| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | from collections.abc import Sequence |
| 4 | from functools import cached_property |
| 5 | from pathlib import Path |
| 6 | from typing import TYPE_CHECKING, Any, cast, overload |
| 7 | |
| 8 | from mistral_common.guidance.grammar_factory import GrammarFactory |
| 9 | from mistral_common.guidance.tokenizer import from_mistral_tokenizer |
| 10 | from mistral_common.protocol.instruct.request import ( |
| 11 | ChatCompletionRequest as MistralChatCompletionRequest, |
| 12 | ) |
| 13 | from mistral_common.protocol.instruct.request import ( |
| 14 | ReasoningEffort, |
| 15 | ) |
| 16 | from mistral_common.protocol.instruct.validator import ValidationMode |
| 17 | from mistral_common.tokens.tokenizers.base import ( |
| 18 | SpecialTokenPolicy, |
| 19 | SpecialTokens, |
| 20 | Tokenizer, |
| 21 | ) |
| 22 | from mistral_common.tokens.tokenizers.instruct import ( |
| 23 | InstructTokenizerBase, |
| 24 | ) |
| 25 | from mistral_common.tokens.tokenizers.mistral import ( |
| 26 | MistralTokenizer as MistralCommonTokenizer, |
| 27 | ) |
| 28 | from mistral_common.tokens.tokenizers.sentencepiece import ( |
| 29 | SentencePieceTokenizer, |
| 30 | ) |
| 31 | from mistral_common.tokens.tokenizers.tekken import Tekkenizer |
| 32 | from pydantic import ValidationError |
| 33 | from transformers.tokenization_mistral_common import MistralCommonBackend |
| 34 | |
| 35 | from vllm.entrypoints.chat_utils import ChatCompletionMessageParam |
| 36 | from vllm.entrypoints.openai.chat_completion.protocol import ChatCompletionRequest |
| 37 | from vllm.exceptions import VLLMValidationError |
| 38 | from vllm.logger import init_logger |
| 39 | from vllm.tokenizers.protocol import TokenizerLike |
| 40 | |
| 41 | if TYPE_CHECKING: |
| 42 | import llguidance |
| 43 | from transformers import BatchEncoding |
| 44 | |
| 45 | logger = init_logger(__name__) |
| 46 | |
| 47 | |
| 48 | def maybe_serialize_tool_calls(request: "MistralChatCompletionRequest"): |
| 49 | # SEE: https://github.com/vllm-project/vllm/pull/9951 |
| 50 | # Credits go to: @gcalmettes |
| 51 | # NOTE: There is currently a bug in pydantic where attributes |
| 52 | # declared as iterables are replaced in the instances by |
| 53 | # pydantic-core ValidatorIterator instance. In particular, this |
| 54 | # affects tool_calls defined in ChatCompletionAssistantMessageParam |
| 55 | # model: |
| 56 | # see: |
| 57 | # - https://github.com/pydantic/pydantic/issues/9467 |
| 58 | # As a result, tool_calls from assistant messages are never |
| 59 | # deserialized in the request object if the tool_calls iterator is |
| 60 | # not consumed. This affect messages passed to the MistralTokenizer |
| 61 | # since no chat template is applied and therefore the tools_calls |
| 62 | # iterator is not directly consumed. |
| 63 | # Issue is tracked on Pydantic side, with resolution planned for |
| 64 | # v2.11 release. In the meantime, the official workaround is to |
| 65 | # consume the iterator so the tool_calls are correctly deserialized |
| 66 | # in the OpenAI ChatCompletionAssistantMessageParam object |
| 67 | # https://github.com/pydantic/pydantic/issues/9467#issuecomment-2442097291 # noqa: E501 |
| 68 | # Official Pydantic Issues: |
| 69 | # - https://github.com/pydantic/pydantic/issues/9541 |
| 70 | # TODO: remove when pydantic v2.11 is released |
| 71 | for i, message in enumerate(request.messages): |
| 72 | if message.get("role") == "assistant": |
| 73 | if (tool_calls_validator := message.get("tool_calls", None)) is not None: |
| 74 | try: |
| 75 | validated_tool_calls = list(tool_calls_validator) |
| 76 | except ValidationError as e: |
| 77 | raise ValueError( |
| 78 | "Validating messages' `tool_calls` raised an error. " |
| 79 | "Please ensure `tool_calls` are iterable of tool calls." |
| 80 | ) from e |
| 81 | else: |
| 82 | validated_tool_calls = [] |
| 83 | |
| 84 | request.messages[i]["tool_calls"] = validated_tool_calls |
| 85 | |
| 86 | |
| 87 | def truncate_tool_call_ids(request: "MistralChatCompletionRequest"): |
| 88 | """Truncates tool call IDs for Mistral's ID requirements.""" |
| 89 | for i, message in enumerate(request.messages): |
| 90 | if message.get("role") == "assistant": |
| 91 | tool_calls = message.get("tool_calls", []) |
| 92 | for tool_call in tool_calls: |
| 93 | if len(tool_call["id"]) > 9: |
| 94 | logger.warning( |
| 95 | "Truncating tool call ID: %s to %s", |
| 96 | tool_call["id"], |
| 97 | tool_call["id"][-9:], |
| 98 | ) |
| 99 | tool_call["id"] = tool_call["id"][-9:] |
| 100 | |
| 101 | request.messages[i]["tool_calls"] = tool_calls |
| 102 | |
| 103 | elif message.get("role") in {"tool_results", "tool"}: |
| 104 | if "tool_call_id" in message: |
| 105 | tool_call_id = message["tool_call_id"] |
| 106 | |
| 107 | if len(tool_call_id) > 9: |
| 108 | logger.warning( |
| 109 | "Truncating tool_call_id: %s to %s", |
| 110 | tool_call_id, |
| 111 | tool_call_id[-9:], |
| 112 | ) |
| 113 | tool_call_id = tool_call_id[-9:] |
| 114 | request.messages[i]["tool_call_id"] = tool_call_id |
| 115 | |
| 116 | |
| 117 | def _validate_apply_chat_template_args( |
| 118 | messages: list["ChatCompletionMessageParam"], |
| 119 | continue_final_message: bool = False, |
| 120 | add_generation_prompt: bool = False, |
| 121 | ) -> None: |
| 122 | if add_generation_prompt and continue_final_message: |
| 123 | raise ValueError( |
| 124 | "Cannot set both `add_generation_prompt` and " |
| 125 | "`continue_final_message` to True." |
| 126 | ) |
| 127 | |
| 128 | last_message = cast(dict[str, Any], messages[-1]) |
| 129 | # add_generation_prompt is directly handled by the tokenizer but we |
| 130 | # check if the user is trying to use it with a final assistant message |
| 131 | # which is probably not what they want. |
| 132 | # If add_generation_prompt is False, we don't need to check anything. |
| 133 | if add_generation_prompt and last_message["role"] == "assistant": |
| 134 | raise ValueError( |
| 135 | "Cannot set `add_generation_prompt` to True when " |
| 136 | "the last message is from the assistant. Consider " |
| 137 | "using `continue_final_message` instead." |
| 138 | ) |
| 139 | if continue_final_message and last_message["role"] != "assistant": |
| 140 | raise ValueError( |
| 141 | "Cannot set `continue_final_message` to True when " |
| 142 | "the last message is not from the assistant." |
| 143 | ) |
| 144 | |
| 145 | |
| 146 | def validate_request_params(request: "ChatCompletionRequest"): |
| 147 | if request.chat_template is not None or request.chat_template_kwargs is not None: |
| 148 | raise VLLMValidationError( |
| 149 | "chat_template is not supported for Mistral tokenizers.", |
| 150 | parameter="chat_template", |
| 151 | ) |
| 152 | |
| 153 | if request.reasoning_effort and request.reasoning_effort not in list( |
| 154 | ReasoningEffort |
| 155 | ): |
| 156 | raise VLLMValidationError( |
| 157 | f"reasoning_effort={request.reasoning_effort} is not supported by " |
| 158 | "Mistral models. Supported values are: " |
| 159 | f"{[e.value for e in ReasoningEffort]}.", |
| 160 | parameter="reasoning_effort", |
| 161 | ) |
| 162 | |
| 163 | |
| 164 | def _tekken_token_to_id(tokenizer: "Tekkenizer", t: str | bytes) -> int: |
| 165 | assert isinstance(tokenizer, Tekkenizer), type(tokenizer) |
| 166 | |
| 167 | t_bytes = t.encode("utf-8") if not isinstance(t, bytes) else t |
| 168 | shift = tokenizer.num_special_tokens |
| 169 | try: |
| 170 | return shift + tokenizer._tekken_token2id_nospecial[t_bytes] |
| 171 | except KeyError: |
| 172 | t_str = t_bytes.decode("utf-8") |
| 173 | if t_str in tokenizer._special_tokens_reverse_vocab: |
| 174 | return tokenizer._special_tokens_reverse_vocab[t_str] |
| 175 | logger.warning( |
| 176 | "Failed to convert token %s to id, replacing with <unk>", t_bytes |
| 177 | ) |
| 178 | return tokenizer.unk_id |
| 179 | |
| 180 | |
| 181 | def mistral_common_tekkenizer(tokenizer: object) -> "Tekkenizer | None": |
| 182 | """Return the underlying `Tekkenizer` for a `MistralCommonBackend`.""" |
| 183 | mistral = getattr(tokenizer, "tokenizer", None) |
| 184 | instruct = getattr(mistral, "instruct_tokenizer", None) |
| 185 | tekken = getattr(instruct, "tokenizer", None) |
| 186 | return tekken if isinstance(tekken, Tekkenizer) else None |
| 187 | |
| 188 | |
| 189 | def tekken_convert_ids_to_tokens( |
| 190 | tokenizer: "Tekkenizer", ids: Sequence[int] |
| 191 | ) -> list[str | bytes]: |
| 192 | """Convert ids to pieces, using raw `bytes` for byte-fallback tokens.""" |
| 193 | tokens: list[str | bytes] = [tokenizer.id_to_piece(i) for i in ids] |
| 194 | if any("�" in t for t in tokens): |
| 195 | tokens = [ |
| 196 | tokenizer.id_to_byte_piece(i, SpecialTokenPolicy.KEEP) |
| 197 | if i >= tokenizer.num_special_tokens |
| 198 | else tokenizer.decode([i], SpecialTokenPolicy.KEEP) |
| 199 | for i in ids |
| 200 | ] |
| 201 | return tokens |
| 202 | |
| 203 | |
| 204 | def tekken_convert_tokens_to_string( |
| 205 | tokenizer: "Tekkenizer", tokens: Sequence[str | bytes] |
| 206 | ) -> str: |
| 207 | """Reassemble pieces from `tekken_convert_ids_to_tokens` into text.""" |
| 208 | if any(isinstance(t, bytes) for t in tokens): |
| 209 | ids = [_tekken_token_to_id(tokenizer, t) for t in tokens] |
| 210 | return tokenizer.decode(ids, SpecialTokenPolicy.KEEP) |
| 211 | return "".join(cast(Sequence[str], tokens)) |
| 212 | |
| 213 | |
| 214 | class MistralTokenizer(TokenizerLike): |
| 215 | IS_MISTRAL_TOKENIZER = True # used by vllm.utils.mistral |
| 216 | |
| 217 | @classmethod |
| 218 | def from_pretrained( |
| 219 | cls, |
| 220 | path_or_repo_id: str | Path, |
| 221 | *args, |
| 222 | trust_remote_code: bool = False, |
| 223 | revision: str | None = None, |
| 224 | download_dir: str | None = None, |
| 225 | **kwargs, |
| 226 | ) -> "MistralTokenizer": |
| 227 | tokenizer = MistralCommonBackend.from_pretrained( |
| 228 | path_or_repo_id, |
| 229 | *args, |
| 230 | mode=ValidationMode.test, |
| 231 | cache_dir=download_dir, |
| 232 | revision="main" if revision is None else revision, |
| 233 | **kwargs, |
| 234 | ) |
| 235 | |
| 236 | return cls(tokenizer) |
| 237 | |
| 238 | def __init__(self, tokenizer: MistralCommonBackend) -> None: |
| 239 | super().__init__() |
| 240 | |
| 241 | self.transformers_tokenizer: MistralCommonBackend = tokenizer |
| 242 | self.mistral: MistralCommonTokenizer = tokenizer.tokenizer |
| 243 | self.instruct: InstructTokenizerBase = self.mistral.instruct_tokenizer |
| 244 | self.tokenizer: Tokenizer = self.instruct.tokenizer |
| 245 | |
| 246 | mode = self.mistral._chat_completion_request_validator._mode |
| 247 | if mode != ValidationMode.test: |
| 248 | raise ValueError( |
| 249 | "Mistral tokenizer must be in test mode. Make sure to " |
| 250 | "set `mode='ValidationMode.test'` when creating the " |
| 251 | "Mistral tokenizer." |
| 252 | ) |
| 253 | |
| 254 | _mistral_version_str = str(self.tokenizer.version.value) |
| 255 | self.version: int = int(_mistral_version_str.split("v")[-1]) |
| 256 | |
| 257 | self.is_tekken = isinstance(self.tokenizer, Tekkenizer) |
| 258 | self.is_spm = isinstance(self.tokenizer, SentencePieceTokenizer) |
| 259 | if not (self.is_tekken or self.is_spm): |
| 260 | raise TypeError(f"Unsupported tokenizer: {type(self.tokenizer)}") |
| 261 | |
| 262 | # Reverse order to ensure that the lowest token id is kept. |
| 263 | self._vocab_dict = { |
| 264 | self.convert_ids_to_tokens([i], skip_special_tokens=False)[0]: i |
| 265 | for i in range(self.vocab_size - 1, -1, -1) |
| 266 | } |
| 267 | # Sort the dict for convenience |
| 268 | self._vocab_dict = dict(sorted(self._vocab_dict.items(), key=lambda x: x[1])) |
| 269 | |
| 270 | # Vocab sorted by token id. |
| 271 | self._vocab = self.tokenizer.vocab() |
| 272 | self._max_token_id = self.vocab_size - 1 |
| 273 | self._max_chars_per_token = max(len(tok) for tok in self._vocab) |
| 274 | |
| 275 | # Cache special tokens for faster access. |
| 276 | self._special_token_ids = self._get_special_token_ids() |
| 277 | self._special_token_ids_set = set(self._special_token_ids) |
| 278 | self._special_tokens = self._get_special_tokens(self._special_token_ids) |
| 279 | self._special_tokens_set = set(self._special_tokens) |
| 280 | |
| 281 | def _get_special_token_ids(self) -> list[int]: |
| 282 | return [i for i in range(len(self._vocab)) if self.tokenizer.is_special(i)] |
| 283 | |
| 284 | def _get_special_tokens(self, all_special_ids: list[int]) -> list[str]: |
| 285 | return [ |
| 286 | self.tokenizer.decode([i], special_token_policy=SpecialTokenPolicy.KEEP) |
| 287 | for i in all_special_ids |
| 288 | ] |
| 289 | |
| 290 | def num_special_tokens_to_add(self) -> int: |
| 291 | return len(self.encode("")) |
| 292 | |
| 293 | # the following attributes are set to fit vLLM's design and are used |
| 294 | # by the structured output backends. |
| 295 | @property |
| 296 | def all_special_tokens(self) -> list[str]: |
| 297 | return self._special_tokens |
| 298 | |
| 299 | @property |
| 300 | def all_special_ids(self) -> list[int]: |
| 301 | return self._special_token_ids |
| 302 | |
| 303 | @property |
| 304 | def bos_token_id(self) -> int: |
| 305 | return self.tokenizer.bos_id |
| 306 | |
| 307 | @property |
| 308 | def eos_token_id(self) -> int: |
| 309 | return self.tokenizer.eos_id |
| 310 | |
| 311 | @property |
| 312 | def pad_token_id(self) -> int: |
| 313 | return self.tokenizer.pad_id |
| 314 | |
| 315 | @property |
| 316 | def is_fast(self) -> bool: |
| 317 | return True |
| 318 | |
| 319 | @property |
| 320 | def vocab_size(self) -> int: |
| 321 | return self.transformers_tokenizer.vocab_size |
| 322 | |
| 323 | @property |
| 324 | def max_token_id(self) -> int: |
| 325 | return self._max_token_id |
| 326 | |
| 327 | @property |
| 328 | def max_chars_per_token(self) -> int: |
| 329 | return self._max_chars_per_token |
| 330 | |
| 331 | @property |
| 332 | def truncation_side(self) -> str: |
| 333 | return self.transformers_tokenizer.truncation_side |
| 334 | |
| 335 | def _is_special_token_id(self, token_id: int) -> bool: |
| 336 | return token_id in self._special_token_ids_set |
| 337 | |
| 338 | def __hash__(self) -> int: |
| 339 | return hash(id(self)) |
| 340 | |
| 341 | def __len__(self) -> int: |
| 342 | return self.vocab_size |
| 343 | |
| 344 | def __call__( |
| 345 | self, |
| 346 | text: str | list[str], |
| 347 | text_pair: str | None = None, |
| 348 | add_special_tokens: bool = True, |
| 349 | truncation: bool = False, |
| 350 | max_length: int | None = None, |
| 351 | ) -> "BatchEncoding": |
| 352 | if text_pair is not None: |
| 353 | raise ValueError( |
| 354 | "`text_pair` is not supported by `MistralTokenizer.__call__`." |
| 355 | ) |
| 356 | |
| 357 | return self.transformers_tokenizer( |
| 358 | text=text, |
| 359 | text_pair=text_pair, |
| 360 | add_special_tokens=add_special_tokens, |
| 361 | truncation=truncation, |
| 362 | max_length=max_length, |
| 363 | ) |
| 364 | |
| 365 | @property |
| 366 | def vocab(self) -> list[str]: |
| 367 | return self._vocab |
| 368 | |
| 369 | def get_vocab(self) -> dict[str, int]: |
| 370 | return self._vocab_dict |
| 371 | |
| 372 | def get_added_vocab(self) -> dict[str, int]: |
| 373 | # Mistral tokenizers have no added vocabulary |
| 374 | return {} |
| 375 | |
| 376 | def encode( |
| 377 | self, |
| 378 | text: str, |
| 379 | truncation: bool | None = None, |
| 380 | max_length: int | None = None, |
| 381 | add_special_tokens: bool = True, |
| 382 | ) -> list[int]: |
| 383 | # NOTE: transformers guards truncation with `if max_length and ...`, so |
| 384 | # `max_length=0` returns the full sequence instead of an empty one. vLLM |
| 385 | # treats `truncate_prompt_tokens=0` as an empty prompt, so keep slicing |
| 386 | # here until that is fixed upstream. |
| 387 | encoded = self.tokenizer.encode(text, bos=add_special_tokens, eos=False) |
| 388 | |
| 389 | if truncation is not False and max_length is not None: |
| 390 | return encoded[:max_length] |
| 391 | else: |
| 392 | return encoded |
| 393 | |
| 394 | def apply_chat_template( |
| 395 | self, |
| 396 | messages: list["ChatCompletionMessageParam"], |
| 397 | tools: list[dict[str, Any]] | None = None, |
| 398 | **kwargs, |
| 399 | ) -> list[int]: |
| 400 | add_generation_prompt = kwargs.pop("add_generation_prompt", False) |
| 401 | continue_final_message = kwargs.get("continue_final_message", False) |
| 402 | tokenize = kwargs.get("tokenize", True) |
| 403 | padding = kwargs.get("padding", False) |
| 404 | truncation = kwargs.get("truncation", False) |
| 405 | max_length = kwargs.get("max_length") |
| 406 | |
| 407 | version_kwargs = {} |
| 408 | # NOTE: This is for backward compatibility. |
| 409 | # Transformers should be passed arguments it knows. |
| 410 | if self.version >= 15: |
| 411 | version_kwargs["reasoning_effort"] = kwargs.get("reasoning_effort") |
| 412 | |
| 413 | _validate_apply_chat_template_args( |
| 414 | messages, continue_final_message, add_generation_prompt |
| 415 | ) |
| 416 | |
| 417 | return self.transformers_tokenizer.apply_chat_template( |
| 418 | conversation=messages, |
| 419 | tools=tools, |
| 420 | continue_final_message=continue_final_message, |
| 421 | tokenize=tokenize, |
| 422 | padding=padding, |
| 423 | truncation=truncation, |
| 424 | max_length=max_length, |
| 425 | return_tensors=None, |
| 426 | return_dict=False, |
| 427 | **version_kwargs, |
| 428 | ) |
| 429 | |
| 430 | def decode( |
| 431 | self, ids: Sequence[int] | int, skip_special_tokens: bool = False |
| 432 | ) -> str: |
| 433 | return self.transformers_tokenizer.decode( |
| 434 | ids, skip_special_tokens=skip_special_tokens |
| 435 | ) |
| 436 | |
| 437 | def batch_decode( |
| 438 | self, ids: list[list[int]] | list[int], skip_special_tokens: bool = False |
| 439 | ) -> str: |
| 440 | return self.transformers_tokenizer.batch_decode( |
| 441 | ids, skip_special_tokens=skip_special_tokens |
| 442 | ) |
| 443 | |
| 444 | @overload |
| 445 | def convert_tokens_to_ids(self, tokens: str) -> int: ... |
| 446 | |
| 447 | @overload |
| 448 | def convert_tokens_to_ids(self, tokens: list[str]) -> list[int]: ... |
| 449 | |
| 450 | def convert_tokens_to_ids(self, tokens: str | list[str]) -> int | list[int]: |
| 451 | return self.transformers_tokenizer.convert_tokens_to_ids(tokens) |
| 452 | |
| 453 | def convert_tokens_to_string(self, tokens: list[str]) -> str: |
| 454 | to_decode_special_tokens = { |
| 455 | SpecialTokens.tool_calls, |
| 456 | SpecialTokens.args, |
| 457 | SpecialTokens.begin_think, |
| 458 | SpecialTokens.end_think, |
| 459 | } |
| 460 | if self.is_tekken: |
| 461 | assert isinstance(self.tokenizer, Tekkenizer), type(self.tokenizer) |
| 462 | tokens = [ |
| 463 | t |
| 464 | for t in tokens |
| 465 | if (t in to_decode_special_tokens or t not in self._special_tokens_set) |
| 466 | ] |
| 467 | |
| 468 | # We filtered unwanted special tokens before so we can decode the rest. |
| 469 | decoded = tekken_convert_tokens_to_string(self.tokenizer, tokens) |
| 470 | else: |
| 471 | # make sure certain special tokens like Tool calls are |
| 472 | # not decoded |
| 473 | assert isinstance(self.tokenizer, SentencePieceTokenizer), type( |
| 474 | self.tokenizer |
| 475 | ) |
| 476 | |
| 477 | regular_tokens: list[str] = [] |
| 478 | decoded_list: list[str] = [] |
| 479 | decoded = "" |
| 480 | |
| 481 | for token in tokens: |
| 482 | if token in to_decode_special_tokens: |
| 483 | if regular_tokens: |
| 484 | decoded_list.append( |
| 485 | self.tokenizer.decode( |
| 486 | regular_tokens, SpecialTokenPolicy.IGNORE |
| 487 | ) |
| 488 | ) |
| 489 | regular_tokens = [] |
| 490 | decoded_list.append(token) |
| 491 | else: |
| 492 | regular_tokens.append(token) |
| 493 | |
| 494 | if regular_tokens: |
| 495 | decoded_list.append( |
| 496 | self.tokenizer.decode(regular_tokens, SpecialTokenPolicy.IGNORE) |
| 497 | ) |
| 498 | decoded = "".join(decoded_list) |
| 499 | |
| 500 | return decoded |
| 501 | |
| 502 | def convert_ids_to_tokens( |
| 503 | self, |
| 504 | ids: Sequence[int], |
| 505 | skip_special_tokens: bool = False, |
| 506 | ) -> list[str]: |
| 507 | if not skip_special_tokens: |
| 508 | return [self.tokenizer.id_to_piece(token_id) for token_id in ids] |
| 509 | |
| 510 | non_skip_special_tokens_ids = { |
| 511 | self.tokenizer.get_special_token(SpecialTokens.tool_calls), |
| 512 | } |
| 513 | # [ARGS] only exists in v11+ tool-call tokenizers; older tokenizers |
| 514 | # raise (Tekken) or return unk (SPM) for it. |
| 515 | if self.tokenizer.is_special(SpecialTokens.args): |
| 516 | non_skip_special_tokens_ids.add( |
| 517 | self.tokenizer.get_special_token(SpecialTokens.args) |
| 518 | ) |
| 519 | # [THINK]/[/THINK] only exist in v13+ reasoning tokenizers; use the |
| 520 | # same is_special gate as [ARGS] above so newer versions are covered |
| 521 | # without an isinstance check on the instruct tokenizer. |
| 522 | for think_token in (SpecialTokens.begin_think, SpecialTokens.end_think): |
| 523 | if self.tokenizer.is_special(think_token): |
| 524 | non_skip_special_tokens_ids.add( |
| 525 | self.tokenizer.get_special_token(think_token) |
| 526 | ) |
| 527 | |
| 528 | ids_kept = [ |
| 529 | i |
| 530 | for i in ids |
| 531 | if i in non_skip_special_tokens_ids or not self._is_special_token_id(i) |
| 532 | ] |
| 533 | |
| 534 | # We filtered unwanted special tokens so we can decode the rest. |
| 535 | tokens = [self.tokenizer.id_to_piece(token_id) for token_id in ids_kept] |
| 536 | |
| 537 | if any("�" in t for t in tokens) and self.is_tekken: |
| 538 | # if a decoded token contains the replacement character, then the |
| 539 | # token has an incomplete UTF-8 character so we must use bytes |
| 540 | # See: https://github.com/vllm-project/vllm/pull/8640 |
| 541 | # https://github.com/vllm-project/vllm/pull/9625 |
| 542 | # if underlying tokenizer is sentencepiece, we just add "�". |
| 543 | # We filtered unwanted special tokens so we can decode the rest. |
| 544 | tokens = [ |
| 545 | self.tokenizer.id_to_byte_piece(token_id, SpecialTokenPolicy.KEEP) |
| 546 | if token_id not in self._special_token_ids_set |
| 547 | else self.tokenizer.decode([token_id], SpecialTokenPolicy.KEEP) |
| 548 | for token_id in ids_kept |
| 549 | ] |
| 550 | |
| 551 | return tokens |
| 552 | |
| 553 | @property |
| 554 | def supports_grammar(self) -> bool: |
| 555 | return GrammarFactory.is_supported(self.mistral) |
| 556 | |
| 557 | @cached_property |
| 558 | def grammar_factory(self) -> GrammarFactory: |
| 559 | if not self.supports_grammar: |
| 560 | raise AttributeError( |
| 561 | "This tokenizer does not support `grammar_factory`. " |
| 562 | "This is only supported for tekken tokenizers with " |
| 563 | "version >= 11." |
| 564 | ) |
| 565 | # Cache grammar factory to avoid creating a llguidance tokenizer at every usage. |
| 566 | return GrammarFactory(self.mistral) |
| 567 | |
| 568 | @cached_property |
| 569 | def llg_tokenizer(self) -> "llguidance.LLTokenizer": |
| 570 | if not self.is_tekken: |
| 571 | raise ValueError("`llg_tokenizer` is only supported for Tekkenizers.") |
| 572 | return from_mistral_tokenizer(self.mistral) |