chat_utils.py 47.5 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
import asyncio
5
import json
6
from abc import ABC, abstractmethod
7
from collections import Counter, defaultdict, deque
8
from collections.abc import Awaitable, Iterable
9
from functools import cached_property, lru_cache, partial
10
from pathlib import Path
11
12
from typing import (Any, Callable, Generic, Literal, Optional, TypeVar, Union,
                    cast)
13

14
15
import jinja2.nodes
import transformers.utils.chat_template_utils as hf_chat_utils
16
17
# yapf conflicts with isort for this block
# yapf: disable
18
from openai.types.chat import (ChatCompletionAssistantMessageParam,
19
20
                               ChatCompletionContentPartImageParam,
                               ChatCompletionContentPartInputAudioParam)
21
22
from openai.types.chat import (
    ChatCompletionContentPartParam as OpenAIChatCompletionContentPartParam)
23
24
from openai.types.chat import (ChatCompletionContentPartRefusalParam,
                               ChatCompletionContentPartTextParam)
25
26
from openai.types.chat import (
    ChatCompletionMessageParam as OpenAIChatCompletionMessageParam)
27
28
from openai.types.chat import (ChatCompletionMessageToolCallParam,
                               ChatCompletionToolMessageParam)
29
30
from openai.types.chat.chat_completion_content_part_input_audio_param import (
    InputAudio)
31
from openai.types.responses import ResponseInputImageParam
32
33
from PIL import Image
from pydantic import BaseModel, ConfigDict, TypeAdapter
34
# yapf: enable
35
36
from transformers import (PreTrainedTokenizer, PreTrainedTokenizerFast,
                          ProcessorMixin)
37
# pydantic needs the TypedDict from typing_extensions
38
from typing_extensions import Required, TypeAlias, TypedDict
39

40
from vllm.config import ModelConfig
41
from vllm.logger import init_logger
42
from vllm.model_executor.models import SupportsMultiModal
43
from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalDataDict
44
from vllm.multimodal.utils import MediaConnector
45
46
47
48
# yapf: disable
from vllm.transformers_utils.chat_templates import (
    get_chat_template_fallback_path)
# yapf: enable
49
from vllm.transformers_utils.processor import cached_get_processor
50
from vllm.transformers_utils.tokenizer import AnyTokenizer, MistralTokenizer
51
from vllm.utils import deprecate_kwargs, random_uuid
52
53
54

logger = init_logger(__name__)

55
56
57
58
59
60
MODALITY_PLACEHOLDERS_MAP = {
    "image": "<##IMAGE##>",
    "audio": "<##AUDIO##>",
    "video": "<##VIDEO##>",
}

61

62
63
64
65
66
67
68
69
70
71
72
73
74
75
class AudioURL(TypedDict, total=False):
    url: Required[str]
    """
    Either a URL of the audio or a data URL with base64 encoded audio data.
    """


class ChatCompletionContentPartAudioParam(TypedDict, total=False):
    audio_url: Required[AudioURL]

    type: Required[Literal["audio_url"]]
    """The type of the content part."""


76
77
78
79
80
81
82
83
84
85
86
class ChatCompletionContentPartImageEmbedsParam(TypedDict, total=False):
    image_embeds: Required[Union[str, dict[str, str]]]
    """
    The image embeddings. It can be either:
    - A single base64 string.
    - A dictionary where each value is a base64 string.
    """
    type: Required[Literal["image_embeds"]]
    """The type of the content part."""


87
88
89
90
91
92
93
94
95
96
97
98
99
100
class VideoURL(TypedDict, total=False):
    url: Required[str]
    """
    Either a URL of the video or a data URL with base64 encoded video data.
    """


class ChatCompletionContentPartVideoParam(TypedDict, total=False):
    video_url: Required[VideoURL]

    type: Required[Literal["video_url"]]
    """The type of the content part."""


101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
class PILImage(BaseModel):
    """
    A PIL.Image.Image object.
    """
    image_pil: Image.Image
    model_config = ConfigDict(arbitrary_types_allowed=True)


class CustomChatCompletionContentPILImageParam(TypedDict, total=False):
    """A simpler version of the param that only accepts a PIL image.

    Example:
    {
        "image_pil": ImageAsset('cherry_blossom').pil_image
    }
    """
    image_pil: Required[PILImage]


120
121
122
class CustomChatCompletionContentSimpleImageParam(TypedDict, total=False):
    """A simpler version of the param that only accepts a plain image_url.
    This is supported by OpenAI API, although it is not documented.
123

124
125
126
127
128
129
130
131
132
133
    Example:
    {
        "image_url": "https://example.com/image.jpg"
    }
    """
    image_url: Required[str]


class CustomChatCompletionContentSimpleAudioParam(TypedDict, total=False):
    """A simpler version of the param that only accepts a plain audio_url.
134

135
136
137
138
139
140
141
142
    Example:
    {
        "audio_url": "https://example.com/audio.mp3"
    }
    """
    audio_url: Required[str]


143
144
145
146
147
148
149
150
151
152
153
class CustomChatCompletionContentSimpleVideoParam(TypedDict, total=False):
    """A simpler version of the param that only accepts a plain audio_url.

    Example:
    {
        "video_url": "https://example.com/video.mp4"
    }
    """
    video_url: Required[str]


Julien Denize's avatar
Julien Denize committed
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
class CustomThinkCompletionContentParam(TypedDict, total=False):
    """A Think Completion Content Param that accepts a plain text and a boolean.

    Example:
    {
        "thinking": "I am thinking about the answer",
        "closed": True,
        "type": "thinking"
    }
    """

    thinking: Required[str]
    """The thinking content."""

    closed: bool
    """Whether the thinking is closed."""

    type: Required[Literal["thinking"]]
    """The thinking type."""


175
176
ChatCompletionContentPartParam: TypeAlias = Union[
    OpenAIChatCompletionContentPartParam, ChatCompletionContentPartAudioParam,
177
    ChatCompletionContentPartInputAudioParam,
178
    ChatCompletionContentPartVideoParam, ChatCompletionContentPartRefusalParam,
179
    CustomChatCompletionContentPILImageParam,
180
    CustomChatCompletionContentSimpleImageParam,
181
    ChatCompletionContentPartImageEmbedsParam,
182
    CustomChatCompletionContentSimpleAudioParam,
Julien Denize's avatar
Julien Denize committed
183
184
    CustomChatCompletionContentSimpleVideoParam, str,
    CustomThinkCompletionContentParam]
185
186
187
188
189
190
191


class CustomChatCompletionMessageParam(TypedDict, total=False):
    """Enables custom roles in the Chat Completion API."""
    role: Required[str]
    """The role of the message's author."""

192
    content: Union[str, list[ChatCompletionContentPartParam]]
193
194
195
196
197
198
199
200
201
    """The contents of the message."""

    name: str
    """An optional name for the participant.

    Provides the model information to differentiate between participants of the
    same role.
    """

202
203
204
205
206
207
    tool_call_id: Optional[str]
    """Tool call that this message is responding to."""

    tool_calls: Optional[Iterable[ChatCompletionMessageToolCallParam]]
    """The tool calls generated by the model, such as function calls."""

208
209
210
211
212

ChatCompletionMessageParam = Union[OpenAIChatCompletionMessageParam,
                                   CustomChatCompletionMessageParam]


213
# TODO: Make fields ReadOnly once mypy supports it
214
215
216
217
class ConversationMessage(TypedDict, total=False):
    role: Required[str]
    """The role of the message's author."""

218
    content: Union[Optional[str], list[dict[str, str]]]
219
220
221
222
223
224
225
226
227
228
    """The contents of the message"""

    tool_call_id: Optional[str]
    """Tool call that this message is responding to."""

    name: Optional[str]
    """The name of the function to call"""

    tool_calls: Optional[Iterable[ChatCompletionMessageToolCallParam]]
    """The tool calls generated by the model, such as function calls."""
229
230


231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
# Passed in by user
ChatTemplateContentFormatOption = Literal["auto", "string", "openai"]

# Used internally
_ChatTemplateContentFormat = Literal["string", "openai"]


def _is_var_access(node: jinja2.nodes.Node, varname: str) -> bool:
    if isinstance(node, jinja2.nodes.Name):
        return node.ctx == "load" and node.name == varname

    return False


def _is_attr_access(node: jinja2.nodes.Node, varname: str, key: str) -> bool:
    if isinstance(node, jinja2.nodes.Getitem):
        return (_is_var_access(node.node, varname)
                and isinstance(node.arg, jinja2.nodes.Const)
                and node.arg.value == key)

    if isinstance(node, jinja2.nodes.Getattr):
        return _is_var_access(node.node, varname) and node.attr == key

    return False


def _is_var_or_elems_access(
    node: jinja2.nodes.Node,
    varname: str,
    key: Optional[str] = None,
) -> bool:
    if isinstance(node, jinja2.nodes.Filter):
        return (node.node is not None
                and _is_var_or_elems_access(node.node, varname, key))
    if isinstance(node, jinja2.nodes.Test):
        return _is_var_or_elems_access(node.node, varname, key)

    if (isinstance(node, jinja2.nodes.Getitem)
            and isinstance(node.arg, jinja2.nodes.Slice)):
        return _is_var_or_elems_access(node.node, varname, key)

    # yapf: disable
    return (
        _is_attr_access(node, varname, key) if key
        else _is_var_access(node, varname)
    ) # yapf: enable


def _iter_nodes_assign_var_or_elems(root: jinja2.nodes.Node, varname: str):
    # Global variable that is implicitly defined at the root
    yield root, varname

    # Iterative BFS
    related_varnames = deque([varname])
    while related_varnames:
        related_varname = related_varnames.popleft()

        for assign_ast in root.find_all(jinja2.nodes.Assign):
            lhs = assign_ast.target
            rhs = assign_ast.node

            if _is_var_or_elems_access(rhs, related_varname):
                assert isinstance(lhs, jinja2.nodes.Name)
                yield assign_ast, lhs.name

                # Avoid infinite looping for self-assignment
                if lhs.name != related_varname:
                    related_varnames.append(lhs.name)


# NOTE: The proper way to handle this is to build a CFG so that we can handle
# the scope in which each variable is defined, but that is too complicated
def _iter_nodes_assign_messages_item(root: jinja2.nodes.Node):
    messages_varnames = [
        varname
        for _, varname in _iter_nodes_assign_var_or_elems(root, "messages")
    ]

    # Search for {%- for message in messages -%} loops
    for loop_ast in root.find_all(jinja2.nodes.For):
        loop_iter = loop_ast.iter
        loop_target = loop_ast.target

        for varname in messages_varnames:
            if _is_var_or_elems_access(loop_iter, varname):
                assert isinstance(loop_target, jinja2.nodes.Name)
                yield loop_ast, loop_target.name
                break


def _iter_nodes_assign_content_item(root: jinja2.nodes.Node):
    message_varnames = [
        varname for _, varname in _iter_nodes_assign_messages_item(root)
    ]

    # Search for {%- for content in message['content'] -%} loops
    for loop_ast in root.find_all(jinja2.nodes.For):
        loop_iter = loop_ast.iter
        loop_target = loop_ast.target

        for varname in message_varnames:
            if _is_var_or_elems_access(loop_iter, varname, "content"):
                assert isinstance(loop_target, jinja2.nodes.Name)
                yield loop_ast, loop_target.name
                break


def _try_extract_ast(chat_template: str) -> Optional[jinja2.nodes.Template]:
    try:
        jinja_compiled = hf_chat_utils._compile_jinja_template(chat_template)
        return jinja_compiled.environment.parse(chat_template)
    except Exception:
        logger.exception("Error when compiling Jinja template")
        return None


347
@lru_cache(maxsize=32)
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
def _detect_content_format(
    chat_template: str,
    *,
    default: _ChatTemplateContentFormat,
) -> _ChatTemplateContentFormat:
    jinja_ast = _try_extract_ast(chat_template)
    if jinja_ast is None:
        return default

    try:
        next(_iter_nodes_assign_content_item(jinja_ast))
    except StopIteration:
        return "string"
    except Exception:
        logger.exception("Error when parsing AST of Jinja template")
        return default
    else:
        return "openai"


368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
def resolve_mistral_chat_template(
    chat_template: Optional[str],
    **kwargs: Any,
) -> Optional[str]:
    if chat_template is not None:
        logger.warning_once(
            "'chat_template' cannot be overridden for mistral tokenizer.")
    if "add_generation_prompt" in kwargs:
        logger.warning_once(
            "'add_generation_prompt' is not supported for mistral tokenizer, "
            "so it will be ignored.")
    if "continue_final_message" in kwargs:
        logger.warning_once(
            "'continue_final_message' is not supported for mistral tokenizer, "
            "so it will be ignored.")
    return None

385

386
387
388
389
@deprecate_kwargs(
    "trust_remote_code",
    additional_message="Please use `model_config.trust_remote_code` instead.",
)
390
def resolve_hf_chat_template(
391
392
393
    tokenizer: Union[PreTrainedTokenizer, PreTrainedTokenizerFast],
    chat_template: Optional[str],
    tools: Optional[list[dict[str, Any]]],
394
395
    *,
    model_config: ModelConfig,
396
    trust_remote_code: Optional[bool] = None,
397
398
399
400
401
402
403
404
405
406
407
408
) -> Optional[str]:
    # 1st priority: The given chat template
    if chat_template is not None:
        return chat_template

    # 2nd priority: AutoProcessor chat template, unless tool calling is enabled
    if tools is None:
        try:
            processor = cached_get_processor(
                tokenizer.name_or_path,
                processor_cls=(PreTrainedTokenizer, PreTrainedTokenizerFast,
                               ProcessorMixin),
409
                trust_remote_code=model_config.trust_remote_code,
410
411
            )
            if isinstance(processor, ProcessorMixin) and \
412
                hasattr(processor, 'chat_template') and \
413
414
415
                processor.chat_template is not None:
                return processor.chat_template
        except Exception:
416
            logger.debug("Failed to load AutoProcessor chat template for %s", tokenizer.name_or_path, exc_info=True)  # noqa: E501
417
418
419
420
421
422
423
424

    # 3rd priority: AutoTokenizer chat template
    try:
        return tokenizer.get_chat_template(chat_template, tools=tools)
    except Exception:
        logger.debug("Failed to load AutoTokenizer chat template for %s",
                     tokenizer.name_or_path, exc_info=True)

425
426
427
428
429
430
431
432
433
434
435
436
437
438
    # 4th priority: Predefined fallbacks
    path = get_chat_template_fallback_path(
        model_type=model_config.hf_config.model_type,
        tokenizer_name_or_path=model_config.tokenizer,
    )
    if path is not None:
        logger.info("Loading chat template fallback for %s as there isn't one "
                    "defined on HF Hub.", tokenizer.name_or_path)
        chat_template = load_chat_template(path)
    else:
        logger.debug("There is no chat template fallback for %s",
                     tokenizer.name_or_path)

    return chat_template
439
440


441
442
def _resolve_chat_template_content_format(
    chat_template: Optional[str],
443
    tools: Optional[list[dict[str, Any]]],
444
    tokenizer: AnyTokenizer,
445
446
    *,
    model_config: ModelConfig,
447
448
) -> _ChatTemplateContentFormat:
    if isinstance(tokenizer, (PreTrainedTokenizer, PreTrainedTokenizerFast)):
449
        hf_chat_template = resolve_hf_chat_template(
450
451
452
            tokenizer,
            chat_template=chat_template,
            tools=tools,
453
            model_config=model_config,
454
        )
455
    else:
456
457
458
459
        hf_chat_template = None

    jinja_text = (hf_chat_template if isinstance(hf_chat_template, str)
                  else load_chat_template(chat_template, is_literal=True))
460
461
462
463

    detected_format = ("string" if jinja_text is None else
                       _detect_content_format(jinja_text, default="string"))

464
    return detected_format
465
466
467


@lru_cache
468
def _log_chat_template_content_format(
469
470
    chat_template: Optional[str],
    given_format: ChatTemplateContentFormatOption,
471
472
    detected_format: ChatTemplateContentFormatOption,
):
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
    logger.info(
        "Detected the chat template content format to be '%s'. "
        "You can set `--chat-template-content-format` to override this.",
        detected_format,
    )

    if given_format != "auto" and given_format != detected_format:
        logger.warning(
            "You specified `--chat-template-content-format %s` "
            "which is different from the detected format '%s'. "
            "If our automatic detection is incorrect, please consider "
            "opening a GitHub issue so that we can improve it: "
            "https://github.com/vllm-project/vllm/issues/new/choose",
            given_format,
            detected_format,
        )

490

491
492
493
494
@deprecate_kwargs(
    "trust_remote_code",
    additional_message="Please use `model_config.trust_remote_code` instead.",
)
495
496
497
498
499
def resolve_chat_template_content_format(
    chat_template: Optional[str],
    tools: Optional[list[dict[str, Any]]],
    given_format: ChatTemplateContentFormatOption,
    tokenizer: AnyTokenizer,
500
501
502
    *,
    model_config: ModelConfig,
    trust_remote_code: Optional[bool] = None,
503
) -> _ChatTemplateContentFormat:
504
505
506
    if given_format != "auto":
        return given_format

507
508
509
510
    detected_format = _resolve_chat_template_content_format(
        chat_template,
        tools,
        tokenizer,
511
        model_config=model_config,
512
513
514
515
516
517
518
519
    )

    _log_chat_template_content_format(
        chat_template,
        given_format=given_format,
        detected_format=detected_format,
    )

520
    return detected_format
521

522
523


524
ModalityStr = Literal["image", "audio", "video", "image_embeds"]
525
526
527
528
_T = TypeVar("_T")


class BaseMultiModalItemTracker(ABC, Generic[_T]):
529
530
531
532
533
534
535
    """
    Tracks multi-modal items in a given request and ensures that the number
    of multi-modal items in a given request does not exceed the configured
    maximum per prompt.
    """

    def __init__(self, model_config: ModelConfig, tokenizer: AnyTokenizer):
536
537
        super().__init__()

538
539
        self._model_config = model_config
        self._tokenizer = tokenizer
540

541
        self._items_by_modality = defaultdict[str, list[_T]](list)
542

543
544
545
546
    @property
    def model_config(self) -> ModelConfig:
        return self._model_config

547
548
    @cached_property
    def model_cls(self):
549
        from vllm.model_executor.model_loader import get_model_cls
550
551
        return get_model_cls(self.model_config)

552
553
554
555
    @property
    def allowed_local_media_path(self):
        return self._model_config.allowed_local_media_path

556
557
558
559
    @property
    def mm_registry(self):
        return MULTIMODAL_REGISTRY

560
561
562
563
564
    def add(self, modality: ModalityStr, item: _T) -> Optional[str]:
        """
        Add a multi-modal item to the current prompt and returns the
        placeholder string to use, if any.
        """
565
566
        mm_registry = self.mm_registry
        model_config = self.model_config
567
        model_cls = cast(SupportsMultiModal, self.model_cls)
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582

        input_modality = modality.replace("_embeds", "")

        if mm_registry.has_processor(model_config):
            mm_processor = mm_registry.create_processor(model_config)
            allowed_counts = mm_processor.info.get_allowed_mm_limits()
            allowed_count = allowed_counts.get(input_modality, 0)
        else:
            mm_config = model_config.multimodal_config
            if mm_config is None:
                msg = "This model does not support multi-modal inputs"
                raise ValueError(msg)

            allowed_count = mm_config.get_limit_per_prompt(input_modality)

583
        current_count = len(self._items_by_modality[modality]) + 1
584
585
586
        if current_count > allowed_count:
            raise ValueError(
                f"At most {allowed_count} {modality}(s) may be provided in "
587
588
                "one request. You can set `--limit-mm-per-prompt` to "
                "increase this limit if the model supports it.")
589

590
        self._items_by_modality[modality].append(item)
591

592
        return model_cls.get_placeholder_str(modality, current_count)
593
594
595
596
597
598

    @abstractmethod
    def create_parser(self) -> "BaseMultiModalContentParser":
        raise NotImplementedError


599
class MultiModalItemTracker(BaseMultiModalItemTracker[object]):
600
601

    def all_mm_data(self) -> Optional[MultiModalDataDict]:
602
603
604
605
606
607
608
609
610
611
612
613
614
615
        if not self._items_by_modality:
            return None
        mm_inputs = {}
        items_by_modality = dict(self._items_by_modality)
        if "image" in items_by_modality and "image_embeds" in items_by_modality:
            raise ValueError(\
                "Mixing raw image and embedding inputs is not allowed")

        if "image_embeds" in items_by_modality:
            image_embeds_lst = items_by_modality["image_embeds"]
            if len(image_embeds_lst) > 1:
                raise ValueError(\
                    "Only one message can have {'type': 'image_embeds'}")
            mm_inputs["image"] = image_embeds_lst[0]
616
        if "image" in items_by_modality:
617
            mm_inputs["image"] = items_by_modality["image"] # A list of images
618
        if "audio" in items_by_modality:
619
            mm_inputs["audio"] = items_by_modality["audio"] # A list of audios
620
        if "video" in items_by_modality:
621
622
            mm_inputs["video"] = items_by_modality["video"] # A list of videos
        return mm_inputs
623
624
625
626
627

    def create_parser(self) -> "BaseMultiModalContentParser":
        return MultiModalContentParser(self)


628
class AsyncMultiModalItemTracker(BaseMultiModalItemTracker[Awaitable[object]]):
629
630

    async def all_mm_data(self) -> Optional[MultiModalDataDict]:
631
632
633
634
        if not self._items_by_modality:
            return None
        mm_inputs = {}
        items_by_modality = {
635
636
637
                modality: await asyncio.gather(*items)
                for modality, items in self._items_by_modality.items()
            }
638

639
640
641
642
643
644
645
646
647
648
        if "image" in items_by_modality and "image_embeds" in items_by_modality:
            raise ValueError(
                "Mixing raw image and embedding inputs is not allowed")

        if "image_embeds" in items_by_modality:
            image_embeds_lst = items_by_modality["image_embeds"]
            if len(image_embeds_lst) > 1:
                raise ValueError(
                    "Only one message can have {'type': 'image_embeds'}")
            mm_inputs["image"] = image_embeds_lst[0]
649
        if "image" in items_by_modality:
650
            mm_inputs["image"] = items_by_modality["image"] # A list of images
651
        if "audio" in items_by_modality:
652
            mm_inputs["audio"] = items_by_modality["audio"] # A list of audios
653
        if "video" in items_by_modality:
654
655
            mm_inputs["video"] = items_by_modality["video"] # A list of videos
        return mm_inputs
656
657
658
659
660
661
662
663
664
665

    def create_parser(self) -> "BaseMultiModalContentParser":
        return AsyncMultiModalContentParser(self)


class BaseMultiModalContentParser(ABC):

    def __init__(self) -> None:
        super().__init__()

666
667
668
669
670
671
672
673
674
675
676
        # stores model placehodlers list with corresponding
        # general MM placeholder:
        # {
        #   "<##IMAGE##>": ["<image>", "<image>", "<image>"],
        #   "<##AUDIO##>": ["<audio>", "<audio>"]
        # }
        self._placeholder_storage: dict[str, list] = defaultdict(list)

    def _add_placeholder(self, modality: ModalityStr,
                         placeholder: Optional[str]):
        mod_placeholder = MODALITY_PLACEHOLDERS_MAP[modality]
677
        if placeholder:
678
            self._placeholder_storage[mod_placeholder].append(placeholder)
679

680
681
    def mm_placeholder_storage(self) -> dict[str, list]:
        return dict(self._placeholder_storage)
682
683
684
685
686

    @abstractmethod
    def parse_image(self, image_url: str) -> None:
        raise NotImplementedError

687
688
689
690
691
    @abstractmethod
    def parse_image_embeds(self,
                           image_embeds: Union[str, dict[str, str]]) -> None:
        raise NotImplementedError

692
693
694
695
    @abstractmethod
    def parse_image_pil(self, image_pil: Image.Image) -> None:
        raise NotImplementedError

696
697
698
699
    @abstractmethod
    def parse_audio(self, audio_url: str) -> None:
        raise NotImplementedError

700
    @abstractmethod
701
    def parse_input_audio(self, input_audio: InputAudio) -> None:
702
703
        raise NotImplementedError

704
705
706
707
    @abstractmethod
    def parse_video(self, video_url: str) -> None:
        raise NotImplementedError

708
709
710
711
712
713
714
715

class MultiModalContentParser(BaseMultiModalContentParser):

    def __init__(self, tracker: MultiModalItemTracker) -> None:
        super().__init__()

        self._tracker = tracker

716
        self._connector = MediaConnector(
717
            media_io_kwargs=self._tracker._model_config.media_io_kwargs,
718
719
720
            allowed_local_media_path=tracker.allowed_local_media_path,
        )

721
    def parse_image(self, image_url: str) -> None:
722
        image = self._connector.fetch_image(image_url)
723
724

        placeholder = self._tracker.add("image", image)
725
        self._add_placeholder("image", placeholder)
726

727
728
729
730
731
732
733
734
735
736
737
738
739
    def parse_image_embeds(self,
                           image_embeds: Union[str, dict[str, str]]) -> None:
        if isinstance(image_embeds, dict):
            embeds = {
                k: self._connector.fetch_image_embedding(v)
                for k, v in image_embeds.items()
            }
            placeholder = self._tracker.add("image_embeds", embeds)

        if isinstance(image_embeds, str):
            embedding = self._connector.fetch_image_embedding(image_embeds)
            placeholder = self._tracker.add("image_embeds", embedding)

740
        self._add_placeholder("image", placeholder)
741

742
743
    def parse_image_pil(self, image_pil: Image.Image) -> None:
        placeholder = self._tracker.add("image", image_pil)
744
        self._add_placeholder("image", placeholder)
745

746
    def parse_audio(self, audio_url: str) -> None:
747
        audio = self._connector.fetch_audio(audio_url)
748
749

        placeholder = self._tracker.add("audio", audio)
750
        self._add_placeholder("audio", placeholder)
751

752
753
754
755
    def parse_input_audio(self, input_audio: InputAudio) -> None:
        audio_data = input_audio.get("data", "")
        audio_format = input_audio.get("format", "")
        audio_url = f"data:audio/{audio_format};base64,{audio_data}"
756

757
        return self.parse_audio(audio_url)
758

759
    def parse_video(self, video_url: str) -> None:
760
        video = self._connector.fetch_video(video_url=video_url)
761
762

        placeholder = self._tracker.add("video", video)
763
        self._add_placeholder("video", placeholder)
764

765
766
767
768
769
770
771

class AsyncMultiModalContentParser(BaseMultiModalContentParser):

    def __init__(self, tracker: AsyncMultiModalItemTracker) -> None:
        super().__init__()

        self._tracker = tracker
772
        self._connector = MediaConnector(
773
774
            media_io_kwargs=self._tracker._model_config.media_io_kwargs,
            allowed_local_media_path=tracker.allowed_local_media_path
775
        )
776
777

    def parse_image(self, image_url: str) -> None:
778
        image_coro = self._connector.fetch_image_async(image_url)
779
780

        placeholder = self._tracker.add("image", image_coro)
781
        self._add_placeholder("image", placeholder)
782

783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
    def parse_image_embeds(self,
                           image_embeds: Union[str, dict[str, str]]) -> None:
        future: asyncio.Future[Union[str, dict[str, str]]] = asyncio.Future()

        if isinstance(image_embeds, dict):
            embeds = {
                k: self._connector.fetch_image_embedding(v)
                for k, v in image_embeds.items()
            }
            future.set_result(embeds)

        if isinstance(image_embeds, str):
            embedding = self._connector.\
                fetch_image_embedding(image_embeds)
            future.set_result(embedding)

        placeholder = self._tracker.add("image_embeds", future)
800
        self._add_placeholder("image", placeholder)
801

802
803
804
805
806
    def parse_image_pil(self, image_pil: Image.Image) -> None:
        future: asyncio.Future[Image.Image] = asyncio.Future()
        future.set_result(image_pil)

        placeholder = self._tracker.add("image", future)
807
        self._add_placeholder("image", placeholder)
808

809
    def parse_audio(self, audio_url: str) -> None:
810
        audio_coro = self._connector.fetch_audio_async(audio_url)
811
812

        placeholder = self._tracker.add("audio", audio_coro)
813
        self._add_placeholder("audio", placeholder)
814

815
816
817
818
    def parse_input_audio(self, input_audio: InputAudio) -> None:
        audio_data = input_audio.get("data", "")
        audio_format = input_audio.get("format", "")
        audio_url = f"data:audio/{audio_format};base64,{audio_data}"
819

820
        return self.parse_audio(audio_url)
821

822
    def parse_video(self, video_url: str) -> None:
823
        video = self._connector.fetch_video_async(video_url=video_url)
824
825

        placeholder = self._tracker.add("video", video)
826
        self._add_placeholder("video", placeholder)
827

828

829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
def validate_chat_template(chat_template: Optional[Union[Path, str]]):
    """Raises if the provided chat template appears invalid."""
    if chat_template is None:
        return

    elif isinstance(chat_template, Path) and not chat_template.exists():
        raise FileNotFoundError(
            "the supplied chat template path doesn't exist")

    elif isinstance(chat_template, str):
        JINJA_CHARS = "{}\n"
        if not any(c in chat_template
                   for c in JINJA_CHARS) and not Path(chat_template).exists():
            raise ValueError(
                f"The supplied chat template string ({chat_template}) "
                f"appears path-like, but doesn't exist!")

    else:
        raise TypeError(
            f"{type(chat_template)} is not a valid chat template type")


851
def _load_chat_template(
852
853
854
855
    chat_template: Optional[Union[Path, str]],
    *,
    is_literal: bool = False,
) -> Optional[str]:
856
857
    if chat_template is None:
        return None
858
859
860
861
862
863

    if is_literal:
        if isinstance(chat_template, Path):
            raise TypeError("chat_template is expected to be read directly "
                            "from its value")

864
        return chat_template
865

866
    try:
867
        with open(chat_template) as f:
868
            return f.read()
869
    except OSError as e:
870
871
872
        if isinstance(chat_template, Path):
            raise

873
874
875
876
877
878
        JINJA_CHARS = "{}\n"
        if not any(c in chat_template for c in JINJA_CHARS):
            msg = (f"The supplied chat template ({chat_template}) "
                   f"looks like a file path, but it failed to be "
                   f"opened. Reason: {e}")
            raise ValueError(msg) from e
879

880
881
        # If opening a file fails, set chat template to be args to
        # ensure we decode so our escape are interpreted correctly
882
883
884
885
886
887
888
889
890
891
892
893
        return _load_chat_template(chat_template, is_literal=True)


_cached_load_chat_template = lru_cache(_load_chat_template)


def load_chat_template(
    chat_template: Optional[Union[Path, str]],
    *,
    is_literal: bool = False,
) -> Optional[str]:
    return _cached_load_chat_template(chat_template, is_literal=is_literal)
894
895


896
897
898
899
900
901
902
903
904
def _get_interleaved_text_prompt(placeholder_storage: dict[str, list],
                                 texts: list[str]) -> str:
    for idx, elem in enumerate(texts):
        if elem in placeholder_storage:
            texts[idx] = placeholder_storage[elem].pop(0)

    return "\n".join(texts)


905
# TODO: Let user specify how to insert multimodal tokens into prompt
906
# (similar to chat template)
907
908
909
910
def _get_full_multimodal_text_prompt(placeholder_storage: dict[str, list],
                                     texts: list[str],
                                     interleave_strings: bool
                                     ) -> str:
911
    """Combine multimodal prompts for a multimodal language model."""
912

913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
    # flatten storage to make it looks like
    # {
    #   "<|image|>": 2,
    #   "<|audio|>": 1
    # }
    placeholder_counts = Counter(
        [v for elem in placeholder_storage.values() for v in elem]
    )

    if interleave_strings:
        text_prompt = _get_interleaved_text_prompt(placeholder_storage, texts)
    else:
        text_prompt = "\n".join(texts)

    # Pass interleaved text further in case the user used image placeholders
    # himself, but forgot to disable the 'interleave_strings' flag

930
    # Look through the text prompt to check for missing placeholders
931
    missing_placeholders: list[str] = []
932
933
934
935
936
937
    for placeholder in placeholder_counts:

        # For any existing placeholder in the text prompt, we leave it as is
        placeholder_counts[placeholder] -= text_prompt.count(placeholder)

        if placeholder_counts[placeholder] < 0:
938
939
940
941
942
943
944
            logger.error(
                "Placeholder count is negative! "
                "Ensure that the 'interleave_strings' flag is disabled "
                "(current value: %s) "
                "when manually placing image placeholders.", interleave_strings
            )
            logger.debug("Input prompt: %s", text_prompt)
945
946
947
948
949
950
951
            raise ValueError(
                f"Found more '{placeholder}' placeholders in input prompt than "
                "actual multimodal data items.")

        missing_placeholders.extend([placeholder] *
                                    placeholder_counts[placeholder])

952
953
    # NOTE: Default behaviour: we always add missing placeholders
    # at the front of the prompt, if interleave_strings=False
954
    return "\n".join(missing_placeholders + [text_prompt])
955
956


957
958
# No need to validate using Pydantic again
_TextParser = partial(cast, ChatCompletionContentPartTextParam)
959
_ImageEmbedsParser = partial(cast, ChatCompletionContentPartImageEmbedsParam)
960
_InputAudioParser = partial(cast, ChatCompletionContentPartInputAudioParam)
961
_RefusalParser = partial(cast, ChatCompletionContentPartRefusalParam)
962
_PILImageParser = partial(cast, CustomChatCompletionContentPILImageParam)
Julien Denize's avatar
Julien Denize committed
963
_ThinkParser = partial(cast, CustomThinkCompletionContentParam)
964
965
966
967
# Need to validate url objects
_ImageParser = TypeAdapter(ChatCompletionContentPartImageParam).validate_python
_AudioParser = TypeAdapter(ChatCompletionContentPartAudioParam).validate_python
_VideoParser = TypeAdapter(ChatCompletionContentPartVideoParam).validate_python
968

969
970
_ResponsesInputImageParser = TypeAdapter(
    ResponseInputImageParam).validate_python
971
_ContentPart: TypeAlias = Union[str, dict[str, str], InputAudio, PILImage]
972

973
# Define a mapping from part types to their corresponding parsing functions.
974
MM_PARSER_MAP: dict[
975
976
977
    str,
    Callable[[ChatCompletionContentPartParam], _ContentPart],
] = {
978
    "text":
979
    lambda part: _TextParser(part).get("text", None),
Julien Denize's avatar
Julien Denize committed
980
981
    "thinking":
    lambda part: _ThinkParser(part).get("thinking", None),
982
983
    "input_text":
    lambda part: _TextParser(part).get("text", None),
984
985
    "input_image":
    lambda part: _ResponsesInputImageParser(part).get("image_url", None),
986
    "image_url":
987
    lambda part: _ImageParser(part).get("image_url", {}).get("url", None),
988
    "image_embeds":
989
    lambda part: _ImageEmbedsParser(part).get("image_embeds", None),
990
    "image_pil": lambda part: _PILImageParser(part).get("image_pil", None),
991
    "audio_url":
992
    lambda part: _AudioParser(part).get("audio_url", {}).get("url", None),
993
    "input_audio":
994
    lambda part: _InputAudioParser(part).get("input_audio", None),
995
    "refusal":
996
    lambda part: _RefusalParser(part).get("refusal", None),
997
    "video_url":
998
    lambda part: _VideoParser(part).get("video_url", {}).get("url", None),
999
1000
1001
1002
}


def _parse_chat_message_content_mm_part(
1003
        part: ChatCompletionContentPartParam) -> tuple[str, _ContentPart]:
1004
    """
1005
    Parses a given multi-modal content part based on its type.
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025

    Args:
        part: A dict containing the content part, with a potential 'type' field.

    Returns:
        A tuple (part_type, content) where:
        - part_type: Type of the part (e.g., 'text', 'image_url').
        - content: Parsed content (e.g., text, image URL).

    Raises:
        ValueError: If the 'type' field is missing and no direct URL is found.
    """
    assert isinstance(
        part, dict)  # This is needed to avoid mypy errors: part.get() from str
    part_type = part.get("type", None)

    if isinstance(part_type, str) and part_type in MM_PARSER_MAP:
        content = MM_PARSER_MAP[part_type](part)

        # Special case for 'image_url.detail'
1026
1027
        # We only support 'auto', which is the default
        if part_type == "image_url" and part.get("detail", "auto") != "auto":
1028
1029
1030
1031
1032
1033
            logger.warning("'image_url.detail' is currently not supported "
                           "and will be ignored.")

        return part_type, content

    # Handle missing 'type' but provided direct URL fields.
1034
    # 'type' is required field by pydantic
1035
1036
1037
1038
1039
1040
1041
1042
1043
    if part_type is None:
        if part.get("image_url") is not None:
            image_params = cast(CustomChatCompletionContentSimpleImageParam,
                                part)
            return "image_url", image_params.get("image_url", "")
        if part.get("audio_url") is not None:
            audio_params = cast(CustomChatCompletionContentSimpleAudioParam,
                                part)
            return "audio_url", audio_params.get("audio_url", "")
1044
        if part.get("input_audio") is not None:
1045
            input_audio_params = cast(dict[str, str], part)
1046
            return "input_audio", input_audio_params
1047
1048
1049
1050
        if part.get("video_url") is not None:
            video_params = cast(CustomChatCompletionContentSimpleVideoParam,
                                part)
            return "video_url", video_params.get("video_url", "")
1051
1052
1053
1054
1055
1056
1057
1058
1059
        # Raise an error if no 'type' or direct URL is found.
        raise ValueError("Missing 'type' field in multimodal part.")

    if not isinstance(part_type, str):
        raise ValueError("Invalid 'type' field in multimodal part.")
    return part_type, "unknown part_type content"


VALID_MESSAGE_CONTENT_MM_PART_TYPES = ("text", "refusal", "image_url",
1060
                                       "image_embeds", "image_pil",
1061
                                       "audio_url", "input_audio", "video_url")
1062

1063

1064
1065
1066
def _parse_chat_message_content_parts(
    role: str,
    parts: Iterable[ChatCompletionContentPartParam],
1067
    mm_tracker: BaseMultiModalItemTracker,
1068
1069
    *,
    wrap_dicts: bool,
1070
    interleave_strings: bool,
1071
) -> list[ConversationMessage]:
1072
    content = list[_ContentPart]()
1073

1074
    mm_parser = mm_tracker.create_parser()
1075
1076

    for part in parts:
1077
        parse_res = _parse_chat_message_content_part(
1078
1079
1080
            part,
            mm_parser,
            wrap_dicts=wrap_dicts,
1081
            interleave_strings=interleave_strings
1082
        )
1083
1084
        if parse_res:
            content.append(parse_res)
1085

1086
    if wrap_dicts:
1087
        # Parsing wraps images and texts as interleaved dictionaries
1088
        return [ConversationMessage(role=role,
1089
                                    content=content)]  # type: ignore
1090
    texts = cast(list[str], content)
1091
1092
1093
1094
1095
1096
1097
1098
    mm_placeholder_storage = mm_parser.mm_placeholder_storage()
    if mm_placeholder_storage:
        text_prompt = _get_full_multimodal_text_prompt(mm_placeholder_storage,
                                                       texts,
                                                       interleave_strings)
    else:
        text_prompt = "\n".join(texts)

1099
1100
1101
1102
    return [ConversationMessage(role=role, content=text_prompt)]


def _parse_chat_message_content_part(
1103
1104
1105
1106
    part: ChatCompletionContentPartParam,
    mm_parser: BaseMultiModalContentParser,
    *,
    wrap_dicts: bool,
1107
    interleave_strings: bool,
1108
) -> Optional[_ContentPart]:
1109
1110
1111
1112
1113
1114
1115
1116
    """Parses a single part of a conversation. If wrap_dicts is True,
    structured dictionary pieces for texts and images will be
    wrapped in dictionaries, i.e., {"type": "text", "text", ...} and
    {"type": "image"}, respectively. Otherwise multimodal data will be
    handled by mm_parser, and texts will be returned as strings to be joined
    with multimodal placeholders.
    """
    if isinstance(part, str):  # Handle plain text parts
1117
        return part
1118
1119
    # Handle structured dictionary parts
    part_type, content = _parse_chat_message_content_mm_part(part)
1120
    # if part_type is text/refusal/image_url/audio_url/video_url/input_audio but
1121
1122
    # content is None, log a warning and skip
    if part_type in VALID_MESSAGE_CONTENT_MM_PART_TYPES and content is None:
1123
        logger.warning(
1124
1125
            "Skipping multimodal part '%s' (type: '%s') "
            "with empty / unparsable content.", part, part_type)
1126
1127
        return None

Julien Denize's avatar
Julien Denize committed
1128
    if part_type in ("text", "input_text", "refusal", "thinking"):
1129
1130
1131
1132
1133
        str_content = cast(str, content)
        if wrap_dicts:
            return {'type': 'text', 'text': str_content}
        else:
            return str_content
1134

1135
    modality = None
1136
1137
1138
    if part_type == "image_pil":
        image_content = cast(Image.Image, content)
        mm_parser.parse_image_pil(image_content)
1139
        modality = "image"
1140
    elif part_type in ("image_url", "input_image"):
1141
1142
        str_content = cast(str, content)
        mm_parser.parse_image(str_content)
1143
1144
        modality = "image"
    elif part_type == "image_embeds":
1145
1146
        content = cast(Union[str, dict[str, str]], content)
        mm_parser.parse_image_embeds(content)
1147
1148
        modality = "image"
    elif part_type == "audio_url":
1149
1150
        str_content = cast(str, content)
        mm_parser.parse_audio(str_content)
1151
1152
        modality = "audio"
    elif part_type == "input_audio":
1153
        dict_content = cast(InputAudio, content)
1154
        mm_parser.parse_input_audio(dict_content)
1155
1156
        modality = "audio"
    elif part_type == "video_url":
1157
1158
        str_content = cast(str, content)
        mm_parser.parse_video(str_content)
1159
1160
1161
        modality = "video"
    else:
        raise NotImplementedError(f"Unknown part type: {part_type}")
1162

1163
1164
1165
    return {'type': modality} if wrap_dicts else (
        MODALITY_PLACEHOLDERS_MAP[modality] if interleave_strings else None
    )
1166
1167


1168
1169
1170
1171
1172
# No need to validate using Pydantic again
_AssistantParser = partial(cast, ChatCompletionAssistantMessageParam)
_ToolParser = partial(cast, ChatCompletionToolMessageParam)


1173
def _parse_chat_message_content(
1174
1175
    message: ChatCompletionMessageParam,
    mm_tracker: BaseMultiModalItemTracker,
1176
    content_format: _ChatTemplateContentFormat,
1177
    interleave_strings: bool,
1178
) -> list[ConversationMessage]:
1179
1180
1181
1182
    role = message["role"]
    content = message.get("content")

    if content is None:
1183
1184
1185
1186
1187
1188
        content = []
    elif isinstance(content, str):
        content = [
            ChatCompletionContentPartTextParam(type="text", text=content)
        ]
    result = _parse_chat_message_content_parts(
1189
1190
        role,
        content,  # type: ignore
1191
        mm_tracker,
1192
        wrap_dicts=(content_format == "openai"),
1193
        interleave_strings=interleave_strings,
1194
    )
1195

1196
1197
1198
1199
    for result_msg in result:
        if role == 'assistant':
            parsed_msg = _AssistantParser(message)

1200
1201
1202
1203
1204
            # The 'tool_calls' is not None check ensures compatibility.
            # It's needed only if downstream code doesn't strictly
            # follow the OpenAI spec.
            if ("tool_calls" in parsed_msg
                and parsed_msg["tool_calls"] is not None):
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
                result_msg["tool_calls"] = list(parsed_msg["tool_calls"])
        elif role == "tool":
            parsed_msg = _ToolParser(message)
            if "tool_call_id" in parsed_msg:
                result_msg["tool_call_id"] = parsed_msg["tool_call_id"]

        if "name" in message and isinstance(message["name"], str):
            result_msg["name"] = message["name"]

    return result

1216

1217
def _postprocess_messages(messages: list[ConversationMessage]) -> None:
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
    # per the Transformers docs & maintainers, tool call arguments in
    # assistant-role messages with tool_calls need to be dicts not JSON str -
    # this is how tool-use chat templates will expect them moving forwards
    # so, for messages that have tool_calls, parse the string (which we get
    # from openAI format) to dict
    for message in messages:
        if (message["role"] == "assistant" and "tool_calls" in message
                and isinstance(message["tool_calls"], list)):

            for item in message["tool_calls"]:
                item["function"]["arguments"] = json.loads(
                    item["function"]["arguments"])


1232
def parse_chat_messages(
1233
    messages: list[ChatCompletionMessageParam],
1234
    model_config: ModelConfig,
1235
    tokenizer: AnyTokenizer,
1236
    content_format: _ChatTemplateContentFormat,
1237
1238
) -> tuple[list[ConversationMessage], Optional[MultiModalDataDict]]:
    conversation: list[ConversationMessage] = []
1239
    mm_tracker = MultiModalItemTracker(model_config, tokenizer)
1240
1241

    for msg in messages:
1242
1243
1244
        sub_messages = _parse_chat_message_content(
            msg,
            mm_tracker,
1245
            content_format,
1246
1247
1248
1249
1250
            interleave_strings=(
                content_format == "string"
                and model_config.multimodal_config is not None
                and model_config.multimodal_config.interleave_mm_strings
            )
1251
        )
1252

1253
        conversation.extend(sub_messages)
1254

1255
1256
    _postprocess_messages(conversation)

1257
    return conversation, mm_tracker.all_mm_data()
1258
1259


1260
def parse_chat_messages_futures(
1261
    messages: list[ChatCompletionMessageParam],
1262
1263
    model_config: ModelConfig,
    tokenizer: AnyTokenizer,
1264
    content_format: _ChatTemplateContentFormat,
1265
1266
) -> tuple[list[ConversationMessage], Awaitable[Optional[MultiModalDataDict]]]:
    conversation: list[ConversationMessage] = []
1267
1268
1269
    mm_tracker = AsyncMultiModalItemTracker(model_config, tokenizer)

    for msg in messages:
1270
1271
1272
        sub_messages = _parse_chat_message_content(
            msg,
            mm_tracker,
1273
            content_format,
1274
1275
1276
1277
1278
            interleave_strings=(
                content_format == "string"
                and model_config.multimodal_config is not None
                and model_config.multimodal_config.interleave_mm_strings
            )
1279
        )
1280
1281
1282

        conversation.extend(sub_messages)

1283
1284
    _postprocess_messages(conversation)

1285
1286
1287
    return conversation, mm_tracker.all_mm_data()


1288
1289
1290
1291
@deprecate_kwargs(
    "trust_remote_code",
    additional_message="Please use `model_config.trust_remote_code` instead.",
)
1292
1293
def apply_hf_chat_template(
    tokenizer: Union[PreTrainedTokenizer, PreTrainedTokenizerFast],
1294
    conversation: list[ConversationMessage],
1295
    chat_template: Optional[str],
1296
    tools: Optional[list[dict[str, Any]]],
1297
    *,
1298
    model_config: ModelConfig,
1299
    tokenize: bool = False,  # Different from HF's default
1300
1301
    # Deprecated, explicitly capture here so it doesn't slit into kwargs.
    trust_remote_code: Optional[bool] = None,
1302
    **kwargs: Any,
1303
) -> str:
1304
    hf_chat_template = resolve_hf_chat_template(
1305
1306
1307
        tokenizer,
        chat_template=chat_template,
        tools=tools,
1308
        model_config=model_config,
1309
    )
1310

1311
    if hf_chat_template is None:
1312
1313
1314
1315
1316
        raise ValueError(
            "As of transformers v4.44, default chat template is no longer "
            "allowed, so you must provide a chat template if the tokenizer "
            "does not define one.")

1317
1318
1319
1320
1321
1322
1323
1324
1325
    try:

        return tokenizer.apply_chat_template(
            conversation=conversation,  # type: ignore[arg-type]
            tools=tools,  # type: ignore[arg-type]
            chat_template=hf_chat_template,
            tokenize=tokenize,
            **kwargs,
        )
1326

1327
1328
1329
1330
1331
1332
1333
1334
    # External library exceptions can sometimes occur despite the framework's
    # internal exception management capabilities.
    except Exception as e:

        # Log and report any library-related exceptions for further
        # investigation.
        logger.exception(
            "An error occurred in `transformers` while applying chat template")
1335
        raise ValueError(str(e)) from e
1336

1337
1338
def apply_mistral_chat_template(
    tokenizer: MistralTokenizer,
1339
    messages: list[ChatCompletionMessageParam],
1340
1341
    chat_template: Optional[str],
    tools: Optional[list[dict[str, Any]]],
1342
    **kwargs: Any,
1343
) -> list[int]:
1344
1345
    from mistral_common.exceptions import MistralCommonException

1346
1347
1348
1349
1350
1351
    # The return value of resolve_mistral_chat_template is always None,
    # and we won't use it.
    resolve_mistral_chat_template(
        chat_template=chat_template,
        **kwargs,
    )
1352

1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
    try:
        return tokenizer.apply_chat_template(
            messages=messages,
            tools=tools,
            **kwargs,
        )
    # mistral-common uses assert statements to stop processing of input
    # if input does not comply with the expected format.
    # We convert those assertion errors to ValueErrors so they can be
    # are properly caught in the preprocessing_input step
1363
    except (AssertionError, MistralCommonException) as e:
1364
        raise ValueError(str(e)) from e
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374

    # External library exceptions can sometimes occur despite the framework's
    # internal exception management capabilities.
    except Exception as e:

        # Log and report any library-related exceptions for further
        # investigation.
        logger.exception(
            "An error occurred in `mistral_common` while applying chat "
            "template")
1375
        raise ValueError(str(e)) from e
1376
1377

def random_tool_call_id() -> str:
1378
    return f"chatcmpl-tool-{random_uuid()}"