chat_utils.py 34.6 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
            if model_type in ("minicpmo", "minicpmv"):
399
                return "(<image>./</image>)"
Patrick von Platen's avatar
Patrick von Platen committed
400
401
            if model_type in ("blip-2", "chatglm", "fuyu", "paligemma",
                              "pixtral"):
402
403
                # These models do not use image tokens in the prompt
                return None
404
405
            if model_type == "qwen":
                return f"Picture {current_count}: <img></img>"
406
            if model_type.startswith("llava"):
407
408
                return self._cached_token_str(self._tokenizer,
                                              hf_config.image_token_index)
409
410
            if model_type in ("chameleon", "deepseek_vl_v2", "internvl_chat",
                              "NVLM_D", "h2ovl_chat"):
411
                return "<image>"
412
413
            if model_type == "mllama":
                return "<|image|>"
Roger Wang's avatar
Roger Wang committed
414
            if model_type in ("qwen2_vl", "qwen2_5_vl"):
415
                return "<|vision_start|><|image_pad|><|vision_end|>"
416
417
            if model_type == "molmo":
                return ""
418
419
            if model_type == "idefics3":
                return "<image>"
420
421
            if model_type == "aria":
                return "<|fim_prefix|><|img|><|fim_suffix|>"
422

423
            raise TypeError(f"Unknown {modality} model type: {model_type}")
424
425
        elif modality == "audio":
            if model_type == "ultravox":
426
                return "<|audio|>"
427
428
429
            if model_type == "qwen2_audio":
                return (f"Audio {current_count}: "
                        f"<|audio_bos|><|AUDIO|><|audio_eos|>")
430
431
            if model_type == "minicpmo":
                return "(<audio>./</audio>)"
432
            raise TypeError(f"Unknown model type: {model_type}")
433
        elif modality == "video":
Roger Wang's avatar
Roger Wang committed
434
            if model_type in ("qwen2_vl", "qwen2_5_vl"):
435
                return "<|vision_start|><|video_pad|><|vision_end|>"
436
437
            if model_type in ("minicpmo", "minicpmv"):
                return "(<video>./</video>)"
438
439
440
            if model_type.startswith("llava"):
                return self._cached_token_str(self._tokenizer,
                                              hf_config.video_token_index)
441
            raise TypeError(f"Unknown {modality} model type: {model_type}")
442
443
444
        else:
            raise TypeError(f"Unknown modality: {modality}")

445
446
447
448
449
450
    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)
451
        current_count = len(self._items_by_modality[modality]) + 1
452
453
454
455
456
        if current_count > allowed_count:
            raise ValueError(
                f"At most {allowed_count} {modality}(s) may be provided in "
                "one request.")

457
        self._items_by_modality[modality].append(item)
458
459
460
461
462
463
464
465

        return self._placeholder_str(modality, current_count)

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


466
class MultiModalItemTracker(BaseMultiModalItemTracker[object]):
467
468

    def all_mm_data(self) -> Optional[MultiModalDataDict]:
469
470
471
472
        if self._items_by_modality:
            return dict(self._items_by_modality)

        return None
473
474
475
476
477

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


478
class AsyncMultiModalItemTracker(BaseMultiModalItemTracker[Awaitable[object]]):
479
480

    async def all_mm_data(self) -> Optional[MultiModalDataDict]:
481
482
483
484
485
        if self._items_by_modality:
            return {
                modality: await asyncio.gather(*items)
                for modality, items in self._items_by_modality.items()
            }
486
487
488
489
490
491
492
493
494
495
496
497
498

        return None

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


class BaseMultiModalContentParser(ABC):

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

        # multimodal placeholder_string : count
499
        self._placeholder_counts: dict[str, int] = defaultdict(lambda: 0)
500
501
502
503
504

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

505
    def mm_placeholder_counts(self) -> dict[str, int]:
506
507
508
509
510
511
512
513
514
515
        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

516
    @abstractmethod
517
    def parse_input_audio(self, input_audio: InputAudio) -> None:
518
519
        raise NotImplementedError

520
521
522
523
    @abstractmethod
    def parse_video(self, video_url: str) -> None:
        raise NotImplementedError

524
525
526
527
528
529
530
531

class MultiModalContentParser(BaseMultiModalContentParser):

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

        self._tracker = tracker

532
533
534
535
        self._connector = MediaConnector(
            allowed_local_media_path=tracker.allowed_local_media_path,
        )

536
    def parse_image(self, image_url: str) -> None:
537
        image = self._connector.fetch_image(image_url)
538
539
540
541
542

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

    def parse_audio(self, audio_url: str) -> None:
543
        audio = self._connector.fetch_audio(audio_url)
544
545
546
547

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

548
549
550
551
    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}"
552

553
        return self.parse_audio(audio_url)
554

555
    def parse_video(self, video_url: str) -> None:
556
        video = self._connector.fetch_video(video_url)
557
558
559
560

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

561
562
563
564
565
566
567

class AsyncMultiModalContentParser(BaseMultiModalContentParser):

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

        self._tracker = tracker
568
569
570
        self._connector = MediaConnector(
            allowed_local_media_path=tracker.allowed_local_media_path,
        )
571
572

    def parse_image(self, image_url: str) -> None:
573
        image_coro = self._connector.fetch_image_async(image_url)
574
575
576
577
578

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

    def parse_audio(self, audio_url: str) -> None:
579
        audio_coro = self._connector.fetch_audio_async(audio_url)
580
581
582

        placeholder = self._tracker.add("audio", audio_coro)
        self._add_placeholder(placeholder)
583

584
585
586
587
    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}"
588

589
        return self.parse_audio(audio_url)
590

591
    def parse_video(self, video_url: str) -> None:
592
        video = self._connector.fetch_video_async(video_url)
593
594
595
596

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

597

598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
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")


620
def load_chat_template(
621
622
623
624
    chat_template: Optional[Union[Path, str]],
    *,
    is_literal: bool = False,
) -> Optional[str]:
625
626
    if chat_template is None:
        return None
627
628
629
630
631
632
633
634

    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")

635
    try:
636
        with open(chat_template) as f:
637
            return f.read()
638
    except OSError as e:
639
640
641
        if isinstance(chat_template, Path):
            raise

642
643
644
645
646
647
        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
648

649
650
        # If opening a file fails, set chat template to be args to
        # ensure we decode so our escape are interpreted correctly
651
        return load_chat_template(chat_template, is_literal=True)
652
653


654
# TODO: Let user specify how to insert multimodal tokens into prompt
655
# (similar to chat template)
656
def _get_full_multimodal_text_prompt(placeholder_counts: dict[str, int],
657
                                     text_prompt: str) -> str:
658
    """Combine multimodal prompts for a multimodal language model."""
659

660
    # Look through the text prompt to check for missing placeholders
661
    missing_placeholders: list[str] = []
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
    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])
678
679


680
681
682
683
# No need to validate using Pydantic again
_TextParser = partial(cast, ChatCompletionContentPartTextParam)
_ImageParser = partial(cast, ChatCompletionContentPartImageParam)
_AudioParser = partial(cast, ChatCompletionContentPartAudioParam)
684
_InputAudioParser = partial(cast, ChatCompletionContentPartInputAudioParam)
685
_RefusalParser = partial(cast, ChatCompletionContentPartRefusalParam)
686
_VideoParser = partial(cast, ChatCompletionContentPartVideoParam)
687

688
_ContentPart: TypeAlias = Union[str, dict[str, str], InputAudio]
689

690
# Define a mapping from part types to their corresponding parsing functions.
691
MM_PARSER_MAP: dict[
692
693
694
    str,
    Callable[[ChatCompletionContentPartParam], _ContentPart],
] = {
695
696
697
698
699
700
    "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", ""),
701
702
    "input_audio":
    lambda part: _InputAudioParser(part).get("input_audio", {}),
703
704
    "refusal":
    lambda part: _RefusalParser(part).get("refusal", ""),
705
706
    "video_url":
    lambda part: _VideoParser(part).get("video_url", {}).get("url", ""),
707
708
709
710
}


def _parse_chat_message_content_mm_part(
711
        part: ChatCompletionContentPartParam) -> tuple[str, _ContentPart]:
712
    """
713
    Parses a given multi-modal content part based on its type.
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733

    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'
734
735
        # We only support 'auto', which is the default
        if part_type == "image_url" and part.get("detail", "auto") != "auto":
736
737
738
739
740
741
            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.
742
    # 'type' is required field by pydantic
743
744
745
746
747
748
749
750
751
    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", "")
752
        if part.get("input_audio") is not None:
753
            input_audio_params = cast(dict[str, str], part)
754
            return "input_audio", input_audio_params
755
756
757
758
        if part.get("video_url") is not None:
            video_params = cast(CustomChatCompletionContentSimpleVideoParam,
                                part)
            return "video_url", video_params.get("video_url", "")
759
760
761
762
763
764
765
766
767
        # 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",
768
                                       "audio_url", "input_audio", "video_url")
769

770

771
772
773
def _parse_chat_message_content_parts(
    role: str,
    parts: Iterable[ChatCompletionContentPartParam],
774
    mm_tracker: BaseMultiModalItemTracker,
775
776
    *,
    wrap_dicts: bool,
777
) -> list[ConversationMessage]:
778
    content = list[_ContentPart]()
779

780
    mm_parser = mm_tracker.create_parser()
781
782

    for part in parts:
783
        parse_res = _parse_chat_message_content_part(
784
785
786
787
            part,
            mm_parser,
            wrap_dicts=wrap_dicts,
        )
788
789
        if parse_res:
            content.append(parse_res)
790

791
    if wrap_dicts:
792
        # Parsing wraps images and texts as interleaved dictionaries
793
        return [ConversationMessage(role=role,
794
                                    content=content)]  # type: ignore
795
    texts = cast(list[str], content)
796
797
798
799
800
801
802
803
804
    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(
805
806
807
808
    part: ChatCompletionContentPartParam,
    mm_parser: BaseMultiModalContentParser,
    *,
    wrap_dicts: bool,
809
) -> Optional[_ContentPart]:
810
811
812
813
814
815
816
817
    """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
818
        return part
819
820
821
822

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

823
    # if part_type is text/refusal/image_url/audio_url/video_url/input_audio but
824
825
826
    # content is empty, log a warning and skip
    if part_type in VALID_MESSAGE_CONTENT_MM_PART_TYPES and not content:
        logger.warning(
827
            "Skipping multimodal part (type: '%s') "
828
829
830
831
            "with empty / unparsable content.", part_type)
        return None

    if part_type in ("text", "refusal"):
832
833
834
835
836
        str_content = cast(str, content)
        if wrap_dicts:
            return {'type': 'text', 'text': str_content}
        else:
            return str_content
837
838

    if part_type == "image_url":
839
840
        str_content = cast(str, content)
        mm_parser.parse_image(str_content)
841
842
843
        return {'type': 'image'} if wrap_dicts else None

    if part_type == "audio_url":
844
845
846
847
848
        str_content = cast(str, content)
        mm_parser.parse_audio(str_content)
        return {'type': 'audio'} if wrap_dicts else None

    if part_type == "input_audio":
849
        dict_content = cast(InputAudio, content)
850
        mm_parser.parse_input_audio(dict_content)
851
852
        return {'type': 'audio'} if wrap_dicts else None

853
    if part_type == "video_url":
854
855
        str_content = cast(str, content)
        mm_parser.parse_video(str_content)
856
857
        return {'type': 'video'} if wrap_dicts else None

858
    raise NotImplementedError(f"Unknown part type: {part_type}")
859
860


861
862
863
864
865
# No need to validate using Pydantic again
_AssistantParser = partial(cast, ChatCompletionAssistantMessageParam)
_ToolParser = partial(cast, ChatCompletionToolMessageParam)


866
def _parse_chat_message_content(
867
868
    message: ChatCompletionMessageParam,
    mm_tracker: BaseMultiModalItemTracker,
869
    content_format: _ChatTemplateContentFormat,
870
) -> list[ConversationMessage]:
871
872
873
874
    role = message["role"]
    content = message.get("content")

    if content is None:
875
876
877
878
879
880
        content = []
    elif isinstance(content, str):
        content = [
            ChatCompletionContentPartTextParam(type="text", text=content)
        ]
    result = _parse_chat_message_content_parts(
881
882
        role,
        content,  # type: ignore
883
        mm_tracker,
884
        wrap_dicts=(content_format == "openai"),
885
    )
886

887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
    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

903

904
def _postprocess_messages(messages: list[ConversationMessage]) -> None:
905
906
907
908
909
910
911
912
913
914
915
916
917
918
    # 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"])


919
def parse_chat_messages(
920
    messages: list[ChatCompletionMessageParam],
921
    model_config: ModelConfig,
922
    tokenizer: AnyTokenizer,
923
    content_format: _ChatTemplateContentFormat,
924
925
) -> tuple[list[ConversationMessage], Optional[MultiModalDataDict]]:
    conversation: list[ConversationMessage] = []
926
    mm_tracker = MultiModalItemTracker(model_config, tokenizer)
927
928

    for msg in messages:
929
930
931
        sub_messages = _parse_chat_message_content(
            msg,
            mm_tracker,
932
            content_format,
933
        )
934

935
        conversation.extend(sub_messages)
936

937
938
    _postprocess_messages(conversation)

939
    return conversation, mm_tracker.all_mm_data()
940
941


942
def parse_chat_messages_futures(
943
    messages: list[ChatCompletionMessageParam],
944
945
    model_config: ModelConfig,
    tokenizer: AnyTokenizer,
946
    content_format: _ChatTemplateContentFormat,
947
948
) -> tuple[list[ConversationMessage], Awaitable[Optional[MultiModalDataDict]]]:
    conversation: list[ConversationMessage] = []
949
950
951
    mm_tracker = AsyncMultiModalItemTracker(model_config, tokenizer)

    for msg in messages:
952
953
954
        sub_messages = _parse_chat_message_content(
            msg,
            mm_tracker,
955
            content_format,
956
        )
957
958
959

        conversation.extend(sub_messages)

960
961
    _postprocess_messages(conversation)

962
963
964
    return conversation, mm_tracker.all_mm_data()


965
966
def apply_hf_chat_template(
    tokenizer: Union[PreTrainedTokenizer, PreTrainedTokenizerFast],
967
    conversation: list[ConversationMessage],
968
969
970
971
    chat_template: Optional[str],
    *,
    tokenize: bool = False,  # Different from HF's default
    **kwargs: Any,
972
) -> str:
973
974
975
976
977
978
    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.")

979
980
981
982
983
984
    return tokenizer.apply_chat_template(
        conversation=conversation,  # type: ignore[arg-type]
        chat_template=chat_template,
        tokenize=tokenize,
        **kwargs,
    )
985
986


987
988
def apply_mistral_chat_template(
    tokenizer: MistralTokenizer,
989
    messages: list[ChatCompletionMessageParam],
990
    chat_template: Optional[str] = None,
991
    **kwargs: Any,
992
) -> list[int]:
993
    if chat_template is not None:
994
        logger.warning_once(
995
            "'chat_template' cannot be overridden for mistral tokenizer.")
996
    if "add_generation_prompt" in kwargs:
997
        logger.warning_once(
998
999
1000
            "'add_generation_prompt' is not supported for mistral tokenizer, "
            "so it will be ignored.")
    if "continue_final_message" in kwargs:
1001
        logger.warning_once(
1002
1003
            "'continue_final_message' is not supported for mistral tokenizer, "
            "so it will be ignored.")
1004

1005
1006
    return tokenizer.apply_chat_template(
        messages=messages,
1007
1008
        **kwargs,
    )