mistral.py 16.2 KB
Newer Older
1
2
# SPDX-License-Identifier: Apache-2.0

3
4
5
6
import os
import re
from dataclasses import dataclass
from pathlib import Path
7
from typing import TYPE_CHECKING, Any, Dict, List, Optional, Union, cast
8

9
import huggingface_hub
10
11
from huggingface_hub import HfApi, hf_hub_download

12
from vllm.logger import init_logger
13
from vllm.utils import is_list_of
14

15
if TYPE_CHECKING:
16
17
18
19
20
21
22
    # 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)

23
    from vllm.entrypoints.chat_utils import ChatCompletionMessageParam
24

25
26
logger = init_logger(__name__)

27
28
29

@dataclass
class Encoding:
30
    input_ids: Union[List[int], List[List[int]]]
31
32


33
def maybe_serialize_tool_calls(request: "ChatCompletionRequest"):
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
67
68
69
    # 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


70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
def list_local_repo_files(repo_id: str, revision: Optional[str]) -> List[str]:
    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 []


90
def find_tokenizer_file(files: List[str]):
91
92
    file_pattern = re.compile(
        r"^tokenizer\.model\.v.*$|^tekken\.json$|^tokenizer\.mm\.model\.v.*$")
93
94
95
96

    matched_files = [file for file in files if file_pattern.match(file)]
    if len(matched_files) > 1:
        raise OSError(f"Found {len(matched_files)} files matching the "
97
98
                      f"pattern: {file_pattern}. Make sure only one Mistral "
                      f"tokenizer is present in {files}.")
99
100
    elif len(matched_files) == 0:
        raise OSError(f"Found {len(matched_files)} files matching the "
101
102
                      f"pattern: {file_pattern}. Make sure that a Mistral "
                      f"tokenizer is present in {files}.")
103
104
105
106

    return matched_files[0]


107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
def make_mistral_chat_completion_request(
        messages: List["ChatCompletionMessageParam"],
        tools: Optional[List[Dict[str,
                                  Any]]] = None) -> "ChatCompletionRequest":
    last_message = cast(Dict[str, Any], messages[-1])
    if last_message["role"] == "assistant":
        last_message["prefix"] = True

        last_message = cast(Dict[str, Any], messages[-1])
        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:
        if message.get("role") == "assistant":
            content = message.get("content")
            if isinstance(content, list):
                content = "\n".join(chunk.get("text") for chunk in content)
                message["content"] = content

    # The Mistral client, in comparison to the OpenAI client, requires the
    # "parameters" dict to be present, even if it's empty.
    if tools:
        for function in [
                tool["function"] for tool in tools
                if tool["type"] == "function"
        ]:
            function.setdefault("parameters", {})

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


143
144
class MistralTokenizer:

145
    def __init__(self, tokenizer: "PublicMistralTokenizer") -> None:
146
147
148
        self.mistral = tokenizer
        self.instruct = tokenizer.instruct_tokenizer

149
        tokenizer_ = tokenizer.instruct_tokenizer.tokenizer
150
151
        from mistral_common.tokens.tokenizers.tekken import (
            SpecialTokenPolicy, Tekkenizer)
152
        self.is_tekken = isinstance(tokenizer_, Tekkenizer)
153
154
        from mistral_common.tokens.tokenizers.sentencepiece import (
            SentencePieceTokenizer)
155
156
        self.is_spm = isinstance(tokenizer_, SentencePieceTokenizer)
        if self.is_tekken:
157
            # Make sure special tokens will not raise
158
            tokenizer_.special_token_policy = SpecialTokenPolicy.IGNORE
159
        elif self.is_spm:
160
            pass
161
162
        else:
            raise TypeError(f"Unsupported tokenizer: {type(tokenizer_)}")
163

164
165
166
167
168
169
170
171
        self._vocab = tokenizer_.vocab()
        # Convert to a Dict[str, int] to match protocol, but this is a lossy
        # 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)
        }
172
        self.tokenizer = tokenizer_
173
        self._max_token_id = self.vocab_size - 1
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193

    @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}"

194
195
        from mistral_common.tokens.tokenizers.mistral import (
            MistralTokenizer as PublicMistralTokenizer)
196
197
198
199
200
201
        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:
202
203
204
205
206
207
208
209
210
211
        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
212
213
214
215
216
217
218
219

        filename = find_tokenizer_file(files)

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

220
221
    # the following attributes are set to fit VLLM's design and are used
    # by the guided structured output backends.
222
223
    @property
    def all_special_tokens_extended(self) -> List[str]:
224
225
        from mistral_common.tokens.tokenizers.base import SpecialTokens

226
227
228
229
230
231
232
233
234
        # 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
        ]
235
236
237

    @property
    def all_special_tokens(self) -> List[str]:
238
        return self.all_special_tokens_extended
239
240
241

    @property
    def all_special_ids(self) -> List[int]:
242
243
244
        return [
            self.all_special_tokens.index(t) for t in self.all_special_tokens
        ]
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261

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

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

    @property
    def is_fast(self) -> bool:
        return True

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

262
263
264
265
    @property
    def max_token_id(self) -> int:
        return self._max_token_id

266
267
268
    def __len__(self) -> int:
        return self.vocab_size

269
270
    def __call__(
        self,
271
        prompt: Union[str, List[str], List[int]],
272
273
274
275
        add_special_tokens: bool = False,
        truncation: bool = False,
        max_length: Optional[int] = None,
    ):
276
277
278
279
280
281
282
283
284
285
286
287
288
289
        input_ids: Union[List[int], List[List[int]]]
        # For List[str], original prompt text
        if is_list_of(prompt, str):
            input_ids_: List[List[int]] = []
            for p in prompt:
                each_input_ids = self.encode_one(p, truncation, max_length)
                input_ids_.append(each_input_ids)
            input_ids = input_ids_
        # For List[int], apply chat template output, already tokens.
        elif is_list_of(prompt, int):
            input_ids = prompt
        # For str, single prompt text
        else:
            input_ids = self.encode_one(prompt, truncation, max_length)
290
291
        return Encoding(input_ids=input_ids)

292
    def get_vocab(self) -> Dict[str, int]:
293
294
295
        # NB: the dictionary form of the vocabulary collapses token ids that map
        # to the same string but have different bytes
        return self._vocab_dict
296
297

    def get_added_vocab(self) -> Dict[str, int]:
298
        # Mistral tokenizers have no added vocabulary
299
        return {}
300

301
302
303
304
305
306
307
308
309
310
311
312
313
    def encode_one(
        self,
        prompt: str,
        truncation: bool = False,
        max_length: Optional[int] = None,
    ) -> List[int]:
        # Mistral Tokenizers should not add special tokens
        input_ids = self.encode(prompt)

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

314
    def encode(self, prompt: str) -> List[int]:
315
        # `encode` should only be used for prompt completion
316
317
318
319
320
        # it should never be used for chat_completion.
        # For chat completion use `apply_chat_template`
        return self.tokenizer.encode(prompt, bos=True, eos=False)

    def apply_chat_template(self,
321
                            messages: List["ChatCompletionMessageParam"],
322
                            tools: Optional[List[Dict[str, Any]]] = None,
323
324
                            **kwargs) -> List[int]:

325
        request = make_mistral_chat_completion_request(messages, tools)
326
327
328
329
330
331
        encoded = self.mistral.encode_chat_completion(request)

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

    def convert_tokens_to_string(self, tokens: List[str]) -> str:
332
        from mistral_common.tokens.tokenizers.base import SpecialTokens
333
        if self.is_tekken:
334
335
            tokens = [
                t for t in tokens
336
337
                if (t is SpecialTokens.tool_calls
                    or t not in self.tokenizer._all_special_tokens)
338
339
340
341
342
            ]

            if any(isinstance(t, bytes) for t in tokens):
                # we need to encode and decode all tokens again
                shift = self.tokenizer.num_special_tokens
343
344
345
346
347
348
349
350
351
352
353
354
355
356

                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]
357
358
359
                decoded = self.tokenizer.decode(ids)
            else:
                decoded = "".join(tokens)
360
        else:
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
            # make sure certain special tokens like Tool calls are
            # not decoded
            special_tokens = {SpecialTokens.tool_calls}
            regular_tokens: List[str] = []
            decoded_list = []

            for token in tokens:
                if token in special_tokens:
                    if regular_tokens:
                        decoded_list.append(
                            self.tokenizer.decode(regular_tokens))
                        regular_tokens = []
                    decoded_list.append(token)
                else:
                    regular_tokens.append(token)

            if regular_tokens:
                decoded_list.append(
379
                    self.tokenizer.decode(regular_tokens))  # type: ignore
380
381

            decoded = ''.join(decoded_list)
382
383

        return decoded
384

385
386
387
    # WARN: Outlines logits processors can overwrite this method.
    # See: guided_decoding/outlines_logits_processors.py::_adapt_tokenizer
    # for more.
388
389
390
391
392
    def decode(self,
               ids: Union[List[int], int],
               skip_special_tokens: bool = True) -> str:
        assert (
            skip_special_tokens
393
        ), "skip_special_tokens=False is not supported for Mistral tokenizers."
394

395
396
397
398
399
        if isinstance(ids, int):
            ids = [ids]
        return self.tokenizer.decode(ids)

    def convert_ids_to_tokens(
400
401
402
403
        self,
        ids: List[int],
        skip_special_tokens: bool = True,
    ) -> List[str]:
404
405
        from mistral_common.tokens.tokenizers.base import SpecialTokens

406
407
408
        # TODO(Patrick) - potentially allow special tokens to not be skipped
        assert (
            skip_special_tokens
409
        ), "skip_special_tokens=False is not supported for Mistral tokenizers."
410

411
        assert self.is_tekken or self.is_spm, type(self.tokenizer)
412

413
        if self.is_tekken:
414
415
416
417
418
            # skip special tokens except tool call
            ids = [
                i for i in ids if i > self.tokenizer.num_special_tokens or i ==
                self.tokenizer.get_control_token(SpecialTokens.tool_calls)
            ]
419

420
        tokens = [self.tokenizer.id_to_piece(id) for id in ids]
421

422
        if any("�" in t for t in tokens) and self.is_tekken:
423
424
            # if a decoded token contains the replacement character, then the
            # token has an incomplete UTF-8 character so we must use bytes
425
            # See: https://github.com/vllm-project/vllm/pull/8640
426
            #      https://github.com/vllm-project/vllm/pull/9625
427
            # if underlying tokenizeir is sentencepiece, we just add "�"
428
429
            tokens = [self.tokenizer.id_to_byte_piece(id) for id in ids]

430
        return tokens