chat_utils.py 56.5 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

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

14
15
16
import jinja2
import jinja2.ext
import jinja2.meta
17
import jinja2.nodes
18
19
import jinja2.parser
import jinja2.sandbox
20
import transformers.utils.chat_template_utils as hf_chat_utils
21
from openai.types.chat import (
22
23
24
25
26
27
28
29
30
31
32
    ChatCompletionAssistantMessageParam,
    ChatCompletionContentPartImageParam,
    ChatCompletionContentPartInputAudioParam,
    ChatCompletionContentPartRefusalParam,
    ChatCompletionContentPartTextParam,
    ChatCompletionMessageToolCallParam,
    ChatCompletionToolMessageParam,
)
from openai.types.chat import (
    ChatCompletionContentPartParam as OpenAIChatCompletionContentPartParam,
)
33
from openai.types.chat import (
34
35
36
    ChatCompletionMessageParam as OpenAIChatCompletionMessageParam,
)
from openai.types.chat.chat_completion_content_part_input_audio_param import InputAudio
37
from openai.types.responses import ResponseInputImageParam
38
from openai_harmony import Message as OpenAIHarmonyMessage
39
40
from PIL import Image
from pydantic import BaseModel, ConfigDict, TypeAdapter
41
42
from transformers import PreTrainedTokenizer, PreTrainedTokenizerFast, ProcessorMixin

43
# pydantic needs the TypedDict from typing_extensions
44
from typing_extensions import Required, TypedDict
45

46
from vllm.config import ModelConfig
47
from vllm.logger import init_logger
48
from vllm.model_executor.models import SupportsMultiModal
49
from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalDataDict, MultiModalUUIDDict
50
from vllm.multimodal.utils import MediaConnector
51
from vllm.transformers_utils.chat_templates import get_chat_template_fallback_path
52
from vllm.transformers_utils.processor import cached_get_processor
53
from vllm.transformers_utils.tokenizer import AnyTokenizer, MistralTokenizer
54
from vllm.utils import random_uuid
55
from vllm.utils.func_utils import supports_kw
56
57
58

logger = init_logger(__name__)

59
60
61
62
63
64
MODALITY_PLACEHOLDERS_MAP = {
    "image": "<##IMAGE##>",
    "audio": "<##AUDIO##>",
    "video": "<##VIDEO##>",
}

65

66
67
68
69
70
71
72
73
74
75
76
77
78
79
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."""


80
class ChatCompletionContentPartImageEmbedsParam(TypedDict, total=False):
81
    image_embeds: str | dict[str, str] | None
82
83
84
85
86
87
88
    """
    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."""
89
    uuid: str | None
90
91
92
93
    """
    User-provided UUID of a media. User must guarantee that it is properly
    generated and unique for different medias.
    """
94
95


96
97
98
99
100
101
102
103
104
105
106
107
108
109
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."""


110
111
112
113
class PILImage(BaseModel):
    """
    A PIL.Image.Image object.
    """
114

115
116
117
118
119
120
121
122
123
124
125
126
    image_pil: Image.Image
    model_config = ConfigDict(arbitrary_types_allowed=True)


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

    Example:
    {
        "image_pil": ImageAsset('cherry_blossom').pil_image
    }
    """
127

128
129
    image_pil: PILImage | None
    uuid: str | None
130
131
132
133
    """
    User-provided UUID of a media. User must guarantee that it is properly
    generated and unique for different medias.
    """
134
135


136
137
138
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.
139

140
141
142
143
144
    Example:
    {
        "image_url": "https://example.com/image.jpg"
    }
    """
145

146
147
    image_url: str | None
    uuid: str | None
148
149
150
151
    """
    User-provided UUID of a media. User must guarantee that it is properly
    generated and unique for different medias.
    """
152
153
154
155


class CustomChatCompletionContentSimpleAudioParam(TypedDict, total=False):
    """A simpler version of the param that only accepts a plain audio_url.
156

157
158
159
160
161
    Example:
    {
        "audio_url": "https://example.com/audio.mp3"
    }
    """
162

163
    audio_url: str | None
164
165


166
167
168
169
170
171
172
173
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"
    }
    """
174

175
176
    video_url: str | None
    uuid: str | None
177
178
179
180
    """
    User-provided UUID of a media. User must guarantee that it is properly
    generated and unique for different medias.
    """
181
182


Julien Denize's avatar
Julien Denize committed
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
class CustomThinkCompletionContentParam(TypedDict, total=False):
    """A Think Completion Content Param that accepts a plain text and a boolean.

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

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

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

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


204
205
206
207
208
209
210
211
212
213
214
215
216
217
ChatCompletionContentPartParam: TypeAlias = (
    OpenAIChatCompletionContentPartParam
    | ChatCompletionContentPartAudioParam
    | ChatCompletionContentPartInputAudioParam
    | ChatCompletionContentPartVideoParam
    | ChatCompletionContentPartRefusalParam
    | CustomChatCompletionContentPILImageParam
    | CustomChatCompletionContentSimpleImageParam
    | ChatCompletionContentPartImageEmbedsParam
    | CustomChatCompletionContentSimpleAudioParam
    | CustomChatCompletionContentSimpleVideoParam
    | str
    | CustomThinkCompletionContentParam
)
218
219
220
221


class CustomChatCompletionMessageParam(TypedDict, total=False):
    """Enables custom roles in the Chat Completion API."""
222

223
224
225
    role: Required[str]
    """The role of the message's author."""

226
    content: str | list[ChatCompletionContentPartParam]
227
228
229
230
231
232
233
234
235
    """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.
    """

236
    tool_call_id: str | None
237
238
    """Tool call that this message is responding to."""

239
    tool_calls: Iterable[ChatCompletionMessageToolCallParam] | None
240
241
    """The tool calls generated by the model, such as function calls."""

242

243
244
245
246
247
ChatCompletionMessageParam: TypeAlias = (
    OpenAIChatCompletionMessageParam
    | CustomChatCompletionMessageParam
    | OpenAIHarmonyMessage
)
248
249


250
# TODO: Make fields ReadOnly once mypy supports it
251
252
253
254
class ConversationMessage(TypedDict, total=False):
    role: Required[str]
    """The role of the message's author."""

255
    content: str | None | list[dict[str, str]]
256
257
    """The contents of the message"""

258
    tool_call_id: str | None
259
260
    """Tool call that this message is responding to."""

261
    name: str | None
262
263
    """The name of the function to call"""

264
    tool_calls: Iterable[ChatCompletionMessageToolCallParam] | None
265
    """The tool calls generated by the model, such as function calls."""
266
267


268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
# 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):
284
285
286
287
288
        return (
            _is_var_access(node.node, varname)
            and isinstance(node.arg, jinja2.nodes.Const)
            and node.arg.value == key
        )
289
290
291
292
293
294
295
296
297
298

    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,
299
    key: str | None = None,
300
301
) -> bool:
    if isinstance(node, jinja2.nodes.Filter):
302
        return node.node is not None and _is_var_or_elems_access(
303
304
            node.node, varname, key
        )
305
306
307
    if isinstance(node, jinja2.nodes.Test):
        return _is_var_or_elems_access(node.node, varname, key)

308
    if isinstance(node, jinja2.nodes.Getitem) and isinstance(
309
310
        node.arg, jinja2.nodes.Slice
    ):
311
312
        return _is_var_or_elems_access(node.node, varname, key)

313
    return _is_attr_access(node, varname, key) if key else _is_var_access(node, varname)
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


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 = [
342
        varname for _, varname in _iter_nodes_assign_var_or_elems(root, "messages")
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
    ]

    # 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


374
def _try_extract_ast(chat_template: str) -> jinja2.nodes.Template | None:
375
376
377
378
379
380
381
382
    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


383
@lru_cache(maxsize=32)
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
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"


404
def resolve_mistral_chat_template(
405
    chat_template: str | None,
406
    **kwargs: Any,
407
) -> str | None:
408
409
410
411
    if chat_template is not None or kwargs.get("chat_template_kwargs") is not None:
        raise ValueError(
            "'chat_template' or 'chat_template_kwargs' cannot be overridden "
            "for mistral tokenizer."
412
        )
413

414
415
    return None

416

417
_PROCESSOR_CHAT_TEMPLATES = dict[tuple[str, bool], str | None]()
418
419
420
421
422
423
424
425
426
"""
Used in `_try_get_processor_chat_template` to avoid calling
`cached_get_processor` again if the processor fails to be loaded.

This is needed because `lru_cache` does not cache when an exception happens.
"""


def _try_get_processor_chat_template(
427
    tokenizer: PreTrainedTokenizer | PreTrainedTokenizerFast,
428
    model_config: ModelConfig,
429
) -> str | None:
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
    cache_key = (tokenizer.name_or_path, model_config.trust_remote_code)
    if cache_key in _PROCESSOR_CHAT_TEMPLATES:
        return _PROCESSOR_CHAT_TEMPLATES[cache_key]

    try:
        processor = cached_get_processor(
            tokenizer.name_or_path,
            processor_cls=(
                PreTrainedTokenizer,
                PreTrainedTokenizerFast,
                ProcessorMixin,
            ),
            trust_remote_code=model_config.trust_remote_code,
        )
        if (
            isinstance(processor, ProcessorMixin)
            and hasattr(processor, "chat_template")
            and (chat_template := processor.chat_template) is not None
        ):
            _PROCESSOR_CHAT_TEMPLATES[cache_key] = chat_template
            return chat_template
    except Exception:
        logger.debug(
            "Failed to load AutoProcessor chat template for %s",
            tokenizer.name_or_path,
            exc_info=True,
        )

    _PROCESSOR_CHAT_TEMPLATES[cache_key] = None
    return None


462
def resolve_hf_chat_template(
463
464
465
    tokenizer: PreTrainedTokenizer | PreTrainedTokenizerFast,
    chat_template: str | None,
    tools: list[dict[str, Any]] | None,
466
467
    *,
    model_config: ModelConfig,
468
) -> str | None:
469
470
471
472
473
474
    # 1st priority: The given chat template
    if chat_template is not None:
        return chat_template

    # 2nd priority: AutoProcessor chat template, unless tool calling is enabled
    if tools is None:
475
        chat_template = _try_get_processor_chat_template(tokenizer, model_config)
476
477
        if chat_template is not None:
            return chat_template
478
479
480
481
482

    # 3rd priority: AutoTokenizer chat template
    try:
        return tokenizer.get_chat_template(chat_template, tools=tools)
    except Exception:
483
484
485
486
487
        logger.debug(
            "Failed to load AutoTokenizer chat template for %s",
            tokenizer.name_or_path,
            exc_info=True,
        )
488

489
490
491
492
493
494
    # 4th priority: Predefined fallbacks
    path = get_chat_template_fallback_path(
        model_type=model_config.hf_config.model_type,
        tokenizer_name_or_path=model_config.tokenizer,
    )
    if path is not None:
495
        logger.info_once(
496
497
498
499
            "Loading chat template fallback for %s as there isn't one "
            "defined on HF Hub.",
            tokenizer.name_or_path,
        )
500
501
        chat_template = load_chat_template(path)
    else:
502
        logger.debug_once(
503
504
            "There is no chat template fallback for %s", tokenizer.name_or_path
        )
505
506

    return chat_template
507
508


509
def _resolve_chat_template_content_format(
510
511
    chat_template: str | None,
    tools: list[dict[str, Any]] | None,
512
    tokenizer: AnyTokenizer,
513
514
    *,
    model_config: ModelConfig,
515
516
) -> _ChatTemplateContentFormat:
    if isinstance(tokenizer, (PreTrainedTokenizer, PreTrainedTokenizerFast)):
517
        hf_chat_template = resolve_hf_chat_template(
518
519
520
            tokenizer,
            chat_template=chat_template,
            tools=tools,
521
            model_config=model_config,
522
        )
523
    else:
524
525
        hf_chat_template = None

526
527
528
529
530
    jinja_text = (
        hf_chat_template
        if isinstance(hf_chat_template, str)
        else load_chat_template(chat_template, is_literal=True)
    )
531

532
533
534
535
536
    detected_format = (
        "string"
        if jinja_text is None
        else _detect_content_format(jinja_text, default="string")
    )
537

538
    return detected_format
539
540
541


@lru_cache
542
def _log_chat_template_content_format(
543
    chat_template: str | None,
544
    given_format: ChatTemplateContentFormatOption,
545
546
    detected_format: ChatTemplateContentFormatOption,
):
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
    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,
        )

564
565

def resolve_chat_template_content_format(
566
567
    chat_template: str | None,
    tools: list[dict[str, Any]] | None,
568
569
    given_format: ChatTemplateContentFormatOption,
    tokenizer: AnyTokenizer,
570
571
    *,
    model_config: ModelConfig,
572
) -> _ChatTemplateContentFormat:
573
574
575
    if given_format != "auto":
        return given_format

576
577
578
579
    detected_format = _resolve_chat_template_content_format(
        chat_template,
        tools,
        tokenizer,
580
        model_config=model_config,
581
582
583
584
585
586
587
588
    )

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

589
    return detected_format
590

591

592
ModalityStr = Literal["image", "audio", "video", "image_embeds"]
593
594
595
596
_T = TypeVar("_T")


class BaseMultiModalItemTracker(ABC, Generic[_T]):
597
598
599
600
601
602
603
    """
    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):
604
605
        super().__init__()

606
607
        self._model_config = model_config
        self._tokenizer = tokenizer
608

609
610
        self._items_by_modality = defaultdict[str, list[_T | None]](list)
        self._uuids_by_modality = defaultdict[str, list[str | None]](list)
611

612
613
614
615
    @property
    def model_config(self) -> ModelConfig:
        return self._model_config

616
    @cached_property
617
    def model_cls(self) -> type[SupportsMultiModal]:
618
        from vllm.model_executor.model_loader import get_model_cls
619

620
621
        model_cls = get_model_cls(self.model_config)
        return cast(type[SupportsMultiModal], model_cls)
622

623
624
625
626
    @property
    def allowed_local_media_path(self):
        return self._model_config.allowed_local_media_path

627
628
629
630
    @property
    def allowed_media_domains(self):
        return self._model_config.allowed_media_domains

631
632
633
634
    @property
    def mm_registry(self):
        return MULTIMODAL_REGISTRY

635
636
637
638
    @cached_property
    def mm_processor(self):
        return self.mm_registry.create_processor(self.model_config)

639
    def add(
640
641
        self,
        modality: ModalityStr,
642
643
644
        item: _T | None,
        uuid: str | None = None,
    ) -> str | None:
645
646
647
        """
        Add a multi-modal item to the current prompt and returns the
        placeholder string to use, if any.
648
649

        An optional uuid can be added which serves as a unique identifier of the
650
        media.
651
        """
652
        input_modality = modality.replace("_embeds", "")
653
        num_items = len(self._items_by_modality[modality]) + 1
654

655
        self.mm_processor.validate_num_items(input_modality, num_items)
656

657
        self._items_by_modality[modality].append(item)
658
        self._uuids_by_modality[modality].append(uuid)
659

660
        return self.model_cls.get_placeholder_str(modality, num_items)
661

662
    def all_mm_uuids(self) -> MultiModalUUIDDict | None:
663
664
665
666
667
        if not self._items_by_modality:
            return None
        mm_uuids = {}
        uuids_by_modality = dict(self._uuids_by_modality)
        if "image" in uuids_by_modality and "image_embeds" in uuids_by_modality:
668
            raise ValueError("Mixing raw image and embedding inputs is not allowed")
669
670
671
672

        if "image_embeds" in uuids_by_modality:
            image_embeds_uuids = uuids_by_modality["image_embeds"]
            if len(image_embeds_uuids) > 1:
673
                raise ValueError("Only one message can have {'type': 'image_embeds'}")
674
675
676
677
678
679
680
681
682
            mm_uuids["image"] = uuids_by_modality["image_embeds"]
        if "image" in uuids_by_modality:
            mm_uuids["image"] = uuids_by_modality["image"]  # UUIDs of images
        if "audio" in uuids_by_modality:
            mm_uuids["audio"] = uuids_by_modality["audio"]  # UUIDs of audios
        if "video" in uuids_by_modality:
            mm_uuids["video"] = uuids_by_modality["video"]  # UUIDs of videos
        return mm_uuids

683
684
685
686
687
    @abstractmethod
    def create_parser(self) -> "BaseMultiModalContentParser":
        raise NotImplementedError


688
class MultiModalItemTracker(BaseMultiModalItemTracker[object]):
689
    def all_mm_data(self) -> MultiModalDataDict | None:
690
691
692
693
694
        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:
695
            raise ValueError("Mixing raw image and embedding inputs is not allowed")
696
697
698
699

        if "image_embeds" in items_by_modality:
            image_embeds_lst = items_by_modality["image_embeds"]
            if len(image_embeds_lst) > 1:
700
                raise ValueError("Only one message can have {'type': 'image_embeds'}")
701
            mm_inputs["image"] = image_embeds_lst[0]
702
        if "image" in items_by_modality:
703
            mm_inputs["image"] = items_by_modality["image"]  # A list of images
704
        if "audio" in items_by_modality:
705
            mm_inputs["audio"] = items_by_modality["audio"]  # A list of audios
706
        if "video" in items_by_modality:
707
            mm_inputs["video"] = items_by_modality["video"]  # A list of videos
708
        return mm_inputs
709
710
711
712
713

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


714
class AsyncMultiModalItemTracker(BaseMultiModalItemTracker[Awaitable[object]]):
715
    async def all_mm_data(self) -> MultiModalDataDict | None:
716
717
718
        if not self._items_by_modality:
            return None
        mm_inputs = {}
719
720
721
722
723
724
725
726
727
        items_by_modality = {}
        for modality, items in self._items_by_modality.items():
            coros = []
            for item in items:
                if item is not None:
                    coros.append(item)
                else:
                    coros.append(asyncio.sleep(0))
            items_by_modality[modality] = await asyncio.gather(*coros)
728

729
        if "image" in items_by_modality and "image_embeds" in items_by_modality:
730
            raise ValueError("Mixing raw image and embedding inputs is not allowed")
731
732
733
734

        if "image_embeds" in items_by_modality:
            image_embeds_lst = items_by_modality["image_embeds"]
            if len(image_embeds_lst) > 1:
735
                raise ValueError("Only one message can have {'type': 'image_embeds'}")
736
            mm_inputs["image"] = image_embeds_lst[0]
737
        if "image" in items_by_modality:
738
            mm_inputs["image"] = items_by_modality["image"]  # A list of images
739
        if "audio" in items_by_modality:
740
            mm_inputs["audio"] = items_by_modality["audio"]  # A list of audios
741
        if "video" in items_by_modality:
742
            mm_inputs["video"] = items_by_modality["video"]  # A list of videos
743
        return mm_inputs
744
745
746
747
748
749
750
751
752

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


class BaseMultiModalContentParser(ABC):
    def __init__(self) -> None:
        super().__init__()

753
        # stores model placeholders list with corresponding
754
755
756
757
758
759
760
        # general MM placeholder:
        # {
        #   "<##IMAGE##>": ["<image>", "<image>", "<image>"],
        #   "<##AUDIO##>": ["<audio>", "<audio>"]
        # }
        self._placeholder_storage: dict[str, list] = defaultdict(list)

761
    def _add_placeholder(self, modality: ModalityStr, placeholder: str | None):
762
        mod_placeholder = MODALITY_PLACEHOLDERS_MAP[modality]
763
        if placeholder:
764
            self._placeholder_storage[mod_placeholder].append(placeholder)
765

766
767
    def mm_placeholder_storage(self) -> dict[str, list]:
        return dict(self._placeholder_storage)
768
769

    @abstractmethod
770
    def parse_image(self, image_url: str | None, uuid: str | None = None) -> None:
771
772
        raise NotImplementedError

773
    @abstractmethod
774
    def parse_image_embeds(
775
        self,
776
777
        image_embeds: str | dict[str, str] | None,
        uuid: str | None = None,
778
    ) -> None:
779
780
        raise NotImplementedError

781
    @abstractmethod
782
    def parse_image_pil(
783
        self, image_pil: Image.Image | None, uuid: str | None = None
784
    ) -> None:
785
786
        raise NotImplementedError

787
    @abstractmethod
788
    def parse_audio(self, audio_url: str | None, uuid: str | None = None) -> None:
789
790
        raise NotImplementedError

791
    @abstractmethod
792
    def parse_input_audio(
793
        self, input_audio: InputAudio | None, uuid: str | None = None
794
    ) -> None:
795
796
        raise NotImplementedError

797
    @abstractmethod
798
    def parse_video(self, video_url: str | None, uuid: str | None = None) -> None:
799
800
        raise NotImplementedError

801
802
803
804
805
806

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

        self._tracker = tracker
807
808
        multimodal_config = self._tracker.model_config.multimodal_config
        media_io_kwargs = getattr(multimodal_config, "media_io_kwargs", None)
809
        self._connector = MediaConnector(
810
            media_io_kwargs=media_io_kwargs,
811
            allowed_local_media_path=tracker.allowed_local_media_path,
812
            allowed_media_domains=tracker.allowed_media_domains,
813
814
        )

815
816
817
818
    @property
    def model_config(self) -> ModelConfig:
        return self._tracker.model_config

819
    def parse_image(self, image_url: str | None, uuid: str | None = None) -> None:
820
        image = self._connector.fetch_image(image_url) if image_url else None
821

822
        placeholder = self._tracker.add("image", image, uuid)
823
        self._add_placeholder("image", placeholder)
824

825
    def parse_image_embeds(
826
        self,
827
828
        image_embeds: str | dict[str, str] | None,
        uuid: str | None = None,
829
    ) -> None:
830
831
832
833
834
835
        mm_config = self.model_config.get_multimodal_config()
        if not mm_config.enable_mm_embeds:
            raise ValueError(
                "You must set `--enable-mm-embeds` to input `image_embeds`"
            )

836
837
838
839
840
        if isinstance(image_embeds, dict):
            embeds = {
                k: self._connector.fetch_image_embedding(v)
                for k, v in image_embeds.items()
            }
841
            placeholder = self._tracker.add("image_embeds", embeds, uuid)
842
843
844

        if isinstance(image_embeds, str):
            embedding = self._connector.fetch_image_embedding(image_embeds)
845
            placeholder = self._tracker.add("image_embeds", embedding, uuid)
846

847
848
849
        if image_embeds is None:
            placeholder = self._tracker.add("image_embeds", None, uuid)

850
        self._add_placeholder("image", placeholder)
851

852
    def parse_image_pil(
853
        self, image_pil: Image.Image | None, uuid: str | None = None
854
855
    ) -> None:
        placeholder = self._tracker.add("image", image_pil, uuid)
856
        self._add_placeholder("image", placeholder)
857

858
    def parse_audio(self, audio_url: str | None, uuid: str | None = None) -> None:
859
        audio = self._connector.fetch_audio(audio_url) if audio_url else None
860

861
        placeholder = self._tracker.add("audio", audio, uuid)
862
        self._add_placeholder("audio", placeholder)
863

864
    def parse_input_audio(
865
        self, input_audio: InputAudio | None, uuid: str | None = None
866
    ) -> None:
867
868
869
870
871
872
873
874
875
876
        if input_audio:
            audio_data = input_audio.get("data", "")
            audio_format = input_audio.get("format", "")
            if audio_data:
                audio_url = f"data:audio/{audio_format};base64,{audio_data}"
            else:
                # If a UUID is provided, audio data may be empty.
                audio_url = None
        else:
            audio_url = None
877

878
        return self.parse_audio(audio_url, uuid)
879

880
    def parse_video(self, video_url: str | None, uuid: str | None = None) -> None:
881
        video = self._connector.fetch_video(video_url=video_url) if video_url else None
882

883
        placeholder = self._tracker.add("video", video, uuid)
884
        self._add_placeholder("video", placeholder)
885

886
887
888
889
890
891

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

        self._tracker = tracker
892
893
        multimodal_config = self._tracker.model_config.multimodal_config
        media_io_kwargs = getattr(multimodal_config, "media_io_kwargs", None)
894
        self._connector = MediaConnector(
895
            media_io_kwargs=media_io_kwargs,
896
            allowed_local_media_path=tracker.allowed_local_media_path,
897
            allowed_media_domains=tracker.allowed_media_domains,
898
        )
899

900
901
902
903
    @property
    def model_config(self) -> ModelConfig:
        return self._tracker.model_config

904
    def parse_image(self, image_url: str | None, uuid: str | None = None) -> None:
905
        image_coro = self._connector.fetch_image_async(image_url) if image_url else None
906

907
        placeholder = self._tracker.add("image", image_coro, uuid)
908
        self._add_placeholder("image", placeholder)
909

910
    def parse_image_embeds(
911
        self,
912
913
        image_embeds: str | dict[str, str] | None,
        uuid: str | None = None,
914
    ) -> None:
915
916
917
918
919
920
        mm_config = self.model_config.get_multimodal_config()
        if not mm_config.enable_mm_embeds:
            raise ValueError(
                "You must set `--enable-mm-embeds` to input `image_embeds`"
            )

921
        future: asyncio.Future[str | dict[str, str] | None] = asyncio.Future()
922
923
924
925
926
927
928
929
930

        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):
931
            embedding = self._connector.fetch_image_embedding(image_embeds)
932
933
            future.set_result(embedding)

934
935
936
        if image_embeds is None:
            future.set_result(None)

937
        placeholder = self._tracker.add("image_embeds", future, uuid)
938
        self._add_placeholder("image", placeholder)
939

940
    def parse_image_pil(
941
        self, image_pil: Image.Image | None, uuid: str | None = None
942
    ) -> None:
943
        future: asyncio.Future[Image.Image | None] = asyncio.Future()
944
945
946
947
        if image_pil:
            future.set_result(image_pil)
        else:
            future.set_result(None)
948

949
        placeholder = self._tracker.add("image", future, uuid)
950
        self._add_placeholder("image", placeholder)
951

952
    def parse_audio(self, audio_url: str | None, uuid: str | None = None) -> None:
953
        audio_coro = self._connector.fetch_audio_async(audio_url) if audio_url else None
954

955
        placeholder = self._tracker.add("audio", audio_coro, uuid)
956
        self._add_placeholder("audio", placeholder)
957

958
    def parse_input_audio(
959
        self, input_audio: InputAudio | None, uuid: str | None = None
960
    ) -> None:
961
962
963
964
965
966
967
968
969
970
        if input_audio:
            audio_data = input_audio.get("data", "")
            audio_format = input_audio.get("format", "")
            if audio_data:
                audio_url = f"data:audio/{audio_format};base64,{audio_data}"
            else:
                # If a UUID is provided, audio data may be empty.
                audio_url = None
        else:
            audio_url = None
971

972
        return self.parse_audio(audio_url, uuid)
973

974
    def parse_video(self, video_url: str | None, uuid: str | None = None) -> None:
975
976
977
978
979
        video = (
            self._connector.fetch_video_async(video_url=video_url)
            if video_url
            else None
        )
980

981
        placeholder = self._tracker.add("video", video, uuid)
982
        self._add_placeholder("video", placeholder)
983

984

985
def validate_chat_template(chat_template: Path | str | None):
986
987
988
989
990
    """Raises if the provided chat template appears invalid."""
    if chat_template is None:
        return

    elif isinstance(chat_template, Path) and not chat_template.exists():
991
        raise FileNotFoundError("the supplied chat template path doesn't exist")
992
993
994

    elif isinstance(chat_template, str):
        JINJA_CHARS = "{}\n"
995
996
997
998
        if (
            not any(c in chat_template for c in JINJA_CHARS)
            and not Path(chat_template).exists()
        ):
999
1000
            raise ValueError(
                f"The supplied chat template string ({chat_template}) "
1001
1002
                f"appears path-like, but doesn't exist!"
            )
1003
1004

    else:
1005
        raise TypeError(f"{type(chat_template)} is not a valid chat template type")
1006
1007


1008
def _load_chat_template(
1009
    chat_template: Path | str | None,
1010
1011
    *,
    is_literal: bool = False,
1012
) -> str | None:
1013
1014
    if chat_template is None:
        return None
1015
1016
1017

    if is_literal:
        if isinstance(chat_template, Path):
1018
1019
1020
            raise TypeError(
                "chat_template is expected to be read directly from its value"
            )
1021

1022
        return chat_template
1023

1024
    try:
1025
        with open(chat_template) as f:
1026
            return f.read()
1027
    except OSError as e:
1028
1029
1030
        if isinstance(chat_template, Path):
            raise

1031
1032
        JINJA_CHARS = "{}\n"
        if not any(c in chat_template for c in JINJA_CHARS):
1033
1034
1035
1036
1037
            msg = (
                f"The supplied chat template ({chat_template}) "
                f"looks like a file path, but it failed to be "
                f"opened. Reason: {e}"
            )
1038
            raise ValueError(msg) from e
1039

1040
1041
        # If opening a file fails, set chat template to be args to
        # ensure we decode so our escape are interpreted correctly
1042
1043
1044
1045
1046
1047
1048
        return _load_chat_template(chat_template, is_literal=True)


_cached_load_chat_template = lru_cache(_load_chat_template)


def load_chat_template(
1049
    chat_template: Path | str | None,
1050
1051
    *,
    is_literal: bool = False,
1052
) -> str | None:
1053
    return _cached_load_chat_template(chat_template, is_literal=is_literal)
1054
1055


1056
1057
1058
def _get_interleaved_text_prompt(
    placeholder_storage: dict[str, list], texts: list[str]
) -> str:
1059
1060
1061
1062
1063
1064
1065
    for idx, elem in enumerate(texts):
        if elem in placeholder_storage:
            texts[idx] = placeholder_storage[elem].pop(0)

    return "\n".join(texts)


1066
# TODO: Let user specify how to insert multimodal tokens into prompt
1067
# (similar to chat template)
1068
1069
1070
1071
1072
def _get_full_multimodal_text_prompt(
    placeholder_storage: dict[str, list],
    texts: list[str],
    interleave_strings: bool,
) -> str:
1073
    """Combine multimodal prompts for a multimodal language model."""
1074

1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
    # flatten storage to make it looks like
    # {
    #   "<|image|>": 2,
    #   "<|audio|>": 1
    # }
    placeholder_counts = Counter(
        [v for elem in placeholder_storage.values() for v in elem]
    )

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

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

1092
    # Look through the text prompt to check for missing placeholders
1093
    missing_placeholders: list[str] = []
1094
1095
1096
1097
1098
    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:
1099
1100
1101
1102
            logger.error(
                "Placeholder count is negative! "
                "Ensure that the 'interleave_strings' flag is disabled "
                "(current value: %s) "
1103
1104
                "when manually placing image placeholders.",
                interleave_strings,
1105
1106
            )
            logger.debug("Input prompt: %s", text_prompt)
1107
1108
            raise ValueError(
                f"Found more '{placeholder}' placeholders in input prompt than "
1109
1110
                "actual multimodal data items."
            )
1111

1112
        missing_placeholders.extend([placeholder] * placeholder_counts[placeholder])
1113

1114
1115
    # NOTE: Default behaviour: we always add missing placeholders
    # at the front of the prompt, if interleave_strings=False
1116
    return "\n".join(missing_placeholders + [text_prompt])
1117
1118


1119
1120
# No need to validate using Pydantic again
_TextParser = partial(cast, ChatCompletionContentPartTextParam)
1121
_ImageEmbedsParser = partial(cast, ChatCompletionContentPartImageEmbedsParam)
1122
_InputAudioParser = partial(cast, ChatCompletionContentPartInputAudioParam)
1123
_RefusalParser = partial(cast, ChatCompletionContentPartRefusalParam)
1124
_PILImageParser = partial(cast, CustomChatCompletionContentPILImageParam)
Julien Denize's avatar
Julien Denize committed
1125
_ThinkParser = partial(cast, CustomThinkCompletionContentParam)
1126
1127
1128
1129
# Need to validate url objects
_ImageParser = TypeAdapter(ChatCompletionContentPartImageParam).validate_python
_AudioParser = TypeAdapter(ChatCompletionContentPartAudioParam).validate_python
_VideoParser = TypeAdapter(ChatCompletionContentPartVideoParam).validate_python
1130

1131
_ResponsesInputImageParser = TypeAdapter(ResponseInputImageParam).validate_python
1132
_ContentPart: TypeAlias = str | dict[str, str] | InputAudio | PILImage
1133

1134
# Define a mapping from part types to their corresponding parsing functions.
1135
MM_PARSER_MAP: dict[
1136
1137
1138
    str,
    Callable[[ChatCompletionContentPartParam], _ContentPart],
] = {
1139
1140
1141
    "text": lambda part: _TextParser(part).get("text", None),
    "thinking": lambda part: _ThinkParser(part).get("thinking", None),
    "input_text": lambda part: _TextParser(part).get("text", None),
1142
1143
1144
    "input_image": lambda part: _ResponsesInputImageParser(part).get("image_url", None),
    "image_url": lambda part: _ImageParser(part).get("image_url", {}).get("url", None),
    "image_embeds": lambda part: _ImageEmbedsParser(part).get("image_embeds", None),
1145
    "image_pil": lambda part: _PILImageParser(part).get("image_pil", None),
1146
1147
    "audio_url": lambda part: _AudioParser(part).get("audio_url", {}).get("url", None),
    "input_audio": lambda part: _InputAudioParser(part).get("input_audio", None),
1148
    "refusal": lambda part: _RefusalParser(part).get("refusal", None),
1149
    "video_url": lambda part: _VideoParser(part).get("video_url", {}).get("url", None),
1150
1151
1152
1153
}


def _parse_chat_message_content_mm_part(
1154
1155
    part: ChatCompletionContentPartParam,
) -> tuple[str, _ContentPart]:
1156
    """
1157
    Parses a given multi-modal content part based on its type.
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170

    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(
1171
1172
        part, dict
    )  # This is needed to avoid mypy errors: part.get() from str
1173
    part_type = part.get("type", None)
1174
    uuid = part.get("uuid", None)
1175

1176
    if isinstance(part_type, str) and part_type in MM_PARSER_MAP and uuid is None:  # noqa: E501
1177
1178
1179
        content = MM_PARSER_MAP[part_type](part)

        # Special case for 'image_url.detail'
1180
1181
        # We only support 'auto', which is the default
        if part_type == "image_url" and part.get("detail", "auto") != "auto":
1182
            logger.warning(
1183
                "'image_url.detail' is currently not supported and will be ignored."
1184
            )
1185
1186
1187
1188

        return part_type, content

    # Handle missing 'type' but provided direct URL fields.
1189
    # 'type' is required field by pydantic
1190
1191
    if part_type is None or uuid is not None:
        if "image_url" in part:
1192
            image_params = cast(CustomChatCompletionContentSimpleImageParam, part)
1193
1194
1195
1196
1197
1198
1199
1200
            image_url = image_params.get("image_url", None)
            if isinstance(image_url, dict):
                # Can potentially happen if user provides a uuid
                # with url as a dict of {"url": url}
                image_url = image_url.get("url", None)
            return "image_url", image_url
        if "image_pil" in part:
            # "image_pil" could be None if UUID is provided.
1201
            image_params = cast(  # type: ignore
1202
1203
1204
1205
1206
1207
                CustomChatCompletionContentPILImageParam, part
            )
            image_pil = image_params.get("image_pil", None)
            return "image_pil", image_pil
        if "image_embeds" in part:
            # "image_embeds" could be None if UUID is provided.
1208
            image_params = cast(  # type: ignore
1209
1210
1211
1212
1213
                ChatCompletionContentPartImageEmbedsParam, part
            )
            image_embeds = image_params.get("image_embeds", None)
            return "image_embeds", image_embeds
        if "audio_url" in part:
1214
            audio_params = cast(CustomChatCompletionContentSimpleAudioParam, part)
1215
1216
1217
1218
1219
1220
            audio_url = audio_params.get("audio_url", None)
            if isinstance(audio_url, dict):
                # Can potentially happen if user provides a uuid
                # with url as a dict of {"url": url}
                audio_url = audio_url.get("url", None)
            return "audio_url", audio_url
1221
        if part.get("input_audio") is not None:
1222
            input_audio_params = cast(dict[str, str], part)
1223
            return "input_audio", input_audio_params
1224
        if "video_url" in part:
1225
            video_params = cast(CustomChatCompletionContentSimpleVideoParam, part)
1226
1227
1228
1229
1230
1231
            video_url = video_params.get("video_url", None)
            if isinstance(video_url, dict):
                # Can potentially happen if user provides a uuid
                # with url as a dict of {"url": url}
                video_url = video_url.get("url", None)
            return "video_url", video_url
1232
1233
1234
1235
1236
1237
1238
1239
        # 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"


1240
PART_TYPES_TO_SKIP_NONE_CONTENT = (
1241
1242
1243
    "text",
    "refusal",
)
1244

1245

1246
1247
1248
def _parse_chat_message_content_parts(
    role: str,
    parts: Iterable[ChatCompletionContentPartParam],
1249
    mm_tracker: BaseMultiModalItemTracker,
1250
1251
    *,
    wrap_dicts: bool,
1252
    interleave_strings: bool,
1253
) -> list[ConversationMessage]:
1254
    content = list[_ContentPart]()
1255

1256
    mm_parser = mm_tracker.create_parser()
1257
1258

    for part in parts:
1259
        parse_res = _parse_chat_message_content_part(
1260
1261
1262
            part,
            mm_parser,
            wrap_dicts=wrap_dicts,
1263
            interleave_strings=interleave_strings,
1264
        )
1265
1266
        if parse_res:
            content.append(parse_res)
1267

1268
    if wrap_dicts:
1269
        # Parsing wraps images and texts as interleaved dictionaries
1270
        return [ConversationMessage(role=role, content=content)]  # type: ignore
1271
    texts = cast(list[str], content)
1272
1273
    mm_placeholder_storage = mm_parser.mm_placeholder_storage()
    if mm_placeholder_storage:
1274
1275
1276
        text_prompt = _get_full_multimodal_text_prompt(
            mm_placeholder_storage, texts, interleave_strings
        )
1277
1278
1279
    else:
        text_prompt = "\n".join(texts)

1280
1281
1282
1283
    return [ConversationMessage(role=role, content=text_prompt)]


def _parse_chat_message_content_part(
1284
1285
1286
1287
    part: ChatCompletionContentPartParam,
    mm_parser: BaseMultiModalContentParser,
    *,
    wrap_dicts: bool,
1288
    interleave_strings: bool,
1289
) -> _ContentPart | None:
1290
1291
1292
1293
1294
1295
1296
1297
    """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
1298
        return part
1299
1300
    # Handle structured dictionary parts
    part_type, content = _parse_chat_message_content_mm_part(part)
1301
    # if part_type is text/refusal/image_url/audio_url/video_url/input_audio but
1302
    # content is None, log a warning and skip
1303
    if part_type in PART_TYPES_TO_SKIP_NONE_CONTENT and content is None:
1304
        logger.warning(
1305
            "Skipping multimodal part '%s' (type: '%s') "
1306
1307
1308
1309
            "with empty / unparsable content.",
            part,
            part_type,
        )
1310
1311
        return None

Julien Denize's avatar
Julien Denize committed
1312
    if part_type in ("text", "input_text", "refusal", "thinking"):
1313
1314
        str_content = cast(str, content)
        if wrap_dicts:
1315
            return {"type": "text", "text": str_content}
1316
1317
        else:
            return str_content
1318

1319
1320
1321
1322
1323
1324
    # For media items, if a user has provided one, use it. Otherwise, insert
    # a placeholder empty uuid.
    uuid = part.get("uuid", None)
    if uuid is not None:
        uuid = str(uuid)

1325
    modality = None
1326
    if part_type == "image_pil":
1327
        image_content = cast(Image.Image, content) if content is not None else None
1328
        mm_parser.parse_image_pil(image_content, uuid)
1329
        modality = "image"
1330
    elif part_type in ("image_url", "input_image"):
1331
        str_content = cast(str, content)
1332
        mm_parser.parse_image(str_content, uuid)
1333
1334
        modality = "image"
    elif part_type == "image_embeds":
1335
        content = cast(str | dict[str, str], content) if content is not None else None
1336
        mm_parser.parse_image_embeds(content, uuid)
1337
1338
        modality = "image"
    elif part_type == "audio_url":
1339
        str_content = cast(str, content)
1340
        mm_parser.parse_audio(str_content, uuid)
1341
1342
        modality = "audio"
    elif part_type == "input_audio":
1343
        dict_content = cast(InputAudio, content)
1344
        mm_parser.parse_input_audio(dict_content, uuid)
1345
1346
        modality = "audio"
    elif part_type == "video_url":
1347
        str_content = cast(str, content)
1348
        mm_parser.parse_video(str_content, uuid)
1349
1350
1351
        modality = "video"
    else:
        raise NotImplementedError(f"Unknown part type: {part_type}")
1352

1353
1354
1355
    return (
        {"type": modality}
        if wrap_dicts
1356
        else (MODALITY_PLACEHOLDERS_MAP[modality] if interleave_strings else None)
1357
    )
1358
1359


1360
1361
1362
1363
1364
# No need to validate using Pydantic again
_AssistantParser = partial(cast, ChatCompletionAssistantMessageParam)
_ToolParser = partial(cast, ChatCompletionToolMessageParam)


1365
def _parse_chat_message_content(
1366
1367
    message: ChatCompletionMessageParam,
    mm_tracker: BaseMultiModalItemTracker,
1368
    content_format: _ChatTemplateContentFormat,
1369
    interleave_strings: bool,
1370
) -> list[ConversationMessage]:
1371
1372
1373
1374
    role = message["role"]
    content = message.get("content")

    if content is None:
1375
1376
        content = []
    elif isinstance(content, str):
1377
        content = [ChatCompletionContentPartTextParam(type="text", text=content)]
1378
    result = _parse_chat_message_content_parts(
1379
1380
        role,
        content,  # type: ignore
1381
        mm_tracker,
1382
        wrap_dicts=(content_format == "openai"),
1383
        interleave_strings=interleave_strings,
1384
    )
1385

1386
    for result_msg in result:
1387
        if role == "assistant":
1388
1389
            parsed_msg = _AssistantParser(message)

1390
1391
1392
            # The 'tool_calls' is not None check ensures compatibility.
            # It's needed only if downstream code doesn't strictly
            # follow the OpenAI spec.
1393
            if "tool_calls" in parsed_msg and parsed_msg["tool_calls"] is not None:
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
                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

1405

1406
def _postprocess_messages(messages: list[ConversationMessage]) -> None:
1407
1408
1409
1410
1411
1412
    # 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:
1413
1414
1415
1416
1417
        if (
            message["role"] == "assistant"
            and "tool_calls" in message
            and isinstance(message["tool_calls"], list)
        ):
1418
            for item in message["tool_calls"]:
1419
1420
1421
1422
1423
                # if arguments is None or empty string, set to {}
                if content := item["function"].get("arguments"):
                    item["function"]["arguments"] = json.loads(content)
                else:
                    item["function"]["arguments"] = {}
1424
1425


1426
def parse_chat_messages(
1427
    messages: list[ChatCompletionMessageParam],
1428
    model_config: ModelConfig,
1429
    tokenizer: AnyTokenizer,
1430
    content_format: _ChatTemplateContentFormat,
1431
1432
) -> tuple[
    list[ConversationMessage],
1433
1434
    MultiModalDataDict | None,
    MultiModalUUIDDict | None,
1435
]:
1436
    conversation: list[ConversationMessage] = []
1437
    mm_tracker = MultiModalItemTracker(model_config, tokenizer)
1438
1439

    for msg in messages:
1440
1441
1442
        sub_messages = _parse_chat_message_content(
            msg,
            mm_tracker,
1443
            content_format,
1444
1445
1446
1447
            interleave_strings=(
                content_format == "string"
                and model_config.multimodal_config is not None
                and model_config.multimodal_config.interleave_mm_strings
1448
            ),
1449
        )
1450

1451
        conversation.extend(sub_messages)
1452

1453
1454
    _postprocess_messages(conversation)

1455
    return conversation, mm_tracker.all_mm_data(), mm_tracker.all_mm_uuids()
1456
1457


1458
def parse_chat_messages_futures(
1459
    messages: list[ChatCompletionMessageParam],
1460
1461
    model_config: ModelConfig,
    tokenizer: AnyTokenizer,
1462
    content_format: _ChatTemplateContentFormat,
1463
1464
) -> tuple[
    list[ConversationMessage],
1465
1466
    Awaitable[MultiModalDataDict | None],
    MultiModalUUIDDict | None,
1467
]:
1468
    conversation: list[ConversationMessage] = []
1469
1470
1471
    mm_tracker = AsyncMultiModalItemTracker(model_config, tokenizer)

    for msg in messages:
1472
1473
1474
        sub_messages = _parse_chat_message_content(
            msg,
            mm_tracker,
1475
            content_format,
1476
1477
1478
1479
            interleave_strings=(
                content_format == "string"
                and model_config.multimodal_config is not None
                and model_config.multimodal_config.interleave_mm_strings
1480
            ),
1481
        )
1482
1483
1484

        conversation.extend(sub_messages)

1485
1486
    _postprocess_messages(conversation)

1487
    return conversation, mm_tracker.all_mm_data(), mm_tracker.all_mm_uuids()
1488
1489


1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
# adapted from https://github.com/huggingface/transformers/blob/v4.56.2/src/transformers/utils/chat_template_utils.py#L398-L412
# only preserve the parse function used to resolve chat template kwargs
class AssistantTracker(jinja2.ext.Extension):
    tags = {"generation"}

    def parse(self, parser: jinja2.parser.Parser) -> jinja2.nodes.CallBlock:
        lineno = next(parser.stream).lineno
        body = parser.parse_statements(["name:endgeneration"], drop_needle=True)
        call = self.call_method("_generation_support")
        call_block = jinja2.nodes.CallBlock(call, [], [], body)
        return call_block.set_lineno(lineno)


1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
def _resolve_chat_template_kwargs(
    chat_template: str,
):
    env = jinja2.sandbox.ImmutableSandboxedEnvironment(
        trim_blocks=True,
        lstrip_blocks=True,
        extensions=[AssistantTracker, jinja2.ext.loopcontrols],
    )
    parsed_content = env.parse(chat_template)
    template_vars = jinja2.meta.find_undeclared_variables(parsed_content)
    return template_vars


_cached_resolve_chat_template_kwargs = lru_cache(_resolve_chat_template_kwargs)


1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
@lru_cache
def _get_hf_base_chat_template_params() -> frozenset[str]:
    # Get standard parameters from HuggingFace's base tokenizer class.
    # This dynamically extracts parameters from PreTrainedTokenizer's
    # apply_chat_template method, ensuring compatibility with tokenizers
    # that use **kwargs to receive standard parameters.

    # Read signature from HF's base class - the single source of truth
    base_sig = inspect.signature(PreTrainedTokenizer.apply_chat_template)
    # Exclude VAR_KEYWORD (**kwargs) and VAR_POSITIONAL (*args) placeholders
    return frozenset(
        p.name
        for p in base_sig.parameters.values()
        if p.kind
        not in (inspect.Parameter.VAR_KEYWORD, inspect.Parameter.VAR_POSITIONAL)
    )


1537
def resolve_chat_template_kwargs(
1538
    tokenizer: PreTrainedTokenizer | PreTrainedTokenizerFast,
1539
1540
    chat_template: str,
    chat_template_kwargs: dict[str, Any],
1541
    raise_on_unexpected: bool = True,
1542
) -> dict[str, Any]:
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
    # We exclude chat_template from kwargs here, because
    # chat template has been already resolved at this stage
    unexpected_vars = {"chat_template", "tokenize"}
    if raise_on_unexpected and (
        unexpected_in_kwargs := unexpected_vars & chat_template_kwargs.keys()
    ):
        raise ValueError(
            "Found unexpected chat template kwargs from request: "
            f"{unexpected_in_kwargs}"
        )

1554
    fn_kw = {
1555
1556
        k
        for k in chat_template_kwargs
1557
1558
        if supports_kw(tokenizer.apply_chat_template, k, allow_var_kwargs=False)
    }
1559
    template_vars = _cached_resolve_chat_template_kwargs(chat_template)
1560
1561
1562
1563
1564

    # Allow standard HF parameters even if tokenizer uses **kwargs to receive them
    hf_base_params = _get_hf_base_chat_template_params()

    accept_vars = (fn_kw | template_vars | hf_base_params) - unexpected_vars
1565
    return {k: v for k, v in chat_template_kwargs.items() if k in accept_vars}
1566
1567


1568
def apply_hf_chat_template(
1569
    tokenizer: PreTrainedTokenizer | PreTrainedTokenizerFast,
1570
    conversation: list[ConversationMessage],
1571
1572
    chat_template: str | None,
    tools: list[dict[str, Any]] | None,
1573
    *,
1574
    model_config: ModelConfig,
1575
    **kwargs: Any,
1576
) -> str:
1577
    hf_chat_template = resolve_hf_chat_template(
1578
1579
1580
        tokenizer,
        chat_template=chat_template,
        tools=tools,
1581
        model_config=model_config,
1582
    )
1583

1584
    if hf_chat_template is None:
1585
1586
1587
        raise ValueError(
            "As of transformers v4.44, default chat template is no longer "
            "allowed, so you must provide a chat template if the tokenizer "
1588
1589
            "does not define one."
        )
1590

1591
1592
1593
1594
1595
1596
    resolved_kwargs = resolve_chat_template_kwargs(
        tokenizer=tokenizer,
        chat_template=hf_chat_template,
        chat_template_kwargs=kwargs,
    )

1597
1598
1599
1600
1601
    try:
        return tokenizer.apply_chat_template(
            conversation=conversation,  # type: ignore[arg-type]
            tools=tools,  # type: ignore[arg-type]
            chat_template=hf_chat_template,
1602
            tokenize=False,
1603
            **resolved_kwargs,
1604
        )
1605

1606
1607
1608
1609
1610
1611
    # External library exceptions can sometimes occur despite the framework's
    # internal exception management capabilities.
    except Exception as e:
        # Log and report any library-related exceptions for further
        # investigation.
        logger.exception(
1612
1613
            "An error occurred in `transformers` while applying chat template"
        )
1614
        raise ValueError(str(e)) from e
1615

1616

1617
1618
def apply_mistral_chat_template(
    tokenizer: MistralTokenizer,
1619
    messages: list[ChatCompletionMessageParam],
1620
1621
    chat_template: str | None,
    tools: list[dict[str, Any]] | None,
1622
    **kwargs: Any,
1623
) -> list[int]:
1624
1625
    from mistral_common.exceptions import MistralCommonException

1626
1627
1628
1629
1630
1631
    # The return value of resolve_mistral_chat_template is always None,
    # and we won't use it.
    resolve_mistral_chat_template(
        chat_template=chat_template,
        **kwargs,
    )
1632

1633
1634
1635
1636
1637
1638
1639
1640
1641
    try:
        return tokenizer.apply_chat_template(
            messages=messages,
            tools=tools,
            **kwargs,
        )
    # mistral-common uses assert statements to stop processing of input
    # if input does not comply with the expected format.
    # We convert those assertion errors to ValueErrors so they can be
1642
    # properly caught in the preprocessing_input step
1643
    except (AssertionError, MistralCommonException) as e:
1644
        raise ValueError(str(e)) from e
1645
1646
1647
1648
1649
1650
1651

    # External library exceptions can sometimes occur despite the framework's
    # internal exception management capabilities.
    except Exception as e:
        # Log and report any library-related exceptions for further
        # investigation.
        logger.exception(
1652
1653
            "An error occurred in `mistral_common` while applying chat template"
        )
1654
        raise ValueError(str(e)) from e
1655

1656

1657
1658
1659
def get_history_tool_calls_cnt(conversation: list[ConversationMessage]):
    idx = 0
    for msg in conversation:
1660
1661
1662
        if msg["role"] == "assistant":
            tool_calls = msg.get("tool_calls")
            idx += len(list(tool_calls)) if tool_calls is not None else 0  # noqa
1663
1664
1665
    return idx


1666
1667
1668
def make_tool_call_id(id_type: str = "random", func_name=None, idx=None):
    if id_type == "kimi_k2":
        return f"functions.{func_name}:{idx}"
1669
1670
1671
    else:
        # by default return random
        return f"chatcmpl-tool-{random_uuid()}"