chat_utils.py 17.9 KB
Newer Older
1
import asyncio
2
import codecs
3
import json
4
from abc import ABC, abstractmethod
5
from collections import defaultdict
6
from functools import lru_cache, partial
7
from pathlib import Path
8
from typing import (Any, Awaitable, Dict, Generic, Iterable, List, Literal,
9
                    Mapping, Optional, Tuple, TypeVar, Union, cast)
10

11
12
# yapf conflicts with isort for this block
# yapf: disable
13
14
from openai.types.chat import (ChatCompletionAssistantMessageParam,
                               ChatCompletionContentPartImageParam)
15
16
from openai.types.chat import (
    ChatCompletionContentPartParam as OpenAIChatCompletionContentPartParam)
17
18
from openai.types.chat import (ChatCompletionContentPartRefusalParam,
                               ChatCompletionContentPartTextParam)
19
20
from openai.types.chat import (
    ChatCompletionMessageParam as OpenAIChatCompletionMessageParam)
21
22
from openai.types.chat import (ChatCompletionMessageToolCallParam,
                               ChatCompletionToolMessageParam)
23
24
# yapf: enable
# pydantic needs the TypedDict from typing_extensions
25
from pydantic import ConfigDict
26
from transformers import PreTrainedTokenizer, PreTrainedTokenizerFast
27
from typing_extensions import Required, TypeAlias, TypedDict
28

29
from vllm.config import ModelConfig
30
31
from vllm.logger import init_logger
from vllm.multimodal import MultiModalDataDict
32
from vllm.multimodal.utils import (async_get_and_parse_audio,
33
34
                                   async_get_and_parse_image,
                                   get_and_parse_audio, get_and_parse_image)
35
from vllm.transformers_utils.tokenizer import AnyTokenizer, MistralTokenizer
36
37
38
39

logger = init_logger(__name__)


40
41
42
43
44
45
46
47
48
49
50
51
52
53
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."""


54
55
56
57
58
59
60
class CustomChatCompletionContentPartParam(TypedDict, total=False):
    __pydantic_config__ = ConfigDict(extra="allow")  # type: ignore

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


61
62
ChatCompletionContentPartParam: TypeAlias = Union[
    OpenAIChatCompletionContentPartParam, ChatCompletionContentPartAudioParam,
63
64
    ChatCompletionContentPartRefusalParam,
    CustomChatCompletionContentPartParam]
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81


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

    content: Union[str, List[ChatCompletionContentPartParam]]
    """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.
    """

82
83
84
85
86
87
    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."""

88
89
90
91
92

ChatCompletionMessageParam = Union[OpenAIChatCompletionMessageParam,
                                   CustomChatCompletionMessageParam]


93
# TODO: Make fields ReadOnly once mypy supports it
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
class ConversationMessage(TypedDict, total=False):
    role: Required[str]
    """The role of the message's author."""

    content: Optional[str]
    """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."""
109
110


111
112
113
114
115
ModalityStr = Literal["image", "audio"]
_T = TypeVar("_T")


class BaseMultiModalItemTracker(ABC, Generic[_T]):
116
117
118
119
120
121
122
    """
    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):
123
124
        super().__init__()

125
126
127
128
129
        self._model_config = model_config
        self._tokenizer = tokenizer
        self._allowed_items = (model_config.multimodal_config.limit_per_prompt
                               if model_config.multimodal_config else {})
        self._consumed_items = {k: 0 for k in self._allowed_items}
130
131

        self._items: List[_T] = []
132
133
134

    @staticmethod
    @lru_cache(maxsize=None)
135
    def _cached_token_str(tokenizer: AnyTokenizer, token_index: int) -> str:
136
137
        return tokenizer.decode(token_index)

138
139
    def _placeholder_str(self, modality: ModalityStr,
                         current_count: int) -> Optional[str]:
140
141
        # TODO: Let user specify how to insert image tokens into prompt
        # (similar to chat template)
142
143
144
        hf_config = self._model_config.hf_config
        model_type = hf_config.model_type

145
146
147
148
149
150
151
152
153
        if modality == "image":
            if model_type == "phi3_v":
                # Workaround since this token is not defined in the tokenizer
                return f"<|image_{current_count}|>"
            if model_type == "minicpmv":
                return "(<image>./</image>)"
            if model_type in ("blip-2", "chatglm", "fuyu", "paligemma"):
                # These models do not use image tokens in the prompt
                return None
154
155
            if model_type == "qwen":
                return f"Picture {current_count}: <img></img>"
156
            if model_type.startswith("llava"):
157
158
                return self._cached_token_str(self._tokenizer,
                                              hf_config.image_token_index)
159
160
161
162
163
164
165
166
167
168
169
170
            if model_type in ("chameleon", "internvl_chat"):
                return "<image>"

            raise TypeError(f"Unknown model type: {model_type}")
        elif modality == "audio":
            if model_type == "ultravox":
                return "<|reserved_special_token_0|>"
            raise TypeError(f"Unknown model type: {model_type}")
        else:
            raise TypeError(f"Unknown modality: {modality}")

    @staticmethod
171
    def _combine(items: List[MultiModalDataDict]) -> MultiModalDataDict:
172
173
174
        mm_lists: Mapping[str, List[object]] = defaultdict(list)

        # Merge all the multi-modal items
175
        for single_mm_data in items:
176
177
178
179
180
181
182
183
184
185
186
187
            for mm_key, mm_item in single_mm_data.items():
                if isinstance(mm_item, list):
                    mm_lists[mm_key].extend(mm_item)
                else:
                    mm_lists[mm_key].append(mm_item)

        # Unpack any single item lists for models that don't expect multiple.
        return {
            mm_key: mm_list[0] if len(mm_list) == 1 else mm_list
            for mm_key, mm_list in mm_lists.items()
        }

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
    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)
        current_count = self._consumed_items.get(modality, 0) + 1
        if current_count > allowed_count:
            raise ValueError(
                f"At most {allowed_count} {modality}(s) may be provided in "
                "one request.")

        self._consumed_items[modality] = current_count
        self._items.append(item)

        return self._placeholder_str(modality, current_count)

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


class MultiModalItemTracker(BaseMultiModalItemTracker[MultiModalDataDict]):

    def all_mm_data(self) -> Optional[MultiModalDataDict]:
        return self._combine(self._items) if self._items else None

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


class AsyncMultiModalItemTracker(
        BaseMultiModalItemTracker[Awaitable[MultiModalDataDict]]):

    async def all_mm_data(self) -> Optional[MultiModalDataDict]:
        if self._items:
            items = await asyncio.gather(*self._items)
            return self._combine(items)

        return None

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


class BaseMultiModalContentParser(ABC):

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

        # multimodal placeholder_string : count
        self._placeholder_counts: Dict[str, int] = defaultdict(lambda: 0)

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

    def mm_placeholder_counts(self) -> Dict[str, int]:
        return dict(self._placeholder_counts)

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

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


class MultiModalContentParser(BaseMultiModalContentParser):

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

        self._tracker = tracker

    def parse_image(self, image_url: str) -> None:
        image = get_and_parse_image(image_url)

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

    def parse_audio(self, audio_url: str) -> None:
        audio = get_and_parse_audio(audio_url)

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


class AsyncMultiModalContentParser(BaseMultiModalContentParser):

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

        self._tracker = tracker

    def parse_image(self, image_url: str) -> None:
        image_coro = async_get_and_parse_image(image_url)

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

    def parse_audio(self, audio_url: str) -> None:
        audio_coro = async_get_and_parse_audio(audio_url)

        placeholder = self._tracker.add("audio", audio_coro)
        self._add_placeholder(placeholder)
295
296


297
298
def load_chat_template(
        chat_template: Optional[Union[Path, str]]) -> Optional[str]:
299
300
301
302
303
304
    if chat_template is None:
        return None
    try:
        with open(chat_template, "r") as f:
            resolved_chat_template = f.read()
    except OSError as e:
305
306
307
        if isinstance(chat_template, Path):
            raise

308
309
310
311
312
313
        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
314

315
316
317
        # If opening a file fails, set chat template to be args to
        # ensure we decode so our escape are interpreted correctly
        resolved_chat_template = codecs.decode(chat_template, "unicode_escape")
318

319
320
    logger.info("Using supplied chat template:\n%s", resolved_chat_template)
    return resolved_chat_template
321
322


323
# TODO: Let user specify how to insert multimodal tokens into prompt
324
# (similar to chat template)
325
def _get_full_multimodal_text_prompt(placeholder_counts: Dict[str, int],
326
                                     text_prompt: str) -> str:
327
    """Combine multimodal prompts for a multimodal language model."""
328

329
    # Look through the text prompt to check for missing placeholders
330
    missing_placeholders: List[str] = []
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
    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])
347
348


349
350
351
352
353
# No need to validate using Pydantic again
_TextParser = partial(cast, ChatCompletionContentPartTextParam)
_ImageParser = partial(cast, ChatCompletionContentPartImageParam)
_AudioParser = partial(cast, ChatCompletionContentPartAudioParam)
_RefusalParser = partial(cast, ChatCompletionContentPartRefusalParam)
354
355


356
357
358
def _parse_chat_message_content_parts(
    role: str,
    parts: Iterable[ChatCompletionContentPartParam],
359
    mm_tracker: BaseMultiModalItemTracker,
360
) -> List[ConversationMessage]:
361
    texts: List[str] = []
362

363
    mm_parser = mm_tracker.create_parser()
364
365
366
367

    for part in parts:
        part_type = part["type"]
        if part_type == "text":
368
            text = _TextParser(part)["text"]
369
370
            texts.append(text)
        elif part_type == "image_url":
371
            image_url = _ImageParser(part)["image_url"]
372
373
374
375
376
377

            if image_url.get("detail", "auto") != "auto":
                logger.warning(
                    "'image_url.detail' is currently not supported and "
                    "will be ignored.")

378
            mm_parser.parse_image(image_url["url"])
379
        elif part_type == "audio_url":
380
            audio_url = _AudioParser(part)["audio_url"]
381
382

            mm_parser.parse_audio(audio_url["url"])
383
384
385
        elif part_type == "refusal":
            text = _RefusalParser(part)["refusal"]
            texts.append(text)
386
387
388
389
        else:
            raise NotImplementedError(f"Unknown part type: {part_type}")

    text_prompt = "\n".join(texts)
390
    mm_placeholder_counts = mm_parser.mm_placeholder_counts()
391
392
393
    if mm_placeholder_counts:
        text_prompt = _get_full_multimodal_text_prompt(mm_placeholder_counts,
                                                       text_prompt)
394

395
    return [ConversationMessage(role=role, content=text_prompt)]
396
397


398
399
400
401
402
# No need to validate using Pydantic again
_AssistantParser = partial(cast, ChatCompletionAssistantMessageParam)
_ToolParser = partial(cast, ChatCompletionToolMessageParam)


403
def _parse_chat_message_content(
404
405
406
    message: ChatCompletionMessageParam,
    mm_tracker: BaseMultiModalItemTracker,
) -> List[ConversationMessage]:
407
408
409
410
    role = message["role"]
    content = message.get("content")

    if content is None:
411
412
413
414
415
        content = []
    elif isinstance(content, str):
        content = [
            ChatCompletionContentPartTextParam(type="text", text=content)
        ]
416

417
    result = _parse_chat_message_content_parts(
418
419
        role,
        content,  # type: ignore
420
        mm_tracker,
421
    )
422

423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
    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

439

440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
def _postprocess_messages(messages: List[ConversationMessage]) -> None:
    # 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"])


455
456
457
def parse_chat_messages(
    messages: List[ChatCompletionMessageParam],
    model_config: ModelConfig,
458
    tokenizer: AnyTokenizer,
459
) -> Tuple[List[ConversationMessage], Optional[MultiModalDataDict]]:
460
    conversation: List[ConversationMessage] = []
461
    mm_tracker = MultiModalItemTracker(model_config, tokenizer)
462
463

    for msg in messages:
464
        sub_messages = _parse_chat_message_content(msg, mm_tracker)
465

466
        conversation.extend(sub_messages)
467

468
469
    _postprocess_messages(conversation)

470
    return conversation, mm_tracker.all_mm_data()
471
472


473
474
475
476
477
478
479
480
481
482
483
484
485
def parse_chat_messages_futures(
    messages: List[ChatCompletionMessageParam],
    model_config: ModelConfig,
    tokenizer: AnyTokenizer,
) -> Tuple[List[ConversationMessage], Awaitable[Optional[MultiModalDataDict]]]:
    conversation: List[ConversationMessage] = []
    mm_tracker = AsyncMultiModalItemTracker(model_config, tokenizer)

    for msg in messages:
        sub_messages = _parse_chat_message_content(msg, mm_tracker)

        conversation.extend(sub_messages)

486
487
    _postprocess_messages(conversation)

488
489
490
    return conversation, mm_tracker.all_mm_data()


491
492
def apply_hf_chat_template(
    tokenizer: Union[PreTrainedTokenizer, PreTrainedTokenizerFast],
493
494
495
496
497
    conversation: List[ConversationMessage],
    chat_template: Optional[str],
    *,
    tokenize: bool = False,  # Different from HF's default
    **kwargs: Any,
498
) -> str:
499
500
501
502
503
504
    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.")

505
506
507
508
509
510
    return tokenizer.apply_chat_template(
        conversation=conversation,  # type: ignore[arg-type]
        chat_template=chat_template,
        tokenize=tokenize,
        **kwargs,
    )
511
512


513
514
515
516
517
518
519
520
def apply_mistral_chat_template(
    tokenizer: MistralTokenizer,
    messages: List[ChatCompletionMessageParam],
    chat_template: Optional[str],
    **kwargs: Any,
) -> List[int]:
    return tokenizer.apply_chat_template(
        messages=messages,
521
522
523
        chat_template=chat_template,
        **kwargs,
    )