chat_utils.py 34.9 KB
Newer Older
1
2
# SPDX-License-Identifier: Apache-2.0

3
import asyncio
4
import codecs
5
import json
6
from abc import ABC, abstractmethod
7
from collections import defaultdict, deque
8
from collections.abc import Awaitable, Iterable
9
from functools import cache, 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
32
# yapf: enable
# pydantic needs the TypedDict from typing_extensions
33
from transformers import PreTrainedTokenizer, PreTrainedTokenizerFast
34
from typing_extensions import Required, TypeAlias, TypedDict
35

36
from vllm.config import ModelConfig
37
38
from vllm.logger import init_logger
from vllm.multimodal import MultiModalDataDict
39
from vllm.multimodal.utils import MediaConnector
40
from vllm.transformers_utils.tokenizer import AnyTokenizer, MistralTokenizer
41
42
43
44

logger = init_logger(__name__)


45
46
47
48
49
50
51
52
53
54
55
56
57
58
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."""


59
60
61
62
63
64
65
66
67
68
69
70
71
72
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."""


73
74
75
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.
76

77
78
79
80
81
82
83
84
85
86
    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.
87

88
89
90
91
92
93
94
95
    Example:
    {
        "audio_url": "https://example.com/audio.mp3"
    }
    """
    audio_url: Required[str]


96
97
98
99
100
101
102
103
104
105
106
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]


107
108
ChatCompletionContentPartParam: TypeAlias = Union[
    OpenAIChatCompletionContentPartParam, ChatCompletionContentPartAudioParam,
109
    ChatCompletionContentPartInputAudioParam,
110
    ChatCompletionContentPartVideoParam, ChatCompletionContentPartRefusalParam,
111
    CustomChatCompletionContentSimpleImageParam,
112
113
    CustomChatCompletionContentSimpleAudioParam,
    CustomChatCompletionContentSimpleVideoParam, str]
114
115
116
117
118
119
120


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

121
    content: Union[str, list[ChatCompletionContentPartParam]]
122
123
124
125
126
127
128
129
130
    """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.
    """

131
132
133
134
135
136
    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."""

137
138
139
140
141

ChatCompletionMessageParam = Union[OpenAIChatCompletionMessageParam,
                                   CustomChatCompletionMessageParam]


142
# TODO: Make fields ReadOnly once mypy supports it
143
144
145
146
class ConversationMessage(TypedDict, total=False):
    role: Required[str]
    """The role of the message's author."""

147
    content: Union[Optional[str], list[dict[str, str]]]
148
149
150
151
152
153
154
155
156
157
    """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."""
158
159


160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
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
347
348
349
350
351
352
# 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


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"


def _resolve_chat_template_content_format(
    chat_template: Optional[str],
    given_format: ChatTemplateContentFormatOption,
    tokenizer: AnyTokenizer,
) -> _ChatTemplateContentFormat:
    if isinstance(tokenizer, (PreTrainedTokenizer, PreTrainedTokenizerFast)):
        tokenizer_chat_template = tokenizer.chat_template
    else:
        tokenizer_chat_template = None

    jinja_text: Optional[str]
    if isinstance(tokenizer_chat_template, str) and chat_template is None:
        jinja_text = tokenizer_chat_template
    elif (isinstance(tokenizer_chat_template, dict)
            and chat_template in tokenizer_chat_template):
        jinja_text = tokenizer_chat_template[chat_template]
    else:
        jinja_text = load_chat_template(chat_template, is_literal=True)

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

    return detected_format if given_format == "auto" else given_format


@lru_cache
def resolve_chat_template_content_format(
    chat_template: Optional[str],
    given_format: ChatTemplateContentFormatOption,
    tokenizer: AnyTokenizer,
) -> _ChatTemplateContentFormat:
    detected_format = _resolve_chat_template_content_format(
        chat_template,
        given_format,
        tokenizer,
    )

    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,
        )

    return detected_format


353
ModalityStr = Literal["image", "audio", "video"]
354
355
356
357
_T = TypeVar("_T")


class BaseMultiModalItemTracker(ABC, Generic[_T]):
358
359
360
361
362
363
364
    """
    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):
365
366
        super().__init__()

367
368
369
370
        self._model_config = model_config
        self._tokenizer = tokenizer
        self._allowed_items = (model_config.multimodal_config.limit_per_prompt
                               if model_config.multimodal_config else {})
371

372
        self._items_by_modality = defaultdict[str, list[_T]](list)
373

374
375
376
377
    @property
    def model_config(self) -> ModelConfig:
        return self._model_config

378
379
380
381
    @property
    def allowed_local_media_path(self):
        return self._model_config.allowed_local_media_path

382
    @staticmethod
383
    @cache
384
    def _cached_token_str(tokenizer: AnyTokenizer, token_index: int) -> str:
385
386
        return tokenizer.decode(token_index)

387
388
    def _placeholder_str(self, modality: ModalityStr,
                         current_count: int) -> Optional[str]:
389
390
        # TODO: Let user specify how to insert image tokens into prompt
        # (similar to chat template)
391
392
393
        hf_config = self._model_config.hf_config
        model_type = hf_config.model_type

394
395
396
397
        if modality == "image":
            if model_type == "phi3_v":
                # Workaround since this token is not defined in the tokenizer
                return f"<|image_{current_count}|>"
398
399
            if model_type == "phi4mm":
                return "<|endoftext10|>"  # 200010 (see vocab.json in hf model)
400
            if model_type in ("minicpmo", "minicpmv"):
401
                return "(<image>./</image>)"
Patrick von Platen's avatar
Patrick von Platen committed
402
403
            if model_type in ("blip-2", "chatglm", "fuyu", "paligemma",
                              "pixtral"):
404
405
                # These models do not use image tokens in the prompt
                return None
406
407
            if model_type == "qwen":
                return f"Picture {current_count}: <img></img>"
408
            if model_type.startswith("llava"):
409
410
                return self._cached_token_str(self._tokenizer,
                                              hf_config.image_token_index)
411
412
            if model_type in ("chameleon", "deepseek_vl_v2", "internvl_chat",
                              "NVLM_D", "h2ovl_chat"):
413
                return "<image>"
414
415
            if model_type == "mllama":
                return "<|image|>"
Roger Wang's avatar
Roger Wang committed
416
            if model_type in ("qwen2_vl", "qwen2_5_vl"):
417
                return "<|vision_start|><|image_pad|><|vision_end|>"
418
419
            if model_type == "molmo":
                return ""
420
421
            if model_type == "idefics3":
                return "<image>"
422
423
            if model_type == "aria":
                return "<|fim_prefix|><|img|><|fim_suffix|>"
424

425
            raise TypeError(f"Unknown {modality} model type: {model_type}")
426
427
        elif modality == "audio":
            if model_type == "ultravox":
428
                return "<|audio|>"
429
430
            if model_type == "phi4mm":
                return "<|endoftext11|>"  # 200011 (see vocab.json in hf model)
431
432
433
            if model_type == "qwen2_audio":
                return (f"Audio {current_count}: "
                        f"<|audio_bos|><|AUDIO|><|audio_eos|>")
434
435
            if model_type == "minicpmo":
                return "(<audio>./</audio>)"
436
            raise TypeError(f"Unknown model type: {model_type}")
437
        elif modality == "video":
Roger Wang's avatar
Roger Wang committed
438
            if model_type in ("qwen2_vl", "qwen2_5_vl"):
439
                return "<|vision_start|><|video_pad|><|vision_end|>"
440
441
            if model_type in ("minicpmo", "minicpmv"):
                return "(<video>./</video>)"
442
443
444
            if model_type.startswith("llava"):
                return self._cached_token_str(self._tokenizer,
                                              hf_config.video_token_index)
445
            raise TypeError(f"Unknown {modality} model type: {model_type}")
446
447
448
        else:
            raise TypeError(f"Unknown modality: {modality}")

449
450
451
452
453
454
    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.
        """
        allowed_count = self._allowed_items.get(modality, 1)
455
        current_count = len(self._items_by_modality[modality]) + 1
456
457
458
459
460
        if current_count > allowed_count:
            raise ValueError(
                f"At most {allowed_count} {modality}(s) may be provided in "
                "one request.")

461
        self._items_by_modality[modality].append(item)
462
463
464
465
466
467
468
469

        return self._placeholder_str(modality, current_count)

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


470
class MultiModalItemTracker(BaseMultiModalItemTracker[object]):
471
472

    def all_mm_data(self) -> Optional[MultiModalDataDict]:
473
474
475
476
        if self._items_by_modality:
            return dict(self._items_by_modality)

        return None
477
478
479
480
481

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


482
class AsyncMultiModalItemTracker(BaseMultiModalItemTracker[Awaitable[object]]):
483
484

    async def all_mm_data(self) -> Optional[MultiModalDataDict]:
485
486
487
488
489
        if self._items_by_modality:
            return {
                modality: await asyncio.gather(*items)
                for modality, items in self._items_by_modality.items()
            }
490
491
492
493
494
495
496
497
498
499
500
501
502

        return None

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


class BaseMultiModalContentParser(ABC):

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

        # multimodal placeholder_string : count
503
        self._placeholder_counts: dict[str, int] = defaultdict(lambda: 0)
504
505
506
507
508

    def _add_placeholder(self, placeholder: Optional[str]):
        if placeholder:
            self._placeholder_counts[placeholder] += 1

509
    def mm_placeholder_counts(self) -> dict[str, int]:
510
511
512
513
514
515
516
517
518
519
        return dict(self._placeholder_counts)

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

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

520
    @abstractmethod
521
    def parse_input_audio(self, input_audio: InputAudio) -> None:
522
523
        raise NotImplementedError

524
525
526
527
    @abstractmethod
    def parse_video(self, video_url: str) -> None:
        raise NotImplementedError

528
529
530
531
532
533
534
535

class MultiModalContentParser(BaseMultiModalContentParser):

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

        self._tracker = tracker

536
537
538
539
        self._connector = MediaConnector(
            allowed_local_media_path=tracker.allowed_local_media_path,
        )

540
    def parse_image(self, image_url: str) -> None:
541
        image = self._connector.fetch_image(image_url)
542
543
544
545
546

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

    def parse_audio(self, audio_url: str) -> None:
547
        audio = self._connector.fetch_audio(audio_url)
548
549
550
551

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

552
553
554
555
    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}"
556

557
        return self.parse_audio(audio_url)
558

559
    def parse_video(self, video_url: str) -> None:
560
        video = self._connector.fetch_video(video_url)
561
562
563
564

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

565
566
567
568
569
570
571

class AsyncMultiModalContentParser(BaseMultiModalContentParser):

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

        self._tracker = tracker
572
573
574
        self._connector = MediaConnector(
            allowed_local_media_path=tracker.allowed_local_media_path,
        )
575
576

    def parse_image(self, image_url: str) -> None:
577
        image_coro = self._connector.fetch_image_async(image_url)
578
579
580
581
582

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

    def parse_audio(self, audio_url: str) -> None:
583
        audio_coro = self._connector.fetch_audio_async(audio_url)
584
585
586

        placeholder = self._tracker.add("audio", audio_coro)
        self._add_placeholder(placeholder)
587

588
589
590
591
    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}"
592

593
        return self.parse_audio(audio_url)
594

595
    def parse_video(self, video_url: str) -> None:
596
        video = self._connector.fetch_video_async(video_url)
597
598
599
600

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

601

602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
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")


624
def load_chat_template(
625
626
627
628
    chat_template: Optional[Union[Path, str]],
    *,
    is_literal: bool = False,
) -> Optional[str]:
629
630
    if chat_template is None:
        return None
631
632
633
634
635
636
637
638

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

        return codecs.decode(chat_template, "unicode_escape")

639
    try:
640
        with open(chat_template) as f:
641
            return f.read()
642
    except OSError as e:
643
644
645
        if isinstance(chat_template, Path):
            raise

646
647
648
649
650
651
        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
652

653
654
        # If opening a file fails, set chat template to be args to
        # ensure we decode so our escape are interpreted correctly
655
        return load_chat_template(chat_template, is_literal=True)
656
657


658
# TODO: Let user specify how to insert multimodal tokens into prompt
659
# (similar to chat template)
660
def _get_full_multimodal_text_prompt(placeholder_counts: dict[str, int],
661
                                     text_prompt: str) -> str:
662
    """Combine multimodal prompts for a multimodal language model."""
663

664
    # Look through the text prompt to check for missing placeholders
665
    missing_placeholders: list[str] = []
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
    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:
            raise ValueError(
                f"Found more '{placeholder}' placeholders in input prompt than "
                "actual multimodal data items.")

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

    # NOTE: For now we always add missing placeholders at the front of
    # the prompt. This may change to be customizable in the future.
    return "\n".join(missing_placeholders + [text_prompt])
682
683


684
685
686
687
# No need to validate using Pydantic again
_TextParser = partial(cast, ChatCompletionContentPartTextParam)
_ImageParser = partial(cast, ChatCompletionContentPartImageParam)
_AudioParser = partial(cast, ChatCompletionContentPartAudioParam)
688
_InputAudioParser = partial(cast, ChatCompletionContentPartInputAudioParam)
689
_RefusalParser = partial(cast, ChatCompletionContentPartRefusalParam)
690
_VideoParser = partial(cast, ChatCompletionContentPartVideoParam)
691

692
_ContentPart: TypeAlias = Union[str, dict[str, str], InputAudio]
693

694
# Define a mapping from part types to their corresponding parsing functions.
695
MM_PARSER_MAP: dict[
696
697
698
    str,
    Callable[[ChatCompletionContentPartParam], _ContentPart],
] = {
699
700
701
702
703
704
    "text":
    lambda part: _TextParser(part).get("text", ""),
    "image_url":
    lambda part: _ImageParser(part).get("image_url", {}).get("url", ""),
    "audio_url":
    lambda part: _AudioParser(part).get("audio_url", {}).get("url", ""),
705
706
    "input_audio":
    lambda part: _InputAudioParser(part).get("input_audio", {}),
707
708
    "refusal":
    lambda part: _RefusalParser(part).get("refusal", ""),
709
710
    "video_url":
    lambda part: _VideoParser(part).get("video_url", {}).get("url", ""),
711
712
713
714
}


def _parse_chat_message_content_mm_part(
715
        part: ChatCompletionContentPartParam) -> tuple[str, _ContentPart]:
716
    """
717
    Parses a given multi-modal content part based on its type.
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737

    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'
738
739
        # We only support 'auto', which is the default
        if part_type == "image_url" and part.get("detail", "auto") != "auto":
740
741
742
743
744
745
            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.
746
    # 'type' is required field by pydantic
747
748
749
750
751
752
753
754
755
    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", "")
756
        if part.get("input_audio") is not None:
757
            input_audio_params = cast(dict[str, str], part)
758
            return "input_audio", input_audio_params
759
760
761
762
        if part.get("video_url") is not None:
            video_params = cast(CustomChatCompletionContentSimpleVideoParam,
                                part)
            return "video_url", video_params.get("video_url", "")
763
764
765
766
767
768
769
770
771
        # 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",
772
                                       "audio_url", "input_audio", "video_url")
773

774

775
776
777
def _parse_chat_message_content_parts(
    role: str,
    parts: Iterable[ChatCompletionContentPartParam],
778
    mm_tracker: BaseMultiModalItemTracker,
779
780
    *,
    wrap_dicts: bool,
781
) -> list[ConversationMessage]:
782
    content = list[_ContentPart]()
783

784
    mm_parser = mm_tracker.create_parser()
785
786

    for part in parts:
787
        parse_res = _parse_chat_message_content_part(
788
789
790
791
            part,
            mm_parser,
            wrap_dicts=wrap_dicts,
        )
792
793
        if parse_res:
            content.append(parse_res)
794

795
    if wrap_dicts:
796
        # Parsing wraps images and texts as interleaved dictionaries
797
        return [ConversationMessage(role=role,
798
                                    content=content)]  # type: ignore
799
    texts = cast(list[str], content)
800
801
802
803
804
805
806
807
808
    text_prompt = "\n".join(texts)
    mm_placeholder_counts = mm_parser.mm_placeholder_counts()
    if mm_placeholder_counts:
        text_prompt = _get_full_multimodal_text_prompt(mm_placeholder_counts,
                                                       text_prompt)
    return [ConversationMessage(role=role, content=text_prompt)]


def _parse_chat_message_content_part(
809
810
811
812
    part: ChatCompletionContentPartParam,
    mm_parser: BaseMultiModalContentParser,
    *,
    wrap_dicts: bool,
813
) -> Optional[_ContentPart]:
814
815
816
817
818
819
820
821
    """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
822
        return part
823
824
825
826

    # Handle structured dictionary parts
    part_type, content = _parse_chat_message_content_mm_part(part)

827
    # if part_type is text/refusal/image_url/audio_url/video_url/input_audio but
828
829
830
    # content is empty, log a warning and skip
    if part_type in VALID_MESSAGE_CONTENT_MM_PART_TYPES and not content:
        logger.warning(
831
            "Skipping multimodal part (type: '%s') "
832
833
834
835
            "with empty / unparsable content.", part_type)
        return None

    if part_type in ("text", "refusal"):
836
837
838
839
840
        str_content = cast(str, content)
        if wrap_dicts:
            return {'type': 'text', 'text': str_content}
        else:
            return str_content
841
842

    if part_type == "image_url":
843
844
        str_content = cast(str, content)
        mm_parser.parse_image(str_content)
845
846
847
        return {'type': 'image'} if wrap_dicts else None

    if part_type == "audio_url":
848
849
850
851
852
        str_content = cast(str, content)
        mm_parser.parse_audio(str_content)
        return {'type': 'audio'} if wrap_dicts else None

    if part_type == "input_audio":
853
        dict_content = cast(InputAudio, content)
854
        mm_parser.parse_input_audio(dict_content)
855
856
        return {'type': 'audio'} if wrap_dicts else None

857
    if part_type == "video_url":
858
859
        str_content = cast(str, content)
        mm_parser.parse_video(str_content)
860
861
        return {'type': 'video'} if wrap_dicts else None

862
    raise NotImplementedError(f"Unknown part type: {part_type}")
863
864


865
866
867
868
869
# No need to validate using Pydantic again
_AssistantParser = partial(cast, ChatCompletionAssistantMessageParam)
_ToolParser = partial(cast, ChatCompletionToolMessageParam)


870
def _parse_chat_message_content(
871
872
    message: ChatCompletionMessageParam,
    mm_tracker: BaseMultiModalItemTracker,
873
    content_format: _ChatTemplateContentFormat,
874
) -> list[ConversationMessage]:
875
876
877
878
    role = message["role"]
    content = message.get("content")

    if content is None:
879
880
881
882
883
884
        content = []
    elif isinstance(content, str):
        content = [
            ChatCompletionContentPartTextParam(type="text", text=content)
        ]
    result = _parse_chat_message_content_parts(
885
886
        role,
        content,  # type: ignore
887
        mm_tracker,
888
        wrap_dicts=(content_format == "openai"),
889
    )
890

891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
    for result_msg in result:
        if role == 'assistant':
            parsed_msg = _AssistantParser(message)

            if "tool_calls" in parsed_msg:
                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

907

908
def _postprocess_messages(messages: list[ConversationMessage]) -> None:
909
910
911
912
913
914
915
916
917
918
919
920
921
922
    # 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"])


923
def parse_chat_messages(
924
    messages: list[ChatCompletionMessageParam],
925
    model_config: ModelConfig,
926
    tokenizer: AnyTokenizer,
927
    content_format: _ChatTemplateContentFormat,
928
929
) -> tuple[list[ConversationMessage], Optional[MultiModalDataDict]]:
    conversation: list[ConversationMessage] = []
930
    mm_tracker = MultiModalItemTracker(model_config, tokenizer)
931
932

    for msg in messages:
933
934
935
        sub_messages = _parse_chat_message_content(
            msg,
            mm_tracker,
936
            content_format,
937
        )
938

939
        conversation.extend(sub_messages)
940

941
942
    _postprocess_messages(conversation)

943
    return conversation, mm_tracker.all_mm_data()
944
945


946
def parse_chat_messages_futures(
947
    messages: list[ChatCompletionMessageParam],
948
949
    model_config: ModelConfig,
    tokenizer: AnyTokenizer,
950
    content_format: _ChatTemplateContentFormat,
951
952
) -> tuple[list[ConversationMessage], Awaitable[Optional[MultiModalDataDict]]]:
    conversation: list[ConversationMessage] = []
953
954
955
    mm_tracker = AsyncMultiModalItemTracker(model_config, tokenizer)

    for msg in messages:
956
957
958
        sub_messages = _parse_chat_message_content(
            msg,
            mm_tracker,
959
            content_format,
960
        )
961
962
963

        conversation.extend(sub_messages)

964
965
    _postprocess_messages(conversation)

966
967
968
    return conversation, mm_tracker.all_mm_data()


969
970
def apply_hf_chat_template(
    tokenizer: Union[PreTrainedTokenizer, PreTrainedTokenizerFast],
971
    conversation: list[ConversationMessage],
972
973
974
975
    chat_template: Optional[str],
    *,
    tokenize: bool = False,  # Different from HF's default
    **kwargs: Any,
976
) -> str:
977
978
979
980
981
982
    if chat_template is None and tokenizer.chat_template is None:
        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.")

983
984
985
986
987
988
    return tokenizer.apply_chat_template(
        conversation=conversation,  # type: ignore[arg-type]
        chat_template=chat_template,
        tokenize=tokenize,
        **kwargs,
    )
989
990


991
992
def apply_mistral_chat_template(
    tokenizer: MistralTokenizer,
993
    messages: list[ChatCompletionMessageParam],
994
    chat_template: Optional[str] = None,
995
    **kwargs: Any,
996
) -> list[int]:
997
    if chat_template is not None:
998
        logger.warning_once(
999
            "'chat_template' cannot be overridden for mistral tokenizer.")
1000
    if "add_generation_prompt" in kwargs:
1001
        logger.warning_once(
1002
1003
1004
            "'add_generation_prompt' is not supported for mistral tokenizer, "
            "so it will be ignored.")
    if "continue_final_message" in kwargs:
1005
        logger.warning_once(
1006
1007
            "'continue_final_message' is not supported for mistral tokenizer, "
            "so it will be ignored.")
1008

1009
1010
    return tokenizer.apply_chat_template(
        messages=messages,
1011
1012
        **kwargs,
    )