chat_utils.py 39 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
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."""


70
71
72
73
74
75
76
77
78
79
80
81
82
83
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."""


84
85
86
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.
87

88
89
90
91
92
93
94
95
96
97
    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.
98

99
100
101
102
103
104
105
106
    Example:
    {
        "audio_url": "https://example.com/audio.mp3"
    }
    """
    audio_url: Required[str]


107
108
109
110
111
112
113
114
115
116
117
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]


118
119
ChatCompletionContentPartParam: TypeAlias = Union[
    OpenAIChatCompletionContentPartParam, ChatCompletionContentPartAudioParam,
120
    ChatCompletionContentPartInputAudioParam,
121
    ChatCompletionContentPartVideoParam, ChatCompletionContentPartRefusalParam,
122
    CustomChatCompletionContentSimpleImageParam,
123
    ChatCompletionContentPartImageEmbedsParam,
124
125
    CustomChatCompletionContentSimpleAudioParam,
    CustomChatCompletionContentSimpleVideoParam, str]
126
127
128
129
130
131
132


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

133
    content: Union[str, list[ChatCompletionContentPartParam]]
134
135
136
137
138
139
140
141
142
    """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.
    """

143
144
145
146
147
148
    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."""

149
150
151
152
153

ChatCompletionMessageParam = Union[OpenAIChatCompletionMessageParam,
                                   CustomChatCompletionMessageParam]


154
# TODO: Make fields ReadOnly once mypy supports it
155
156
157
158
class ConversationMessage(TypedDict, total=False):
    role: Required[str]
    """The role of the message's author."""

159
    content: Union[Optional[str], list[dict[str, str]]]
160
161
162
163
164
165
166
167
168
169
    """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."""
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
353
354
355
356
357
358
359
360
361
362
363
364
# 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


365
ModalityStr = Literal["image", "audio", "video", "image_embeds"]
366
367
368
369
_T = TypeVar("_T")


class BaseMultiModalItemTracker(ABC, Generic[_T]):
370
371
372
373
374
375
376
    """
    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):
377
378
        super().__init__()

379
380
381
382
        self._model_config = model_config
        self._tokenizer = tokenizer
        self._allowed_items = (model_config.multimodal_config.limit_per_prompt
                               if model_config.multimodal_config else {})
383

384
        self._items_by_modality = defaultdict[str, list[_T]](list)
385

386
387
388
389
    @property
    def model_config(self) -> ModelConfig:
        return self._model_config

390
391
392
393
    @property
    def allowed_local_media_path(self):
        return self._model_config.allowed_local_media_path

394
    @staticmethod
395
    @cache
396
    def _cached_token_str(tokenizer: AnyTokenizer, token_index: int) -> str:
397
398
        return tokenizer.decode(token_index)

399
400
    def _placeholder_str(self, modality: ModalityStr,
                         current_count: int) -> Optional[str]:
401
402
        # TODO: Let user specify how to insert image tokens into prompt
        # (similar to chat template)
403
404
405
        hf_config = self._model_config.hf_config
        model_type = hf_config.model_type

406
        if modality in ["image", "image_embeds"]:
407
408
409
            if model_type == "phi3_v":
                # Workaround since this token is not defined in the tokenizer
                return f"<|image_{current_count}|>"
410
411
            if model_type == "phi4mm":
                return "<|endoftext10|>"  # 200010 (see vocab.json in hf model)
412
            if model_type in ("minicpmo", "minicpmv"):
413
                return "(<image>./</image>)"
Patrick von Platen's avatar
Patrick von Platen committed
414
415
            if model_type in ("blip-2", "chatglm", "fuyu", "paligemma",
                              "pixtral"):
416
417
                # These models do not use image tokens in the prompt
                return None
418
419
            if model_type == "qwen":
                return f"Picture {current_count}: <img></img>"
420
            if model_type.startswith("llava"):
421
422
                return self._cached_token_str(self._tokenizer,
                                              hf_config.image_token_index)
423
424
            if model_type in ("chameleon", "deepseek_vl_v2", "internvl_chat",
                              "NVLM_D", "h2ovl_chat"):
425
                return "<image>"
426
427
            if model_type == "mllama":
                return "<|image|>"
Roger Wang's avatar
Roger Wang committed
428
            if model_type in ("qwen2_vl", "qwen2_5_vl"):
429
                return "<|vision_start|><|image_pad|><|vision_end|>"
430
431
            if model_type == "molmo":
                return ""
432
433
            if model_type == "idefics3":
                return "<image>"
434
435
            if model_type == "aria":
                return "<|fim_prefix|><|img|><|fim_suffix|>"
436
437
            if model_type == "gemma3":
                return "<start_of_image>"
438

439
            raise TypeError(f"Unknown {modality} model type: {model_type}")
440
441
        elif modality == "audio":
            if model_type == "ultravox":
442
                return "<|audio|>"
443
444
            if model_type == "phi4mm":
                return "<|endoftext11|>"  # 200011 (see vocab.json in hf model)
445
446
447
            if model_type == "qwen2_audio":
                return (f"Audio {current_count}: "
                        f"<|audio_bos|><|AUDIO|><|audio_eos|>")
448
449
            if model_type == "minicpmo":
                return "(<audio>./</audio>)"
450
            raise TypeError(f"Unknown model type: {model_type}")
451
        elif modality == "video":
Roger Wang's avatar
Roger Wang committed
452
            if model_type in ("qwen2_vl", "qwen2_5_vl"):
453
                return "<|vision_start|><|video_pad|><|vision_end|>"
454
455
            if model_type in ("minicpmo", "minicpmv"):
                return "(<video>./</video>)"
456
457
458
            if model_type.startswith("llava"):
                return self._cached_token_str(self._tokenizer,
                                              hf_config.video_token_index)
459
            raise TypeError(f"Unknown {modality} model type: {model_type}")
460
461
462
        else:
            raise TypeError(f"Unknown modality: {modality}")

463
464
465
466
467
468
    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)
469
        current_count = len(self._items_by_modality[modality]) + 1
470
471
472
473
474
        if current_count > allowed_count:
            raise ValueError(
                f"At most {allowed_count} {modality}(s) may be provided in "
                "one request.")

475
        self._items_by_modality[modality].append(item)
476
477
478
479
480
481
482
483

        return self._placeholder_str(modality, current_count)

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


484
class MultiModalItemTracker(BaseMultiModalItemTracker[object]):
485
486

    def all_mm_data(self) -> Optional[MultiModalDataDict]:
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
        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]
        elif "image" in items_by_modality:
            mm_inputs["image"] = items_by_modality["image"] # A list of images
        elif "audio" in items_by_modality:
            mm_inputs["audio"] = items_by_modality["audio"] # A list of audios
        elif "video" in items_by_modality:
            mm_inputs["video"] = items_by_modality["video"] # A list of videos
        return mm_inputs
508
509
510
511
512

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


513
class AsyncMultiModalItemTracker(BaseMultiModalItemTracker[Awaitable[object]]):
514
515

    async def all_mm_data(self) -> Optional[MultiModalDataDict]:
516
517
518
519
        if not self._items_by_modality:
            return None
        mm_inputs = {}
        items_by_modality = {
520
521
522
                modality: await asyncio.gather(*items)
                for modality, items in self._items_by_modality.items()
            }
523

524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
        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]
        elif "image" in items_by_modality:
            mm_inputs["image"] = items_by_modality["image"] # A list of images
        elif "audio" in items_by_modality:
            mm_inputs["audio"] = items_by_modality["audio"] # A list of audios
        elif "video" in items_by_modality:
            mm_inputs["video"] = items_by_modality["video"] # A list of videos
        return mm_inputs
541
542
543
544
545
546
547
548
549
550
551

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


class BaseMultiModalContentParser(ABC):

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

        # multimodal placeholder_string : count
552
        self._placeholder_counts: dict[str, int] = defaultdict(lambda: 0)
553
554
555
556
557

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

558
    def mm_placeholder_counts(self) -> dict[str, int]:
559
560
561
562
563
564
        return dict(self._placeholder_counts)

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

565
566
567
568
569
    @abstractmethod
    def parse_image_embeds(self,
                           image_embeds: Union[str, dict[str, str]]) -> None:
        raise NotImplementedError

570
571
572
573
    @abstractmethod
    def parse_audio(self, audio_url: str) -> None:
        raise NotImplementedError

574
    @abstractmethod
575
    def parse_input_audio(self, input_audio: InputAudio) -> None:
576
577
        raise NotImplementedError

578
579
580
581
    @abstractmethod
    def parse_video(self, video_url: str) -> None:
        raise NotImplementedError

582
583
584
585
586
587
588
589

class MultiModalContentParser(BaseMultiModalContentParser):

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

        self._tracker = tracker

590
591
592
593
        self._connector = MediaConnector(
            allowed_local_media_path=tracker.allowed_local_media_path,
        )

594
    def parse_image(self, image_url: str) -> None:
595
        image = self._connector.fetch_image(image_url)
596
597
598
599

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

600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
    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)

        self._add_placeholder(placeholder)

615
    def parse_audio(self, audio_url: str) -> None:
616
        audio = self._connector.fetch_audio(audio_url)
617
618
619
620

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

621
622
623
624
    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}"
625

626
        return self.parse_audio(audio_url)
627

628
    def parse_video(self, video_url: str) -> None:
629
        video = self._connector.fetch_video(video_url)
630
631
632
633

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

634
635
636
637
638
639
640

class AsyncMultiModalContentParser(BaseMultiModalContentParser):

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

        self._tracker = tracker
641
642
643
        self._connector = MediaConnector(
            allowed_local_media_path=tracker.allowed_local_media_path,
        )
644
645

    def parse_image(self, image_url: str) -> None:
646
        image_coro = self._connector.fetch_image_async(image_url)
647
648
649
650

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

651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
    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)
        self._add_placeholder(placeholder)

670
    def parse_audio(self, audio_url: str) -> None:
671
        audio_coro = self._connector.fetch_audio_async(audio_url)
672
673
674

        placeholder = self._tracker.add("audio", audio_coro)
        self._add_placeholder(placeholder)
675

676
677
678
679
    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}"
680

681
        return self.parse_audio(audio_url)
682

683
    def parse_video(self, video_url: str) -> None:
684
        video = self._connector.fetch_video_async(video_url)
685
686
687
688

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

689

690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
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")


712
def load_chat_template(
713
714
715
716
    chat_template: Optional[Union[Path, str]],
    *,
    is_literal: bool = False,
) -> Optional[str]:
717
718
    if chat_template is None:
        return None
719
720
721
722
723
724
725
726

    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")

727
    try:
728
        with open(chat_template) as f:
729
            return f.read()
730
    except OSError as e:
731
732
733
        if isinstance(chat_template, Path):
            raise

734
735
736
737
738
739
        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
740

741
742
        # If opening a file fails, set chat template to be args to
        # ensure we decode so our escape are interpreted correctly
743
        return load_chat_template(chat_template, is_literal=True)
744
745


746
# TODO: Let user specify how to insert multimodal tokens into prompt
747
# (similar to chat template)
748
def _get_full_multimodal_text_prompt(placeholder_counts: dict[str, int],
749
                                     text_prompt: str) -> str:
750
    """Combine multimodal prompts for a multimodal language model."""
751

752
    # Look through the text prompt to check for missing placeholders
753
    missing_placeholders: list[str] = []
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
    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])
770
771


772
773
774
# No need to validate using Pydantic again
_TextParser = partial(cast, ChatCompletionContentPartTextParam)
_ImageParser = partial(cast, ChatCompletionContentPartImageParam)
775
_ImageEmbedsParser = partial(cast, ChatCompletionContentPartImageEmbedsParam)
776
_AudioParser = partial(cast, ChatCompletionContentPartAudioParam)
777
_InputAudioParser = partial(cast, ChatCompletionContentPartInputAudioParam)
778
_RefusalParser = partial(cast, ChatCompletionContentPartRefusalParam)
779
_VideoParser = partial(cast, ChatCompletionContentPartVideoParam)
780

781
_ContentPart: TypeAlias = Union[str, dict[str, str], InputAudio]
782

783
# Define a mapping from part types to their corresponding parsing functions.
784
MM_PARSER_MAP: dict[
785
786
787
    str,
    Callable[[ChatCompletionContentPartParam], _ContentPart],
] = {
788
789
790
791
    "text":
    lambda part: _TextParser(part).get("text", ""),
    "image_url":
    lambda part: _ImageParser(part).get("image_url", {}).get("url", ""),
792
793
    "image_embeds":
    lambda part: _ImageEmbedsParser(part).get("image_embeds", {}),
794
795
    "audio_url":
    lambda part: _AudioParser(part).get("audio_url", {}).get("url", ""),
796
797
    "input_audio":
    lambda part: _InputAudioParser(part).get("input_audio", {}),
798
799
    "refusal":
    lambda part: _RefusalParser(part).get("refusal", ""),
800
801
    "video_url":
    lambda part: _VideoParser(part).get("video_url", {}).get("url", ""),
802
803
804
805
}


def _parse_chat_message_content_mm_part(
806
        part: ChatCompletionContentPartParam) -> tuple[str, _ContentPart]:
807
    """
808
    Parses a given multi-modal content part based on its type.
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828

    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'
829
830
        # We only support 'auto', which is the default
        if part_type == "image_url" and part.get("detail", "auto") != "auto":
831
832
833
834
835
836
            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.
837
    # 'type' is required field by pydantic
838
839
840
841
842
843
844
845
846
    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", "")
847
        if part.get("input_audio") is not None:
848
            input_audio_params = cast(dict[str, str], part)
849
            return "input_audio", input_audio_params
850
851
852
853
        if part.get("video_url") is not None:
            video_params = cast(CustomChatCompletionContentSimpleVideoParam,
                                part)
            return "video_url", video_params.get("video_url", "")
854
855
856
857
858
859
860
861
862
        # 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",
863
                                       "image_embeds",
864
                                       "audio_url", "input_audio", "video_url")
865

866

867
868
869
def _parse_chat_message_content_parts(
    role: str,
    parts: Iterable[ChatCompletionContentPartParam],
870
    mm_tracker: BaseMultiModalItemTracker,
871
872
    *,
    wrap_dicts: bool,
873
) -> list[ConversationMessage]:
874
    content = list[_ContentPart]()
875

876
    mm_parser = mm_tracker.create_parser()
877
878

    for part in parts:
879
        parse_res = _parse_chat_message_content_part(
880
881
882
883
            part,
            mm_parser,
            wrap_dicts=wrap_dicts,
        )
884
885
        if parse_res:
            content.append(parse_res)
886

887
    if wrap_dicts:
888
        # Parsing wraps images and texts as interleaved dictionaries
889
        return [ConversationMessage(role=role,
890
                                    content=content)]  # type: ignore
891
    texts = cast(list[str], content)
892
893
894
895
896
897
898
899
900
    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(
901
902
903
904
    part: ChatCompletionContentPartParam,
    mm_parser: BaseMultiModalContentParser,
    *,
    wrap_dicts: bool,
905
) -> Optional[_ContentPart]:
906
907
908
909
910
911
912
913
    """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
914
        return part
915
916
917
918

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

919
    # if part_type is text/refusal/image_url/audio_url/video_url/input_audio but
920
921
922
    # content is empty, log a warning and skip
    if part_type in VALID_MESSAGE_CONTENT_MM_PART_TYPES and not content:
        logger.warning(
923
            "Skipping multimodal part (type: '%s') "
924
925
926
927
            "with empty / unparsable content.", part_type)
        return None

    if part_type in ("text", "refusal"):
928
929
930
931
932
        str_content = cast(str, content)
        if wrap_dicts:
            return {'type': 'text', 'text': str_content}
        else:
            return str_content
933
934

    if part_type == "image_url":
935
936
        str_content = cast(str, content)
        mm_parser.parse_image(str_content)
937
        return {'type': 'image'} if wrap_dicts else None
938
939
940
941
    if part_type == "image_embeds":
        content = cast(Union[str, dict[str, str]], content)
        mm_parser.parse_image_embeds(content)
        return {'type': 'image'} if wrap_dicts else None
942
    if part_type == "audio_url":
943
944
945
946
947
        str_content = cast(str, content)
        mm_parser.parse_audio(str_content)
        return {'type': 'audio'} if wrap_dicts else None

    if part_type == "input_audio":
948
        dict_content = cast(InputAudio, content)
949
        mm_parser.parse_input_audio(dict_content)
950
951
        return {'type': 'audio'} if wrap_dicts else None

952
    if part_type == "video_url":
953
954
        str_content = cast(str, content)
        mm_parser.parse_video(str_content)
955
956
        return {'type': 'video'} if wrap_dicts else None

957
    raise NotImplementedError(f"Unknown part type: {part_type}")
958
959


960
961
962
963
964
# No need to validate using Pydantic again
_AssistantParser = partial(cast, ChatCompletionAssistantMessageParam)
_ToolParser = partial(cast, ChatCompletionToolMessageParam)


965
def _parse_chat_message_content(
966
967
    message: ChatCompletionMessageParam,
    mm_tracker: BaseMultiModalItemTracker,
968
    content_format: _ChatTemplateContentFormat,
969
) -> list[ConversationMessage]:
970
971
972
973
    role = message["role"]
    content = message.get("content")

    if content is None:
974
975
976
977
978
979
        content = []
    elif isinstance(content, str):
        content = [
            ChatCompletionContentPartTextParam(type="text", text=content)
        ]
    result = _parse_chat_message_content_parts(
980
981
        role,
        content,  # type: ignore
982
        mm_tracker,
983
        wrap_dicts=(content_format == "openai"),
984
    )
985

986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
    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

1002

1003
def _postprocess_messages(messages: list[ConversationMessage]) -> None:
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
    # 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"])


1018
def parse_chat_messages(
1019
    messages: list[ChatCompletionMessageParam],
1020
    model_config: ModelConfig,
1021
    tokenizer: AnyTokenizer,
1022
    content_format: _ChatTemplateContentFormat,
1023
1024
) -> tuple[list[ConversationMessage], Optional[MultiModalDataDict]]:
    conversation: list[ConversationMessage] = []
1025
    mm_tracker = MultiModalItemTracker(model_config, tokenizer)
1026
1027

    for msg in messages:
1028
1029
1030
        sub_messages = _parse_chat_message_content(
            msg,
            mm_tracker,
1031
            content_format,
1032
        )
1033

1034
        conversation.extend(sub_messages)
1035

1036
1037
    _postprocess_messages(conversation)

1038
    return conversation, mm_tracker.all_mm_data()
1039
1040


1041
def parse_chat_messages_futures(
1042
    messages: list[ChatCompletionMessageParam],
1043
1044
    model_config: ModelConfig,
    tokenizer: AnyTokenizer,
1045
    content_format: _ChatTemplateContentFormat,
1046
1047
) -> tuple[list[ConversationMessage], Awaitable[Optional[MultiModalDataDict]]]:
    conversation: list[ConversationMessage] = []
1048
1049
1050
    mm_tracker = AsyncMultiModalItemTracker(model_config, tokenizer)

    for msg in messages:
1051
1052
1053
        sub_messages = _parse_chat_message_content(
            msg,
            mm_tracker,
1054
            content_format,
1055
        )
1056
1057
1058

        conversation.extend(sub_messages)

1059
1060
    _postprocess_messages(conversation)

1061
1062
1063
    return conversation, mm_tracker.all_mm_data()


1064
1065
def apply_hf_chat_template(
    tokenizer: Union[PreTrainedTokenizer, PreTrainedTokenizerFast],
1066
    conversation: list[ConversationMessage],
1067
1068
1069
1070
    chat_template: Optional[str],
    *,
    tokenize: bool = False,  # Different from HF's default
    **kwargs: Any,
1071
) -> str:
1072
1073
1074
1075
1076
1077
    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.")

1078
1079
1080
1081
1082
1083
    return tokenizer.apply_chat_template(
        conversation=conversation,  # type: ignore[arg-type]
        chat_template=chat_template,
        tokenize=tokenize,
        **kwargs,
    )
1084
1085


1086
1087
def apply_mistral_chat_template(
    tokenizer: MistralTokenizer,
1088
    messages: list[ChatCompletionMessageParam],
1089
    chat_template: Optional[str] = None,
1090
    **kwargs: Any,
1091
) -> list[int]:
1092
    if chat_template is not None:
1093
        logger.warning_once(
1094
            "'chat_template' cannot be overridden for mistral tokenizer.")
1095
    if "add_generation_prompt" in kwargs:
1096
        logger.warning_once(
1097
1098
1099
            "'add_generation_prompt' is not supported for mistral tokenizer, "
            "so it will be ignored.")
    if "continue_final_message" in kwargs:
1100
        logger.warning_once(
1101
1102
            "'continue_final_message' is not supported for mistral tokenizer, "
            "so it will be ignored.")
1103

1104
1105
    return tokenizer.apply_chat_template(
        messages=messages,
1106
1107
        **kwargs,
    )