mistral.py 20.2 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
5
import os
from pathlib import Path
6
from typing import TYPE_CHECKING, Any, Optional, Union, cast
7

8
import huggingface_hub
9
import regex as re
10
from huggingface_hub import HfApi, hf_hub_download
11
from transformers.tokenization_utils_base import BatchEncoding
12

13
from vllm.logger import init_logger
14
from vllm.transformers_utils.tokenizer_base import TokenizerBase
15
from vllm.utils import is_list_of
16

17
if TYPE_CHECKING:
18
19
20
21
22
23
24
    # make sure `mistral_common` is lazy imported,
    # so that users who only use non-mistral models
    # will not be bothered by the dependency.
    from mistral_common.protocol.instruct.request import ChatCompletionRequest
    from mistral_common.tokens.tokenizers.mistral import (
        MistralTokenizer as PublicMistralTokenizer)

25
    from vllm.entrypoints.chat_utils import ChatCompletionMessageParam
26

27
28
logger = init_logger(__name__)

29

30
def maybe_serialize_tool_calls(request: "ChatCompletionRequest"):
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
    # SEE: https://github.com/vllm-project/vllm/pull/9951
    # Credits go to: @gcalmettes
    # NOTE: There is currently a bug in pydantic where attributes
    # declared as iterables are replaced in in the instances by
    # pydantic-core ValidatorIterator instance. In particular, this
    # affects tool_calls defined in ChatCompletionAssistantMessageParam
    # model:
    # see:
    #   - https://github.com/pydantic/pydantic/issues/9467
    # As a result, tool_calls from assistant messages are never
    # deserialized in the request object if the tool_calls iterator is
    # not consumed. This affect messages passed to the MistralTokenizer
    # since no chat template is applied and therefore the tools_calls
    # iterator is not directly consumed.
    # Issue is tracked on Pydantic side, with resolution planned for
    # v2.11 release. In the meantime, the official workaround is to
    # consume the iterator so the tool_calls are correctly deserialized
    # in the OpenAI ChatCompletionAssistantMessageParam object
    # https://github.com/pydantic/pydantic/issues/9467#issuecomment-2442097291 # noqa: E501
    # Official Pydantic Issues:
    #   - https://github.com/pydantic/pydantic/issues/9541
    # TODO: remove when pydantic v2.11 is released
    for i, message in enumerate(request.messages):
        if message.get("role") == 'assistant':
            tool_calls_validator = message.get("tool_calls", ().__iter__())
            validated_tool_calls = []
            while True:
                try:
                    tool_call = next(tool_calls_validator)  # type: ignore
                    validated_tool_calls.append(tool_call)
                except StopIteration:
                    break

            request.messages[i]["tool_calls"] = validated_tool_calls


67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
def truncate_tool_call_ids(request: "ChatCompletionRequest"):
    """Truncates tool call IDs for Mistral's ID requirements."""
    for i, message in enumerate(request.messages):
        if message.get("role") == 'assistant':
            tool_calls = message.get("tool_calls", [])
            for tool_call in tool_calls:
                if len(tool_call["id"]) > 9:
                    logger.warning(
                        "Truncating tool call ID: %s to %s",
                        tool_call["id"],
                        tool_call["id"][-9:],
                    )
                    tool_call["id"] = tool_call["id"][-9:]

            request.messages[i]["tool_calls"] = tool_calls

        elif message.get("role") in {"tool_results", "tool"}:
            if "tool_call_id" in message:
                tool_call_id = message["tool_call_id"]

                if len(tool_call_id) > 9:
                    logger.warning(
                        "Truncating tool_call_id: %s to %s",
                        tool_call_id,
                        tool_call_id[-9:],
                    )
                    tool_call_id = tool_call_id[-9:]
                request.messages[i]["tool_call_id"] = tool_call_id


97
98
99
100
101
102
103
def validate_request_params(request: "ChatCompletionRequest"):
    if (request.skip_special_tokens is not None
            and not request.skip_special_tokens):
        raise ValueError("skip_special_tokens=False is not supported "
                         "for Mistral tokenizers.")


104
def list_local_repo_files(repo_id: str, revision: Optional[str]) -> list[str]:
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
    repo_cache = os.path.join(
        huggingface_hub.constants.HF_HUB_CACHE,
        huggingface_hub.constants.REPO_ID_SEPARATOR.join(
            ["models", *repo_id.split("/")]))

    if revision is None:
        revision_file = os.path.join(repo_cache, "refs", "main")
        if os.path.isfile(revision_file):
            with open(revision_file) as file:
                revision = file.read()

    if revision:
        revision_dir = os.path.join(repo_cache, "snapshots", revision)
        if os.path.isdir(revision_dir):
            return os.listdir(revision_dir)

    return []


124
def find_tokenizer_file(files: list[str]):
125
126
127
    # Accept both versioned (tokenizer.model.v3) and unversioned
    # (tokenizer.model) forms, plus tekken.json and tokenizer.mm.model
    # variants. Previous pattern only matched the versioned variants.
128
    file_pattern = re.compile(
129
130
        r"^tokenizer\.model(\.v.*)?|tekken\.json|tokenizer\.mm\.model(\.v.*)?$"
    )
131
132
133

    matched_files = [file for file in files if file_pattern.match(file)]
    if len(matched_files) > 1:
134
135
136
137
138
139
        logger.warning(
            "Multiple files matched pattern `%s`: %s. Using %s.",
            file_pattern.pattern,
            matched_files,
            matched_files[0],
        )
140
    elif len(matched_files) == 0:
141
142
143
144
        raise OSError(
            f"Found {len(matched_files)} files matching the "
            f"pattern: `{file_pattern.pattern}`. Make sure that a Mistral "
            f"tokenizer is present in {files}.")
145
146
147
148

    return matched_files[0]


Julien Denize's avatar
Julien Denize committed
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
def _aggregate_content(content: list) -> list[dict[str, Any]]:
    aggregated_content: list[dict[str, Any]] = []
    for chunk in content:
        if chunk.get("type"
                     ) == "text" and aggregated_content and aggregated_content[
                         -1].get("type") == "text":
            aggregated_content[-1]["text"] += "\n\n" + chunk.get("text")
        else:
            aggregated_content.append(chunk)
    if len(aggregated_content) == 1 and aggregated_content[0].get(
            "type") == "text":
        content = aggregated_content[0]["text"]
    return content


164
def make_mistral_chat_completion_request(
165
166
        messages: list["ChatCompletionMessageParam"],
        tools: Optional[list[dict[str,
167
                                  Any]]] = None) -> "ChatCompletionRequest":
168
    last_message = cast(dict[str, Any], messages[-1])
169
170
171
172
173
174
175
    if last_message["role"] == "assistant":
        last_message["prefix"] = True

    # mistral-common requires AssistantMessage content to be string [1].
    #
    # [1]: https://github.com/mistralai/mistral-common/blob/f4a06998b75ed78bbf5aaf569590b772ea26c9f6/src/mistral_common/protocol/instruct/messages.py#L80
    for message in messages:
176
177
178
179
180
        # Remove reasoning_content as unsupported by Mistral
        _ = message.pop("reasoning_content", None)  # type: ignore

        # Convert list text content to string
        if message.get("role") in ("assistant", "tool"):
Julien Denize's avatar
Julien Denize committed
181
            content: Any = message.get("content")
182
            if isinstance(content, list):
Julien Denize's avatar
Julien Denize committed
183
184
                content = _aggregate_content(content)
            message["content"] = content
185
186

    # The Mistral client, in comparison to the OpenAI client, requires the
187
188
    # "parameters" dict and the "description" string to be present
    # even if they are empty.
189
190
191
192
193
    if tools:
        for function in [
                tool["function"] for tool in tools
                if tool["type"] == "function"
        ]:
194
195
            if function.get("parameters") is None:
                function["parameters"] = {}
196
197
            if function.get("description") is None:
                function["description"] = ""
198
199
200
201
202
203

    from mistral_common.protocol.instruct.request import ChatCompletionRequest
    return ChatCompletionRequest(messages=messages,
                                 tools=tools)  # type: ignore[type-var]


204
class MistralTokenizer(TokenizerBase):
205

206
    def __init__(self, tokenizer: "PublicMistralTokenizer") -> None:
207
208
        self.mistral = tokenizer
        self.instruct = tokenizer.instruct_tokenizer
209
210
        _mistral_version_str = self.instruct.tokenizer.version.value
        self.version: int = int(_mistral_version_str.split("v")[-1])
211

212
        tokenizer_ = tokenizer.instruct_tokenizer.tokenizer
213
214
215
        from mistral_common.tokens.tokenizers.base import SpecialTokenPolicy
        from mistral_common.tokens.tokenizers.tekken import Tekkenizer

216
        self.is_tekken = isinstance(tokenizer_, Tekkenizer)
217
218
        from mistral_common.tokens.tokenizers.sentencepiece import (
            SentencePieceTokenizer)
219
        self.is_spm = isinstance(tokenizer_, SentencePieceTokenizer)
220
221
222
        self._special_token_policy = (SpecialTokenPolicy.IGNORE
                                      if self.is_tekken else None)
        if not (self.is_tekken or self.is_spm):
223
            raise TypeError(f"Unsupported tokenizer: {type(tokenizer_)}")
224

225
        self._vocab = tokenizer_.vocab()
226
        # Convert to a dict[str, int] to match protocol, but this is a lossy
227
228
229
230
231
232
        # conversion. There may be multiple token ids that decode to the same
        # string due to partial UTF-8 byte sequences being converted to �
        self._vocab_dict = {
            token: idx
            for idx, token in enumerate(self._vocab)
        }
233
        self.tokenizer = tokenizer_
234
        self._max_token_id = self.vocab_size - 1
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253

    @classmethod
    def from_pretrained(cls,
                        path_or_repo_id: str,
                        *,
                        revision: Optional[str] = None) -> "MistralTokenizer":
        if not Path(path_or_repo_id).exists():
            assert len(path_or_repo_id.split("/")) == 2, (
                "You have either provided a non-existent path: "
                "{path_or_repo_id} or an invalid HF Hub repo id.")
            tokenizer_file = cls._download_mistral_tokenizer_from_hf(
                path_or_repo_id, revision)
        elif Path(path_or_repo_id).is_dir():
            tokenizer_file_name = find_tokenizer_file(
                os.listdir(path_or_repo_id))
            tokenizer_file = str(Path(path_or_repo_id) / tokenizer_file_name)
        else:
            assert Path(
                path_or_repo_id).is_file(), f"Invalid path: {path_or_repo_id}"
254
            tokenizer_file = str(Path(path_or_repo_id))
255

256
257
        from mistral_common.tokens.tokenizers.mistral import (
            MistralTokenizer as PublicMistralTokenizer)
258
259
260
261
262
263
        mistral_tokenizer = PublicMistralTokenizer.from_file(tokenizer_file)
        return cls(mistral_tokenizer)

    @staticmethod
    def _download_mistral_tokenizer_from_hf(tokenizer_name: str,
                                            revision: Optional[str]) -> str:
264
265
266
267
268
269
270
271
272
273
        try:
            hf_api = HfApi()
            files = hf_api.list_repo_files(repo_id=tokenizer_name,
                                           revision=revision)
        except ConnectionError as exc:
            files = list_local_repo_files(repo_id=tokenizer_name,
                                          revision=revision)

            if len(files) == 0:
                raise exc
274
275
276
277
278
279
280
281

        filename = find_tokenizer_file(files)

        tokenizer_file = hf_hub_download(tokenizer_name,
                                         filename=filename,
                                         revision=revision)
        return tokenizer_file

282
    # the following attributes are set to fit vLLM's design and are used
283
    # by the structured output backends.
284
    @property
285
    def all_special_tokens_extended(self) -> list[str]:
286
287
        from mistral_common.tokens.tokenizers.base import SpecialTokens

288
289
290
291
292
293
294
295
296
        # tekken defines its own extended special tokens list
        if hasattr(self.tokenizer, "SPECIAL_TOKENS"):
            special_tokens = self.tokenizer.SPECIAL_TOKENS
        else:
            special_tokens = list(SpecialTokens)
        return [
            s.value if isinstance(s, SpecialTokens) else s
            for s in special_tokens
        ]
297
298

    @property
299
    def all_special_tokens(self) -> list[str]:
300
        return self.all_special_tokens_extended
301
302

    @property
303
    def all_special_ids(self) -> list[int]:
304
305
306
        return [
            self.all_special_tokens.index(t) for t in self.all_special_tokens
        ]
307
308
309
310
311
312
313
314
315

    @property
    def bos_token_id(self) -> int:
        return self.tokenizer.bos_id

    @property
    def eos_token_id(self) -> int:
        return self.tokenizer.eos_id

316
317
318
319
320
321
322
323
    @property
    def sep_token(self) -> str:
        raise NotImplementedError()

    @property
    def pad_token(self) -> str:
        raise NotImplementedError()

324
325
326
327
328
329
330
331
    @property
    def is_fast(self) -> bool:
        return True

    @property
    def vocab_size(self) -> int:
        return len(self._vocab)

332
333
334
335
    @property
    def max_token_id(self) -> int:
        return self._max_token_id

336
337
338
339
    @property
    def truncation_side(self) -> str:
        raise NotImplementedError()

340
341
342
    def __len__(self) -> int:
        return self.vocab_size

343
344
    def __call__(
        self,
345
        text: Union[str, list[str], list[int]],
346
        text_pair: Optional[str] = None,
347
348
349
350
        add_special_tokens: bool = False,
        truncation: bool = False,
        max_length: Optional[int] = None,
    ):
351
352
        input_ids: Union[list[int], list[list[int]]]
        # For list[str], original prompt text
353
        if is_list_of(text, str):
354
            input_ids_: list[list[int]] = []
355
            for p in text:
356
357
358
                each_input_ids = self.encode_one(p, truncation, max_length)
                input_ids_.append(each_input_ids)
            input_ids = input_ids_
359
        # For list[int], apply chat template output, already tokens.
360
361
        elif is_list_of(text, int):
            input_ids = text
362
363
        # For str, single prompt text
        else:
364
            input_ids = self.encode_one(text, truncation, max_length)
365
        return BatchEncoding({"input_ids": input_ids})
366

367
    def get_vocab(self) -> dict[str, int]:
368
369
370
        # NB: the dictionary form of the vocabulary collapses token ids that map
        # to the same string but have different bytes
        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
377
    def encode_one(
        self,
378
        text: str,
379
380
        truncation: bool = False,
        max_length: Optional[int] = None,
381
    ) -> list[int]:
382
        # Mistral Tokenizers should not add special tokens
383
        input_ids = self.encode(text)
384
385
386
387
388

        if truncation:
            input_ids = input_ids[:max_length]
        return input_ids

389
390
    def encode(self,
               text: str,
391
392
               truncation: Optional[bool] = None,
               max_length: Optional[int] = None,
393
               add_special_tokens: Optional[bool] = None) -> list[int]:
394
        # `encode` should only be used for prompt completion
395
396
        # it should never be used for chat_completion.
        # For chat completion use `apply_chat_template`
397
398
399
400
401
402
        if add_special_tokens is not None:
            return self.tokenizer.encode(text,
                                         bos=add_special_tokens,
                                         eos=add_special_tokens)
        else:
            return self.tokenizer.encode(text, bos=True, eos=False)
403
404

    def apply_chat_template(self,
405
406
407
                            messages: list["ChatCompletionMessageParam"],
                            tools: Optional[list[dict[str, Any]]] = None,
                            **kwargs) -> list[int]:
408

409
        request = make_mistral_chat_completion_request(messages, tools)
410
411
412
413
414
        encoded = self.mistral.encode_chat_completion(request)

        # encode-decode to get clean prompt
        return encoded.tokens

415
    def convert_tokens_to_string(self, tokens: list[str]) -> str:
416
        from mistral_common.tokens.tokenizers.base import SpecialTokens
417
        if self.is_tekken:
418
419
            tokens = [
                t for t in tokens
420
421
                if (t is SpecialTokens.tool_calls
                    or t not in self.tokenizer._all_special_tokens)
422
423
424
425
426
            ]

            if any(isinstance(t, bytes) for t in tokens):
                # we need to encode and decode all tokens again
                shift = self.tokenizer.num_special_tokens
427
428
429
430
431
432
433
434
435
436
437
438
439
440

                def _token_to_id(t: str):
                    t_bytes = t.encode("utf-8") \
                        if not isinstance(t, bytes) else t
                    try:
                        return shift + \
                            self.tokenizer._tekken_token2id_nospecial[t_bytes]
                    except KeyError:
                        logger.warning(
                            "Failed to convert token %s to id,"
                            " replacing with <unk>", t_bytes)
                        return self.tokenizer.unk_id

                ids = [_token_to_id(t) for t in tokens]
441
442
                decoded = self.tokenizer.decode(ids,
                                                self._special_token_policy)
443
444
            else:
                decoded = "".join(tokens)
445
        else:
446
447
448
            # make sure certain special tokens like Tool calls are
            # not decoded
            special_tokens = {SpecialTokens.tool_calls}
449
            regular_tokens: list[str] = []
450
451
452
453
454
455
            decoded_list = []

            for token in tokens:
                if token in special_tokens:
                    if regular_tokens:
                        decoded_list.append(
456
457
                            self.tokenizer.decode(regular_tokens,
                                                  self._special_token_policy))
458
459
460
461
462
463
464
                        regular_tokens = []
                    decoded_list.append(token)
                else:
                    regular_tokens.append(token)

            if regular_tokens:
                decoded_list.append(
465
466
                    self.tokenizer.decode(regular_tokens,
                                          self._special_token_policy))
467
468

            decoded = ''.join(decoded_list)
469
470

        return decoded
471

472
    def decode(self,
473
               ids: Union[list[int], int],
474
475
476
               skip_special_tokens: bool = True) -> str:
        assert (
            skip_special_tokens
477
        ), "skip_special_tokens=False is not supported for Mistral tokenizers."
478

479
480
        if isinstance(ids, int):
            ids = [ids]
481
        return self.tokenizer.decode(ids, self._special_token_policy)
482
483

    def convert_ids_to_tokens(
484
        self,
485
        ids: list[int],
486
        skip_special_tokens: bool = True,
487
    ) -> list[str]:
488
        from mistral_common.tokens.tokenizers.base import SpecialTokens
Julien Denize's avatar
Julien Denize committed
489
490
        from mistral_common.tokens.tokenizers.instruct import (
            InstructTokenizerV13)
491

492
493
494
        # TODO(Patrick) - potentially allow special tokens to not be skipped
        assert (
            skip_special_tokens
495
        ), "skip_special_tokens=False is not supported for Mistral tokenizers."
496

497
        assert self.is_tekken or self.is_spm, type(self.tokenizer)
498

499
        if self.is_tekken:
Julien Denize's avatar
Julien Denize committed
500
501
            # skip special tokens except tool call and think tokens
            non_skip_special_tokens = {
502
                self.tokenizer.get_control_token(SpecialTokens.tool_calls)
Julien Denize's avatar
Julien Denize committed
503
504
505
506
507
508
509
510
511
            }
            if isinstance(self.instruct, InstructTokenizerV13):
                if self.instruct.BEGIN_THINK:
                    non_skip_special_tokens.add(self.instruct.BEGIN_THINK)
                if self.instruct.END_THINK:
                    non_skip_special_tokens.add(self.instruct.END_THINK)
            ids = [
                i for i in ids if i > self.tokenizer.num_special_tokens
                or i in non_skip_special_tokens
512
            ]
513

514
        tokens = [self.tokenizer.id_to_piece(id) for id in ids]
515

516
        if any("�" in t for t in tokens) and self.is_tekken:
517
518
            # if a decoded token contains the replacement character, then the
            # token has an incomplete UTF-8 character so we must use bytes
519
            # See: https://github.com/vllm-project/vllm/pull/8640
520
            #      https://github.com/vllm-project/vllm/pull/9625
521
            # if underlying tokenizeir is sentencepiece, we just add "�"
522
523
524
525
            tokens = [
                self.tokenizer.id_to_byte_piece(id, self._special_token_policy)
                for id in ids
            ]
526

527
        return tokens