripped = True |
| 546 | return delta |
| 547 | chars_of_prefix_in_delta = prefix_len - prev_reasoning_len |
| 548 | stripped = delta.reasoning[chars_of_prefix_in_delta:] |
| 549 | if stripped: |
| 550 | self._prefix_stripped = True |
| 551 | delta.reasoning = stripped |
| 552 | return delta |
| 553 | if len(self._reasoning_text) >= prefix_len: |
| 554 | self._prefix_stripped = True |
| 555 | delta.reasoning = None |
| 556 | if delta.content is not None or delta.tool_calls: |
| 557 | return delta |
| 558 | return None |
| 559 | return None |
| 560 | |
| 561 | if _GEMMA4_THOUGHT_PREFIX.startswith(self._reasoning_text): |
| 562 | if finished: |
| 563 | self._prefix_stripped = True |
| 564 | return None |
| 565 | |
| 566 | self._prefix_stripped = True |
| 567 | delta.reasoning = self._reasoning_text |
| 568 | return delta |
| 569 | |
| 570 | def extract_reasoning( |
| 571 | self, |
| 572 | model_output: str, |
| 573 | request: ChatCompletionRequest | ResponsesRequest, |
| 574 | ) -> tuple[str | None, str | None]: |
| 575 | reasoning, content = super().extract_reasoning(model_output, request) |
| 576 | if reasoning: |
| 577 | if reasoning.startswith(_GEMMA4_THOUGHT_PREFIX): |
| 578 | reasoning = reasoning[len(_GEMMA4_THOUGHT_PREFIX) :] |
| 579 | elif reasoning == _GEMMA4_THOUGHT_PREFIX.rstrip(): |
| 580 | reasoning = None |
| 581 | return reasoning or None, content |