protocol.py 68.4 KB
Newer Older
1
2
# SPDX-License-Identifier: Apache-2.0

3
4
# Adapted from
# https://github.com/lm-sys/FastChat/blob/168ccc29d3f7edc50823016105c024fe2282732a/fastchat/protocol/openai_api_protocol.py
5
import json
6
import re
Zhuohan Li's avatar
Zhuohan Li committed
7
import time
8
from http import HTTPStatus
9
from typing import Annotated, Any, ClassVar, Literal, Optional, Union
Zhuohan Li's avatar
Zhuohan Li committed
10

11
import torch
12
from fastapi import HTTPException, UploadFile
13
14
from pydantic import (BaseModel, ConfigDict, Field, TypeAdapter,
                      ValidationInfo, field_validator, model_validator)
15
from typing_extensions import TypeAlias
Zhuohan Li's avatar
Zhuohan Li committed
16

17
from vllm import envs
18
19
from vllm.entrypoints.chat_utils import (ChatCompletionMessageParam,
                                         random_tool_call_id)
20
from vllm.logger import init_logger
21
from vllm.pooling_params import PoolingParams
22
23
from vllm.sampling_params import (BeamSearchParams, GuidedDecodingParams,
                                  RequestOutputKind, SamplingParams)
24
from vllm.sequence import Logprob
25
from vllm.utils import random_uuid, resolve_obj_by_qualname
26

27
28
logger = init_logger(__name__)

29
_LONG_INFO = torch.iinfo(torch.long)
30

Zhuohan Li's avatar
Zhuohan Li committed
31

32
class OpenAIBaseModel(BaseModel):
33
34
35
    # OpenAI API does allow extra fields
    model_config = ConfigDict(extra="allow")

36
    # Cache class field names
37
    field_names: ClassVar[Optional[set[str]]] = None
38

39
    @model_validator(mode="wrap")
40
    @classmethod
41
42
43
44
    def __log_extra_fields__(cls, data, handler):
        result = handler(data)
        if not isinstance(data, dict):
            return result
45
46
        field_names = cls.field_names
        if field_names is None:
47
48
49
50
            # Get all class field names and their potential aliases
            field_names = set()
            for field_name, field in cls.model_fields.items():
                field_names.add(field_name)
51
                if alias := getattr(field, "alias", None):
52
53
54
55
56
57
58
59
                    field_names.add(alias)
            cls.field_names = field_names

        # Compare against both field names and aliases
        if any(k not in field_names for k in data):
            logger.warning(
                "The following fields were present in the request "
                "but ignored: %s",
60
61
                data.keys() - field_names,
            )
62
        return result
63
64
65


class ErrorResponse(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
66
67
68
69
    object: str = "error"
    message: str
    type: str
    param: Optional[str] = None
70
    code: int
Zhuohan Li's avatar
Zhuohan Li committed
71
72


73
class ModelPermission(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
74
75
76
77
78
79
80
81
82
83
84
    id: str = Field(default_factory=lambda: f"modelperm-{random_uuid()}")
    object: str = "model_permission"
    created: int = Field(default_factory=lambda: int(time.time()))
    allow_create_engine: bool = False
    allow_sampling: bool = True
    allow_logprobs: bool = True
    allow_search_indices: bool = False
    allow_view: bool = True
    allow_fine_tuning: bool = False
    organization: str = "*"
    group: Optional[str] = None
85
    is_blocking: bool = False
Zhuohan Li's avatar
Zhuohan Li committed
86
87


88
class ModelCard(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
89
90
91
    id: str
    object: str = "model"
    created: int = Field(default_factory=lambda: int(time.time()))
Woosuk Kwon's avatar
Woosuk Kwon committed
92
    owned_by: str = "vllm"
Zhuohan Li's avatar
Zhuohan Li committed
93
94
    root: Optional[str] = None
    parent: Optional[str] = None
95
    max_model_len: Optional[int] = None
96
    permission: list[ModelPermission] = Field(default_factory=list)
Zhuohan Li's avatar
Zhuohan Li committed
97
98


99
class ModelList(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
100
    object: str = "list"
101
    data: list[ModelCard] = Field(default_factory=list)
Zhuohan Li's avatar
Zhuohan Li committed
102
103


104
105
106
107
class PromptTokenUsageInfo(OpenAIBaseModel):
    cached_tokens: Optional[int] = None


108
class UsageInfo(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
109
110
111
    prompt_tokens: int = 0
    total_tokens: int = 0
    completion_tokens: Optional[int] = 0
112
    prompt_tokens_details: Optional[PromptTokenUsageInfo] = None
Zhuohan Li's avatar
Zhuohan Li committed
113
114


115
116
117
118
119
class RequestResponseMetadata(BaseModel):
    request_id: str
    final_usage_info: Optional[UsageInfo] = None


120
121
122
123
124
class JsonSchemaResponseFormat(OpenAIBaseModel):
    name: str
    description: Optional[str] = None
    # schema is the field in openai but that causes conflicts with pydantic so
    # instead use json_schema with an alias
125
    json_schema: Optional[dict[str, Any]] = Field(default=None, alias='schema')
126
127
128
    strict: Optional[bool] = None


129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
class StructuralTag(OpenAIBaseModel):
    begin: str
    # schema is the field, but that causes conflicts with pydantic so
    # instead use structural_tag_schema with an alias
    structural_tag_schema: Optional[dict[str, Any]] = Field(default=None,
                                                            alias="schema")
    end: str


class StructuralTagResponseFormat(OpenAIBaseModel):
    type: Literal["structural_tag"]
    structures: list[StructuralTag]
    triggers: list[str]


144
class ResponseFormat(OpenAIBaseModel):
145
    # type must be "json_schema", "json_object", or "text"
146
147
    type: Literal["text", "json_object", "json_schema"]
    json_schema: Optional[JsonSchemaResponseFormat] = None
148
149


150
151
152
AnyResponseFormat = Union[ResponseFormat, StructuralTagResponseFormat]


153
class StreamOptions(OpenAIBaseModel):
154
    include_usage: Optional[bool] = True
155
    continuous_usage_stats: Optional[bool] = False
156
157


158
159
160
class FunctionDefinition(OpenAIBaseModel):
    name: str
    description: Optional[str] = None
161
    parameters: Optional[dict[str, Any]] = None
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177


class ChatCompletionToolsParam(OpenAIBaseModel):
    type: Literal["function"] = "function"
    function: FunctionDefinition


class ChatCompletionNamedFunction(OpenAIBaseModel):
    name: str


class ChatCompletionNamedToolChoiceParam(OpenAIBaseModel):
    function: ChatCompletionNamedFunction
    type: Literal["function"] = "function"


178
179
class LogitsProcessorConstructor(BaseModel):
    qualname: str
180
181
    args: Optional[list[Any]] = None
    kwargs: Optional[dict[str, Any]] = None
182
183


184
LogitsProcessors = list[Union[str, LogitsProcessorConstructor]]
185
186
187


def get_logits_processors(processors: Optional[LogitsProcessors],
188
                          pattern: Optional[str]) -> Optional[list[Any]]:
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
    if processors and pattern:
        logits_processors = []
        for processor in processors:
            qualname = processor if isinstance(processor,
                                               str) else processor.qualname
            if not re.match(pattern, qualname):
                raise ValueError(
                    f"Logits processor '{qualname}' is not allowed by this "
                    "server. See --logits-processor-pattern engine argument "
                    "for more information.")
            try:
                logits_processor = resolve_obj_by_qualname(qualname)
            except Exception as e:
                raise ValueError(
                    f"Logits processor '{qualname}' could not be resolved: {e}"
                ) from e
            if isinstance(processor, LogitsProcessorConstructor):
                logits_processor = logits_processor(*processor.args or [],
                                                    **processor.kwargs or {})
            logits_processors.append(logits_processor)
        return logits_processors
    elif processors:
        raise ValueError(
            "The `logits_processors` argument is not supported by this "
            "server. See --logits-processor-pattern engine argugment "
            "for more information.")
    return None


218
class ChatCompletionRequest(OpenAIBaseModel):
219
220
    # Ordered by official OpenAI API documentation
    # https://platform.openai.com/docs/api-reference/chat/create
221
    messages: list[ChatCompletionMessageParam]
222
    model: Optional[str] = None
223
    frequency_penalty: Optional[float] = 0.0
224
    logit_bias: Optional[dict[str, float]] = None
225
    logprobs: Optional[bool] = False
226
    top_logprobs: Optional[int] = 0
227
228
229
230
231
232
    # TODO(#9845): remove max_tokens when field is removed from OpenAI API
    max_tokens: Optional[int] = Field(
        default=None,
        deprecated=
        'max_tokens is deprecated in favor of the max_completion_tokens field')
    max_completion_tokens: Optional[int] = None
233
234
    n: Optional[int] = 1
    presence_penalty: Optional[float] = 0.0
235
    response_format: Optional[AnyResponseFormat] = None
236
    seed: Optional[int] = Field(None, ge=_LONG_INFO.min, le=_LONG_INFO.max)
237
    stop: Optional[Union[str, list[str]]] = Field(default_factory=list)
Zhuohan Li's avatar
Zhuohan Li committed
238
    stream: Optional[bool] = False
239
    stream_options: Optional[StreamOptions] = None
240
241
    temperature: Optional[float] = None
    top_p: Optional[float] = None
242
    tools: Optional[list[ChatCompletionToolsParam]] = None
243
244
245
246
247
248
    tool_choice: Optional[Union[
        Literal["none"],
        Literal["auto"],
        Literal["required"],
        ChatCompletionNamedToolChoiceParam,
    ]] = "none"
249

250
    # NOTE this will be ignored by vLLM -- the model determines the behavior
251
    parallel_tool_calls: Optional[bool] = False
Zhuohan Li's avatar
Zhuohan Li committed
252
    user: Optional[str] = None
253

254
    # --8<-- [start:chat-completion-sampling-params]
255
    best_of: Optional[int] = None
256
    use_beam_search: bool = False
257
258
259
    top_k: Optional[int] = None
    min_p: Optional[float] = None
    repetition_penalty: Optional[float] = None
260
    length_penalty: float = 1.0
261
    stop_token_ids: Optional[list[int]] = Field(default_factory=list)
262
263
264
265
266
267
    include_stop_str_in_output: bool = False
    ignore_eos: bool = False
    min_tokens: int = 0
    skip_special_tokens: bool = True
    spaces_between_special_tokens: bool = True
    truncate_prompt_tokens: Optional[Annotated[int, Field(ge=1)]] = None
268
    prompt_logprobs: Optional[int] = None
269
    # --8<-- [end:chat-completion-sampling-params]
270

271
    # --8<-- [start:chat-completion-extra-params]
272
    echo: bool = Field(
273
274
275
276
277
        default=False,
        description=(
            "If true, the new message will be prepended with the last message "
            "if they belong to the same role."),
    )
278
    add_generation_prompt: bool = Field(
279
280
281
282
283
284
        default=True,
        description=
        ("If true, the generation prompt will be added to the chat template. "
         "This is a parameter used by chat template in tokenizer config of the "
         "model."),
    )
285
286
287
288
289
290
291
292
293
    continue_final_message: bool = Field(
        default=False,
        description=
        ("If this is set, the chat will be formatted so that the final "
         "message in the chat is open-ended, without any EOS tokens. The "
         "model will continue this message rather than starting a new one. "
         "This allows you to \"prefill\" part of the model's response for it. "
         "Cannot be used at the same time as `add_generation_prompt`."),
    )
294
    add_special_tokens: bool = Field(
295
296
297
298
299
        default=False,
        description=(
            "If true, special tokens (e.g. BOS) will be added to the prompt "
            "on top of what is added by the chat template. "
            "For most models, the chat template takes care of adding the "
300
            "special tokens so this should be set to false (as is the "
301
302
            "default)."),
    )
303
    documents: Optional[list[dict[str, str]]] = Field(
304
305
306
307
308
309
310
311
312
313
314
315
        default=None,
        description=
        ("A list of dicts representing documents that will be accessible to "
         "the model if it is performing RAG (retrieval-augmented generation)."
         " If the template does not support RAG, this argument will have no "
         "effect. We recommend that each document should be a dict containing "
         "\"title\" and \"text\" keys."),
    )
    chat_template: Optional[str] = Field(
        default=None,
        description=(
            "A Jinja template to use for this conversion. "
316
317
318
            "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."),
319
    )
320
    chat_template_kwargs: Optional[dict[str, Any]] = Field(
321
322
323
324
        default=None,
        description=("Additional kwargs to pass to the template renderer. "
                     "Will be accessible by the chat template."),
    )
325
    mm_processor_kwargs: Optional[dict[str, Any]] = Field(
326
327
328
        default=None,
        description=("Additional kwargs to pass to the HF processor."),
    )
329
330
331
332
333
334
335
336
337
    guided_json: Optional[Union[str, dict, BaseModel]] = Field(
        default=None,
        description=("If specified, the output will follow the JSON schema."),
    )
    guided_regex: Optional[str] = Field(
        default=None,
        description=(
            "If specified, the output will follow the regex pattern."),
    )
338
    guided_choice: Optional[list[str]] = Field(
339
340
341
342
343
344
345
346
347
        default=None,
        description=(
            "If specified, the output will be exactly one of the choices."),
    )
    guided_grammar: Optional[str] = Field(
        default=None,
        description=(
            "If specified, the output will follow the context free grammar."),
    )
348
349
350
351
352
    structural_tag: Optional[str] = Field(
        default=None,
        description=(
            "If specified, the output will follow the structural tag schema."),
    )
353
354
355
356
357
    guided_decoding_backend: Optional[str] = Field(
        default=None,
        description=(
            "If specified, will override the default guided decoding backend "
            "of the server for this specific request. If set, must be either "
358
359
            "'outlines' / 'lm-format-enforcer'"),
    )
360
361
362
363
    guided_whitespace_pattern: Optional[str] = Field(
        default=None,
        description=(
            "If specified, will override the default whitespace pattern "
364
365
            "for guided json decoding."),
    )
366
367
368
369
370
    priority: int = Field(
        default=0,
        description=(
            "The priority of the request (lower means earlier handling; "
            "default: 0). Any priority other than 0 will raise an error "
371
372
            "if the served model does not use priority scheduling."),
    )
373
374
375
376
377
    request_id: str = Field(
        default_factory=lambda: f"{random_uuid()}",
        description=(
            "The request_id related to this request. If the caller does "
            "not set it, a random_uuid will be generated. This id is used "
378
379
            "through out the inference process and return in response."),
    )
380
381
382
383
384
385
386
387
388
389
390
    logits_processors: Optional[LogitsProcessors] = Field(
        default=None,
        description=(
            "A list of either qualified names of logits processors, or "
            "constructor objects, to apply when sampling. A constructor is "
            "a JSON object with a required 'qualname' field specifying the "
            "qualified name of the processor class/factory, and optional "
            "'args' and 'kwargs' fields containing positional and keyword "
            "arguments. For example: {'qualname': "
            "'my_module.MyLogitsProcessor', 'args': [1, 2], 'kwargs': "
            "{'param': 'value'}}."))
391
392
393
394
395
396
    return_tokens_as_token_ids: Optional[bool] = Field(
        default=None,
        description=(
            "If specified with 'logprobs', tokens are represented "
            " as strings of the form 'token_id:{token_id}' so that tokens "
            "that are not JSON-encodable can be identified."))
397
398
399
400
401
402
403
404
405
    cache_salt: Optional[str] = Field(
        default=None,
        description=(
            "If specified, the prefix cache will be salted with the provided "
            "string to prevent an attacker to guess prompts in multi-user "
            "environments. The salt should be random, protected from "
            "access by 3rd parties, and long enough to be "
            "unpredictable (e.g., 43 characters base64-encoded, corresponding "
            "to 256 bit). Not supported by vLLM engine V0."))
Robert Shaw's avatar
Robert Shaw committed
406
407
408
    kv_transfer_params: Optional[dict[str, Any]] = Field(
        default=None,
        description="KVTransfer parameters used for disaggregated serving.")
409

410
    # --8<-- [end:chat-completion-extra-params]
Zhuohan Li's avatar
Zhuohan Li committed
411

412
413
414
415
416
    # Default sampling parameters for chat completion requests
    _DEFAULT_SAMPLING_PARAMS: dict = {
        "repetition_penalty": 1.0,
        "temperature": 1.0,
        "top_p": 1.0,
417
        "top_k": 0,
418
419
420
421
422
423
424
425
        "min_p": 0.0,
    }

    def to_beam_search_params(
            self,
            default_max_tokens: int,
            default_sampling_params: Optional[dict] = None
    ) -> BeamSearchParams:
426
427
        # TODO(#9845): remove max_tokens when field is removed from OpenAI API
        max_tokens = self.max_completion_tokens or self.max_tokens
428

429
430
        if default_sampling_params is None:
            default_sampling_params = {}
431
        n = self.n if self.n is not None else 1
432

433
434
435
436
437
438
        # Use minimum of context window, user request & server limit.
        max_tokens = min(
            val for val in (default_max_tokens, max_tokens,
                            default_sampling_params.get("max_tokens", None))
            if val is not None)

439
440
441
        if (temperature := self.temperature) is None:
            temperature = default_sampling_params.get(
                "temperature", self._DEFAULT_SAMPLING_PARAMS["temperature"])
442
443
444
445
446
447

        return BeamSearchParams(
            beam_width=n,
            max_tokens=max_tokens,
            ignore_eos=self.ignore_eos,
            temperature=temperature,
448
            length_penalty=self.length_penalty,
449
450
            include_stop_str_in_output=self.include_stop_str_in_output,
        )
451

452
    def to_sampling_params(
453
454
455
456
457
        self,
        default_max_tokens: int,
        logits_processor_pattern: Optional[str],
        default_sampling_params: Optional[dict] = None,
    ) -> SamplingParams:
458
459
        # TODO(#9845): remove max_tokens when field is removed from OpenAI API
        max_tokens = self.max_completion_tokens or self.max_tokens
460

461
462
        if default_sampling_params is None:
            default_sampling_params = {}
463
464
465
466
467
468
469

        # Use minimum of context window, user request & server limit.
        max_tokens = min(
            val for val in (default_max_tokens, max_tokens,
                            default_sampling_params.get("max_tokens", None))
            if val is not None)

470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
        # Default parameters
        if (repetition_penalty := self.repetition_penalty) is None:
            repetition_penalty = default_sampling_params.get(
                "repetition_penalty",
                self._DEFAULT_SAMPLING_PARAMS["repetition_penalty"],
            )
        if (temperature := self.temperature) is None:
            temperature = default_sampling_params.get(
                "temperature", self._DEFAULT_SAMPLING_PARAMS["temperature"])
        if (top_p := self.top_p) is None:
            top_p = default_sampling_params.get(
                "top_p", self._DEFAULT_SAMPLING_PARAMS["top_p"])
        if (top_k := self.top_k) is None:
            top_k = default_sampling_params.get(
                "top_k", self._DEFAULT_SAMPLING_PARAMS["top_k"])
        if (min_p := self.min_p) is None:
            min_p = default_sampling_params.get(
                "min_p", self._DEFAULT_SAMPLING_PARAMS["min_p"])

489
490
491
492
        prompt_logprobs = self.prompt_logprobs
        if prompt_logprobs is None and self.echo:
            prompt_logprobs = self.top_logprobs

493
        guided_json_object = None
494
495
496
497
498
499
500
        if self.response_format is not None:
            if self.response_format.type == "json_object":
                guided_json_object = True
            elif self.response_format.type == "json_schema":
                json_schema = self.response_format.json_schema
                assert json_schema is not None
                self.guided_json = json_schema.json_schema
501
502
503
504
505
506
            elif self.response_format.type == "structural_tag":
                structural_tag = self.response_format
                assert structural_tag is not None and isinstance(
                    structural_tag, StructuralTagResponseFormat)
                s_tag_obj = structural_tag.model_dump(by_alias=True)
                self.structural_tag = json.dumps(s_tag_obj)
507
508
509
510
511
512
513
514

        guided_decoding = GuidedDecodingParams.from_optional(
            json=self._get_guided_json_from_tool() or self.guided_json,
            regex=self.guided_regex,
            choice=self.guided_choice,
            grammar=self.guided_grammar,
            json_object=guided_json_object,
            backend=self.guided_decoding_backend,
515
            whitespace_pattern=self.guided_whitespace_pattern,
516
            structural_tag=self.structural_tag,
517
        )
518

519
        return SamplingParams.from_optional(
520
            n=self.n,
521
            best_of=self.best_of,
522
523
            presence_penalty=self.presence_penalty,
            frequency_penalty=self.frequency_penalty,
524
525
526
527
528
            repetition_penalty=repetition_penalty,
            temperature=temperature,
            top_p=top_p,
            top_k=top_k,
            min_p=min_p,
Nick Hill's avatar
Nick Hill committed
529
            seed=self.seed,
530
531
            stop=self.stop,
            stop_token_ids=self.stop_token_ids,
532
            logprobs=self.top_logprobs if self.logprobs else None,
533
            prompt_logprobs=prompt_logprobs,
534
            ignore_eos=self.ignore_eos,
535
            max_tokens=max_tokens,
536
            min_tokens=self.min_tokens,
537
538
            skip_special_tokens=self.skip_special_tokens,
            spaces_between_special_tokens=self.spaces_between_special_tokens,
539
540
            logits_processors=get_logits_processors(self.logits_processors,
                                                    logits_processor_pattern),
541
            include_stop_str_in_output=self.include_stop_str_in_output,
542
            truncate_prompt_tokens=self.truncate_prompt_tokens,
543
544
            output_kind=RequestOutputKind.DELTA if self.stream \
                else RequestOutputKind.FINAL_ONLY,
545
            guided_decoding=guided_decoding,
Robert Shaw's avatar
Robert Shaw committed
546
547
548
            logit_bias=self.logit_bias,
            extra_args=({"kv_transfer_params": self.kv_transfer_params}
                        if self.kv_transfer_params else None))
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565

    def _get_guided_json_from_tool(
            self) -> Optional[Union[str, dict, BaseModel]]:
        # user has chosen to not use any tool
        if self.tool_choice == "none" or self.tools is None:
            return None

        # user has chosen to use a named tool
        if type(self.tool_choice) is ChatCompletionNamedToolChoiceParam:
            tool_name = self.tool_choice.function.name
            tools = {tool.function.name: tool.function for tool in self.tools}
            if tool_name not in tools:
                raise ValueError(
                    f"Tool '{tool_name}' has not been passed in `tools`.")
            tool = tools[tool_name]
            return tool.parameters

566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
        if self.tool_choice == "required":
            # Pydantic schema generation cannot be used since the JSON schema
            # has to be constructed for a specific instantiation of a tool list
            # so that parameters of a function are correctly generated
            # based on the chosen function name
            def get_tool_schema(tool: ChatCompletionToolsParam) -> dict:
                return {
                    "properties": {
                        "name": {
                            "type": "string",
                            "enum": [tool.function.name]
                        },
                        # parameters are always generated as '{}' in the final
                        # output if they are missing from the request
                        # (i.e. are None or '{}') so the schema is
                        # updated to produce an empty object in that case
                        "parameters": tool.function.parameters
                        if tool.function.parameters else {
                            "type": "object",
                            "properties": {}
                        }
                    },
                    "required": ["name", "parameters"]
                }

            json_schema = {
                "type": "array",
                "minItems": 1,
                "items": {
                    "type": "object",
                    "anyOf": [get_tool_schema(tool) for tool in self.tools]
                }
            }
            return json_schema

601
        return None
602

603
    @model_validator(mode="before")
604
    @classmethod
605
606
    def validate_stream_options(cls, data):
        if data.get("stream_options") and not data.get("stream"):
607
            raise ValueError(
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
                "Stream options can only be defined when `stream=True`.")

        return data

    @model_validator(mode="before")
    @classmethod
    def check_logprobs(cls, data):
        if (prompt_logprobs := data.get("prompt_logprobs")) is not None:
            if data.get("stream") and prompt_logprobs > 0:
                raise ValueError(
                    "`prompt_logprobs` are not available when `stream=True`.")

            if prompt_logprobs < 0:
                raise ValueError("`prompt_logprobs` must be a positive value.")

        if (top_logprobs := data.get("top_logprobs")) is not None:
            if top_logprobs < 0:
                raise ValueError("`top_logprobs` must be a positive value.")

627
            if top_logprobs > 0 and not data.get("logprobs"):
628
629
630
631
632
                raise ValueError(
                    "when using `top_logprobs`, `logprobs` must be set to true."
                )

        return data
633

634
635
636
    @model_validator(mode="before")
    @classmethod
    def check_guided_decoding_count(cls, data):
637
638
639
        if isinstance(data, ValueError):
            raise data

640
641
642
643
644
        guide_count = sum([
            "guided_json" in data and data["guided_json"] is not None,
            "guided_regex" in data and data["guided_regex"] is not None,
            "guided_choice" in data and data["guided_choice"] is not None
        ])
645
        # you can only use one kind of guided decoding
646
647
648
649
        if guide_count > 1:
            raise ValueError(
                "You can only use one kind of guided decoding "
                "('guided_json', 'guided_regex' or 'guided_choice').")
650
        # you can only either use guided decoding or tools, not both
651
652
653
654
655
        if guide_count > 1 and data.get("tool_choice", "none") not in (
                "none",
                "auto",
                "required",
        ):
656
657
658
659
660
661
            raise ValueError(
                "You can only either use guided decoding or tools, not both.")
        return data

    @model_validator(mode="before")
    @classmethod
662
663
664
665
    def check_tool_usage(cls, data):

        # if "tool_choice" is not specified but tools are provided,
        # default to "auto" tool_choice
666
        if "tool_choice" not in data and data.get("tools"):
667
668
            data["tool_choice"] = "auto"

669
670
671
672
673
674
        # if "tool_choice" is "none" -- ignore tools if present
        if "tool_choice" in data and data["tool_choice"] == "none":
            # ensure that no tools are present
            data.pop("tools", None)
            return data

675
676
677
678
        # if "tool_choice" is specified -- validation
        if "tool_choice" in data:

            # ensure that if "tool choice" is specified, tools are present
679
680
681
            if "tools" not in data or data["tools"] is None:
                raise ValueError(
                    "When using `tool_choice`, `tools` must be set.")
682
683

            # make sure that tool choice is either a named tool
684
685
686
687
688
689
690
691
692
            # OR that it's set to "auto" or "required"
            if data["tool_choice"] not in [
                    "auto", "required"
            ] and not isinstance(data["tool_choice"], dict):
                raise NotImplementedError(
                    f'Invalid value for `tool_choice`: {data["tool_choice"]}! '\
                    'Only named tools, "none", "auto" or "required" '\
                    'are supported.'
                )
693
694
695
696
697

            # ensure that if "tool_choice" is specified as an object,
            # it matches a valid tool
            if isinstance(data["tool_choice"], dict):
                valid_tool = False
698
                specified_function = data["tool_choice"].get("function")
699
700
                if not specified_function:
                    raise ValueError(
701
702
                        "Expected field `function` in `tool_choice`."
                        " Correct usage: `{\"type\": \"function\","
703
                        " \"function\": {\"name\": \"my_function\"}}`")
704
                specified_function_name = specified_function.get("name")
705
706
                if not specified_function_name:
                    raise ValueError(
707
708
                        "Expected field `name` in `function` in `tool_choice`."
                        "Correct usage: `{\"type\": \"function\", "
709
710
711
712
713
714
715
716
717
                        "\"function\": {\"name\": \"my_function\"}}`")
                for tool in data["tools"]:
                    if tool["function"]["name"] == specified_function_name:
                        valid_tool = True
                        break
                if not valid_tool:
                    raise ValueError(
                        "The tool specified in `tool_choice` does not match any"
                        " of the specified `tools`")
718
719
        return data

720
721
722
723
724
725
726
727
728
    @model_validator(mode="before")
    @classmethod
    def check_generation_prompt(cls, data):
        if data.get("continue_final_message") and data.get(
                "add_generation_prompt"):
            raise ValueError("Cannot set both `continue_final_message` and "
                             "`add_generation_prompt` to True.")
        return data

729
730
731
732
733
734
735
736
737
738
739
740
741
742
    @model_validator(mode="before")
    @classmethod
    def check_cache_salt_support(cls, data):
        if data.get("cache_salt") is not None:
            if not envs.VLLM_USE_V1:
                raise ValueError(
                    "Parameter 'cache_salt' is not supported with "
                    "this instance of vLLM, which uses engine V0.")
            if not isinstance(data["cache_salt"],
                              str) or not data["cache_salt"]:
                raise ValueError("Parameter 'cache_salt' must be a "
                                 "non-empty string if provided.")
        return data

Zhuohan Li's avatar
Zhuohan Li committed
743

744
class CompletionRequest(OpenAIBaseModel):
745
746
    # Ordered by official OpenAI API documentation
    # https://platform.openai.com/docs/api-reference/completions/create
747
    model: Optional[str] = None
748
749
    prompt: Optional[Union[list[int], list[list[int]], str, list[str]]] = None
    prompt_embeds: Optional[Union[bytes, list[bytes]]] = None
750
    best_of: Optional[int] = None
Zhuohan Li's avatar
Zhuohan Li committed
751
752
    echo: Optional[bool] = False
    frequency_penalty: Optional[float] = 0.0
753
    logit_bias: Optional[dict[str, float]] = None
754
755
    logprobs: Optional[int] = None
    max_tokens: Optional[int] = 16
756
    n: int = 1
757
    presence_penalty: Optional[float] = 0.0
758
    seed: Optional[int] = Field(None, ge=_LONG_INFO.min, le=_LONG_INFO.max)
759
    stop: Optional[Union[str, list[str]]] = Field(default_factory=list)
760
    stream: Optional[bool] = False
761
    stream_options: Optional[StreamOptions] = None
762
    suffix: Optional[str] = None
763
764
    temperature: Optional[float] = None
    top_p: Optional[float] = None
Zhuohan Li's avatar
Zhuohan Li committed
765
    user: Optional[str] = None
766

767
    # --8<-- [start:completion-sampling-params]
768
    use_beam_search: bool = False
769
770
771
    top_k: Optional[int] = None
    min_p: Optional[float] = None
    repetition_penalty: Optional[float] = None
772
    length_penalty: float = 1.0
773
    stop_token_ids: Optional[list[int]] = Field(default_factory=list)
774
775
776
777
778
    include_stop_str_in_output: bool = False
    ignore_eos: bool = False
    min_tokens: int = 0
    skip_special_tokens: bool = True
    spaces_between_special_tokens: bool = True
779
    truncate_prompt_tokens: Optional[Annotated[int, Field(ge=1)]] = None
780
    allowed_token_ids: Optional[list[int]] = None
781
    prompt_logprobs: Optional[int] = None
782
    # --8<-- [end:completion-sampling-params]
783

784
    # --8<-- [start:completion-extra-params]
785
786
    add_special_tokens: bool = Field(
        default=True,
787
        description=(
788
789
            "If true (the default), special tokens (e.g. BOS) will be added to "
            "the prompt."),
790
    )
791
    response_format: Optional[AnyResponseFormat] = Field(
792
        default=None,
793
794
795
796
797
        description=(
            "Similar to chat completion, this parameter specifies the format "
            "of output. Only {'type': 'json_object'}, {'type': 'json_schema'}"
            ", {'type': 'structural_tag'}, or {'type': 'text' } is supported."
        ),
798
799
800
    )
    guided_json: Optional[Union[str, dict, BaseModel]] = Field(
        default=None,
801
        description="If specified, the output will follow the JSON schema.",
802
803
804
805
806
807
    )
    guided_regex: Optional[str] = Field(
        default=None,
        description=(
            "If specified, the output will follow the regex pattern."),
    )
808
    guided_choice: Optional[list[str]] = Field(
809
810
811
812
813
814
815
816
817
        default=None,
        description=(
            "If specified, the output will be exactly one of the choices."),
    )
    guided_grammar: Optional[str] = Field(
        default=None,
        description=(
            "If specified, the output will follow the context free grammar."),
    )
818
819
820
821
822
    guided_decoding_backend: Optional[str] = Field(
        default=None,
        description=(
            "If specified, will override the default guided decoding backend "
            "of the server for this specific request. If set, must be one of "
823
824
            "'outlines' / 'lm-format-enforcer'"),
    )
825
826
827
828
    guided_whitespace_pattern: Optional[str] = Field(
        default=None,
        description=(
            "If specified, will override the default whitespace pattern "
829
830
            "for guided json decoding."),
    )
831
832
833
834
835
    priority: int = Field(
        default=0,
        description=(
            "The priority of the request (lower means earlier handling; "
            "default: 0). Any priority other than 0 will raise an error "
836
837
            "if the served model does not use priority scheduling."),
    )
838
839
840
841
842
843
844
845
846
847
848
    logits_processors: Optional[LogitsProcessors] = Field(
        default=None,
        description=(
            "A list of either qualified names of logits processors, or "
            "constructor objects, to apply when sampling. A constructor is "
            "a JSON object with a required 'qualname' field specifying the "
            "qualified name of the processor class/factory, and optional "
            "'args' and 'kwargs' fields containing positional and keyword "
            "arguments. For example: {'qualname': "
            "'my_module.MyLogitsProcessor', 'args': [1, 2], 'kwargs': "
            "{'param': 'value'}}."))
849

850
851
852
853
854
855
    return_tokens_as_token_ids: Optional[bool] = Field(
        default=None,
        description=(
            "If specified with 'logprobs', tokens are represented "
            " as strings of the form 'token_id:{token_id}' so that tokens "
            "that are not JSON-encodable can be identified."))
856

Robert Shaw's avatar
Robert Shaw committed
857
858
859
860
    kv_transfer_params: Optional[dict[str, Any]] = Field(
        default=None,
        description="KVTransfer parameters used for disaggregated serving.")

861
    # --8<-- [end:completion-extra-params]
Zhuohan Li's avatar
Zhuohan Li committed
862

863
864
865
866
867
    # Default sampling parameters for completion requests
    _DEFAULT_SAMPLING_PARAMS: dict = {
        "repetition_penalty": 1.0,
        "temperature": 1.0,
        "top_p": 1.0,
868
        "top_k": 0,
869
870
871
872
873
874
875
876
        "min_p": 0.0,
    }

    def to_beam_search_params(
            self,
            default_max_tokens: int,
            default_sampling_params: Optional[dict] = None
    ) -> BeamSearchParams:
877
878
        max_tokens = self.max_tokens

879
880
        if default_sampling_params is None:
            default_sampling_params = {}
881
        n = self.n if self.n is not None else 1
882

883
884
885
886
887
888
        # Use minimum of context window, user request & server limit.
        max_tokens = min(
            val for val in (default_max_tokens, max_tokens,
                            default_sampling_params.get("max_tokens", None))
            if val is not None)

889
890
        if (temperature := self.temperature) is None:
            temperature = default_sampling_params.get("temperature", 1.0)
891
892
893
894
895
896

        return BeamSearchParams(
            beam_width=n,
            max_tokens=max_tokens,
            ignore_eos=self.ignore_eos,
            temperature=temperature,
897
            length_penalty=self.length_penalty,
898
899
            include_stop_str_in_output=self.include_stop_str_in_output,
        )
900

901
    def to_sampling_params(
902
903
904
905
906
        self,
        default_max_tokens: int,
        logits_processor_pattern: Optional[str],
        default_sampling_params: Optional[dict] = None,
    ) -> SamplingParams:
907
908
        max_tokens = self.max_tokens

909
910
        if default_sampling_params is None:
            default_sampling_params = {}
911
912
913
914
915
916
917

        # Use minimum of context window, user request & server limit.
        max_tokens = min(
            val for val in (default_max_tokens, max_tokens,
                            default_sampling_params.get("max_tokens", None))
            if val is not None)

918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
        # Default parameters
        if (repetition_penalty := self.repetition_penalty) is None:
            repetition_penalty = default_sampling_params.get(
                "repetition_penalty",
                self._DEFAULT_SAMPLING_PARAMS["repetition_penalty"],
            )
        if (temperature := self.temperature) is None:
            temperature = default_sampling_params.get(
                "temperature", self._DEFAULT_SAMPLING_PARAMS["temperature"])
        if (top_p := self.top_p) is None:
            top_p = default_sampling_params.get(
                "top_p", self._DEFAULT_SAMPLING_PARAMS["top_p"])
        if (top_k := self.top_k) is None:
            top_k = default_sampling_params.get(
                "top_k", self._DEFAULT_SAMPLING_PARAMS["top_k"])
        if (min_p := self.min_p) is None:
            min_p = default_sampling_params.get(
                "min_p", self._DEFAULT_SAMPLING_PARAMS["min_p"])

937
938
939
940
        prompt_logprobs = self.prompt_logprobs
        if prompt_logprobs is None and self.echo:
            prompt_logprobs = self.logprobs

941
942
        echo_without_generation = self.echo and self.max_tokens == 0

943
944
945
946
947
948
949
950
951
952
953
954
        guided_json_object = None
        if (self.response_format is not None
                and self.response_format.type == "json_object"):
            guided_json_object = True

        guided_decoding = GuidedDecodingParams.from_optional(
            json=self.guided_json,
            regex=self.guided_regex,
            choice=self.guided_choice,
            grammar=self.guided_grammar,
            json_object=guided_json_object,
            backend=self.guided_decoding_backend,
955
956
            whitespace_pattern=self.guided_whitespace_pattern,
        )
957

958
        return SamplingParams.from_optional(
959
            n=self.n,
960
            best_of=self.best_of,
961
962
            presence_penalty=self.presence_penalty,
            frequency_penalty=self.frequency_penalty,
963
964
965
966
967
            repetition_penalty=repetition_penalty,
            temperature=temperature,
            top_p=top_p,
            top_k=top_k,
            min_p=min_p,
Nick Hill's avatar
Nick Hill committed
968
            seed=self.seed,
969
970
            stop=self.stop,
            stop_token_ids=self.stop_token_ids,
971
            logprobs=self.logprobs,
972
            ignore_eos=self.ignore_eos,
973
            max_tokens=max_tokens if not echo_without_generation else 1,
974
            min_tokens=self.min_tokens,
975
            prompt_logprobs=prompt_logprobs,
976
            skip_special_tokens=self.skip_special_tokens,
977
            spaces_between_special_tokens=self.spaces_between_special_tokens,
978
            include_stop_str_in_output=self.include_stop_str_in_output,
979
980
            logits_processors=get_logits_processors(self.logits_processors,
                                                    logits_processor_pattern),
981
            truncate_prompt_tokens=self.truncate_prompt_tokens,
982
983
            output_kind=RequestOutputKind.DELTA if self.stream \
                else RequestOutputKind.FINAL_ONLY,
984
985
            guided_decoding=guided_decoding,
            logit_bias=self.logit_bias,
Robert Shaw's avatar
Robert Shaw committed
986
987
988
            allowed_token_ids=self.allowed_token_ids,
            extra_args=({"kv_transfer_params": self.kv_transfer_params}
                        if self.kv_transfer_params else None))
989

990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
    @model_validator(mode="before")
    @classmethod
    def check_guided_decoding_count(cls, data):
        guide_count = sum([
            "guided_json" in data and data["guided_json"] is not None,
            "guided_regex" in data and data["guided_regex"] is not None,
            "guided_choice" in data and data["guided_choice"] is not None
        ])
        if guide_count > 1:
            raise ValueError(
                "You can only use one kind of guided decoding "
                "('guided_json', 'guided_regex' or 'guided_choice').")
        return data

1004
1005
1006
    @model_validator(mode="before")
    @classmethod
    def check_logprobs(cls, data):
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
        if (prompt_logprobs := data.get("prompt_logprobs")) is not None:
            if data.get("stream") and prompt_logprobs > 0:
                raise ValueError(
                    "`prompt_logprobs` are not available when `stream=True`.")

            if prompt_logprobs < 0:
                raise ValueError("`prompt_logprobs` must be a positive value.")

        if (logprobs := data.get("logprobs")) is not None and logprobs < 0:
            raise ValueError("`logprobs` must be a positive value.")

1018
1019
        return data

1020
1021
1022
1023
1024
    @model_validator(mode="before")
    @classmethod
    def validate_stream_options(cls, data):
        if data.get("stream_options") and not data.get("stream"):
            raise ValueError(
1025
1026
                "Stream options can only be defined when `stream=True`.")

1027
1028
        return data

1029
1030
1031
1032
1033
1034
1035
1036
    @model_validator(mode="before")
    @classmethod
    def validate_prompt_and_prompt_embeds(cls, data):
        if data.get("prompt") is None and data.get("prompt_embeds") is None:
            raise ValueError(
                "At least one of `prompt` or `prompt_embeds` must be set.")
        return data

Zhuohan Li's avatar
Zhuohan Li committed
1037

1038
class EmbeddingCompletionRequest(OpenAIBaseModel):
1039
1040
    # Ordered by official OpenAI API documentation
    # https://platform.openai.com/docs/api-reference/embeddings
1041
    model: Optional[str] = None
1042
    input: Union[list[int], list[list[int]], str, list[str]]
1043
    encoding_format: Literal["float", "base64"] = "float"
1044
1045
    dimensions: Optional[int] = None
    user: Optional[str] = None
1046
    truncate_prompt_tokens: Optional[Annotated[int, Field(ge=-1)]] = None
1047

1048
    # --8<-- [start:embedding-pooling-params]
1049
    additional_data: Optional[Any] = None
1050
    # --8<-- [end:embedding-pooling-params]
1051

1052
    # --8<-- [start:embedding-extra-params]
1053
1054
1055
1056
1057
1058
    add_special_tokens: bool = Field(
        default=True,
        description=(
            "If true (the default), special tokens (e.g. BOS) will be added to "
            "the prompt."),
    )
1059
1060
1061
1062
1063
    priority: int = Field(
        default=0,
        description=(
            "The priority of the request (lower means earlier handling; "
            "default: 0). Any priority other than 0 will raise an error "
1064
1065
            "if the served model does not use priority scheduling."),
    )
1066

1067
    # --8<-- [end:embedding-extra-params]
1068

1069
    def to_pooling_params(self):
1070
1071
        return PoolingParams(dimensions=self.dimensions,
                             additional_data=self.additional_data)
1072
1073


1074
class EmbeddingChatRequest(OpenAIBaseModel):
1075
    model: Optional[str] = None
1076
    messages: list[ChatCompletionMessageParam]
1077
1078
1079
1080

    encoding_format: Literal["float", "base64"] = "float"
    dimensions: Optional[int] = None
    user: Optional[str] = None
1081
    truncate_prompt_tokens: Optional[Annotated[int, Field(ge=-1)]] = None
1082

1083
    # --8<-- [start:chat-embedding-pooling-params]
1084
    additional_data: Optional[Any] = None
1085
    # --8<-- [end:chat-embedding-pooling-params]
1086

1087
    # --8<-- [start:chat-embedding-extra-params]
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
    add_special_tokens: bool = Field(
        default=False,
        description=(
            "If true, special tokens (e.g. BOS) will be added to the prompt "
            "on top of what is added by the chat template. "
            "For most models, the chat template takes care of adding the "
            "special tokens so this should be set to false (as is the "
            "default)."),
    )
    chat_template: Optional[str] = Field(
        default=None,
        description=(
            "A Jinja template to use for this conversion. "
            "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."),
    )
1105
    chat_template_kwargs: Optional[dict[str, Any]] = Field(
1106
1107
1108
1109
        default=None,
        description=("Additional kwargs to pass to the template renderer. "
                     "Will be accessible by the chat template."),
    )
1110
    mm_processor_kwargs: Optional[dict[str, Any]] = Field(
1111
1112
1113
        default=None,
        description=("Additional kwargs to pass to the HF processor."),
    )
1114
1115
1116
1117
1118
    priority: int = Field(
        default=0,
        description=(
            "The priority of the request (lower means earlier handling; "
            "default: 0). Any priority other than 0 will raise an error "
1119
1120
            "if the served model does not use priority scheduling."),
    )
1121
    # --8<-- [end:chat-embedding-extra-params]
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132

    @model_validator(mode="before")
    @classmethod
    def check_generation_prompt(cls, data):
        if data.get("continue_final_message") and data.get(
                "add_generation_prompt"):
            raise ValueError("Cannot set both `continue_final_message` and "
                             "`add_generation_prompt` to True.")
        return data

    def to_pooling_params(self):
1133
1134
        return PoolingParams(dimensions=self.dimensions,
                             additional_data=self.additional_data)
1135
1136
1137
1138


EmbeddingRequest = Union[EmbeddingCompletionRequest, EmbeddingChatRequest]

1139
1140
1141
1142
PoolingCompletionRequest = EmbeddingCompletionRequest
PoolingChatRequest = EmbeddingChatRequest
PoolingRequest = Union[PoolingCompletionRequest, PoolingChatRequest]

1143

1144
class ScoreRequest(OpenAIBaseModel):
1145
    model: Optional[str] = None
1146
1147
    text_1: Union[list[str], str]
    text_2: Union[list[str], str]
1148
    truncate_prompt_tokens: Optional[Annotated[int, Field(ge=-1)]] = None
1149

1150
    # --8<-- [start:score-pooling-params]
1151
    additional_data: Optional[Any] = None
1152
    # --8<-- [end:score-pooling-params]
1153

1154
    # --8<-- [start:score-extra-params]
1155
1156
1157
1158
1159
    priority: int = Field(
        default=0,
        description=(
            "The priority of the request (lower means earlier handling; "
            "default: 0). Any priority other than 0 will raise an error "
1160
1161
            "if the served model does not use priority scheduling."),
    )
1162

1163
    # --8<-- [end:score-extra-params]
1164

1165
1166
1167
1168
    def to_pooling_params(self):
        return PoolingParams(additional_data=self.additional_data)


1169
class RerankRequest(OpenAIBaseModel):
1170
    model: Optional[str] = None
1171
    query: str
1172
    documents: list[str]
1173
    top_n: int = Field(default_factory=lambda: 0)
1174
    truncate_prompt_tokens: Optional[Annotated[int, Field(ge=-1)]] = None
1175

1176
    # --8<-- [start:rerank-pooling-params]
1177
    additional_data: Optional[Any] = None
1178
    # --8<-- [end:rerank-pooling-params]
1179

1180
    # --8<-- [start:rerank-extra-params]
1181
1182
1183
1184
1185
    priority: int = Field(
        default=0,
        description=(
            "The priority of the request (lower means earlier handling; "
            "default: 0). Any priority other than 0 will raise an error "
1186
1187
            "if the served model does not use priority scheduling."),
    )
1188

1189
    # --8<-- [end:rerank-extra-params]
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212

    def to_pooling_params(self):
        return PoolingParams(additional_data=self.additional_data)


class RerankDocument(BaseModel):
    text: str


class RerankResult(BaseModel):
    index: int
    document: RerankDocument
    relevance_score: float


class RerankUsage(BaseModel):
    total_tokens: int


class RerankResponse(OpenAIBaseModel):
    id: str
    model: str
    usage: RerankUsage
1213
    results: list[RerankResult]
1214
1215


1216
class CompletionLogProbs(OpenAIBaseModel):
1217
1218
1219
1220
    text_offset: list[int] = Field(default_factory=list)
    token_logprobs: list[Optional[float]] = Field(default_factory=list)
    tokens: list[str] = Field(default_factory=list)
    top_logprobs: list[Optional[dict[str,
1221
                                     float]]] = Field(default_factory=list)
Zhuohan Li's avatar
Zhuohan Li committed
1222
1223


1224
class CompletionResponseChoice(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
1225
1226
    index: int
    text: str
1227
    logprobs: Optional[CompletionLogProbs] = None
1228
1229
    finish_reason: Optional[str] = None
    stop_reason: Optional[Union[int, str]] = Field(
1230
1231
1232
1233
1234
1235
        default=None,
        description=(
            "The stop string or token id that caused the completion "
            "to stop, None if the completion finished for some other reason "
            "including encountering the EOS token"),
    )
1236
    prompt_logprobs: Optional[list[Optional[dict[int, Logprob]]]] = None
Zhuohan Li's avatar
Zhuohan Li committed
1237
1238


1239
class CompletionResponse(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
1240
1241
1242
1243
    id: str = Field(default_factory=lambda: f"cmpl-{random_uuid()}")
    object: str = "text_completion"
    created: int = Field(default_factory=lambda: int(time.time()))
    model: str
1244
    choices: list[CompletionResponseChoice]
Zhuohan Li's avatar
Zhuohan Li committed
1245
    usage: UsageInfo
Robert Shaw's avatar
Robert Shaw committed
1246
1247
    kv_transfer_params: Optional[dict[str, Any]] = Field(
        default=None, description="KVTransfer parameters.")
Zhuohan Li's avatar
Zhuohan Li committed
1248
1249


1250
class CompletionResponseStreamChoice(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
1251
1252
    index: int
    text: str
1253
    logprobs: Optional[CompletionLogProbs] = None
1254
1255
    finish_reason: Optional[str] = None
    stop_reason: Optional[Union[int, str]] = Field(
1256
1257
1258
1259
1260
1261
        default=None,
        description=(
            "The stop string or token id that caused the completion "
            "to stop, None if the completion finished for some other reason "
            "including encountering the EOS token"),
    )
Zhuohan Li's avatar
Zhuohan Li committed
1262
1263


1264
class CompletionStreamResponse(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
1265
1266
1267
1268
    id: str = Field(default_factory=lambda: f"cmpl-{random_uuid()}")
    object: str = "text_completion"
    created: int = Field(default_factory=lambda: int(time.time()))
    model: str
1269
    choices: list[CompletionResponseStreamChoice]
1270
    usage: Optional[UsageInfo] = Field(default=None)
1271
1272


1273
class EmbeddingResponseData(OpenAIBaseModel):
1274
1275
    index: int
    object: str = "embedding"
1276
    embedding: Union[list[float], str]
1277
1278


1279
class EmbeddingResponse(OpenAIBaseModel):
1280
    id: str = Field(default_factory=lambda: f"embd-{random_uuid()}")
1281
1282
1283
    object: str = "list"
    created: int = Field(default_factory=lambda: int(time.time()))
    model: str
1284
    data: list[EmbeddingResponseData]
1285
1286
1287
    usage: UsageInfo


1288
1289
1290
class PoolingResponseData(OpenAIBaseModel):
    index: int
    object: str = "pooling"
1291
    data: Union[list[list[float]], list[float], str]
1292
1293
1294
1295
1296
1297
1298


class PoolingResponse(OpenAIBaseModel):
    id: str = Field(default_factory=lambda: f"pool-{random_uuid()}")
    object: str = "list"
    created: int = Field(default_factory=lambda: int(time.time()))
    model: str
1299
    data: list[PoolingResponseData]
1300
1301
1302
    usage: UsageInfo


1303
1304
1305
class ScoreResponseData(OpenAIBaseModel):
    index: int
    object: str = "score"
1306
    score: float
1307
1308
1309
1310
1311
1312
1313


class ScoreResponse(OpenAIBaseModel):
    id: str = Field(default_factory=lambda: f"embd-{random_uuid()}")
    object: str = "list"
    created: int = Field(default_factory=lambda: int(time.time()))
    model: str
1314
    data: list[ScoreResponseData]
1315
1316
1317
    usage: UsageInfo


1318
1319
1320
1321
1322
1323
class ClassificationRequest(OpenAIBaseModel):
    model: Optional[str] = None
    input: Union[list[str], str]
    truncate_prompt_tokens: Optional[int] = None
    user: Optional[str] = None

1324
    # --8<-- [start:classification-pooling-params]
1325
    additional_data: Optional[Any] = None
1326
    # --8<-- [end:classification-pooling-params]
1327

1328
    # --8<-- [start:classification-extra-params]
1329
1330
1331
1332
1333
1334
1335
1336
    priority: int = Field(
        default=0,
        description=(
            "The priority of the request (lower means earlier handling; "
            "default: 0). Any priority other than 0 will raise an error "
            "if the served model does not use priority scheduling."),
    )

1337
    # --8<-- [end:classification-extra-params]
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358

    def to_pooling_params(self):
        return PoolingParams(additional_data=self.additional_data)


class ClassificationData(OpenAIBaseModel):
    index: int
    label: Optional[str]
    probs: list[float]
    num_classes: int


class ClassificationResponse(OpenAIBaseModel):
    id: str = Field(default_factory=lambda: f"classify-{random_uuid()}")
    object: str = "list"
    created: int = Field(default_factory=lambda: int(time.time()))
    model: str
    data: list[ClassificationData]
    usage: UsageInfo


1359
1360
1361
1362
1363
1364
class FunctionCall(OpenAIBaseModel):
    name: str
    arguments: str


class ToolCall(OpenAIBaseModel):
1365
    id: str = Field(default_factory=random_tool_call_id)
1366
1367
1368
1369
    type: Literal["function"] = "function"
    function: FunctionCall


1370
1371
1372
1373
1374
1375
1376
class DeltaFunctionCall(BaseModel):
    name: Optional[str] = None
    arguments: Optional[str] = None


# a tool call delta where everything is optional
class DeltaToolCall(OpenAIBaseModel):
1377
1378
    id: Optional[str] = None
    type: Optional[Literal["function"]] = None
1379
1380
1381
1382
1383
1384
1385
1386
1387
    index: int
    function: Optional[DeltaFunctionCall] = None


class ExtractedToolCallInformation(BaseModel):
    # indicate if tools were called
    tools_called: bool

    # extracted tool calls
1388
    tool_calls: list[ToolCall]
1389
1390
1391
1392
1393
1394

    # content - per OpenAI spec, content AND tool calls can be returned rarely
    # But some models will do this intentionally
    content: Optional[str] = None


1395
class ChatMessage(OpenAIBaseModel):
1396
    role: str
1397
    reasoning_content: Optional[str] = None
1398
    content: Optional[str] = None
1399
    tool_calls: list[ToolCall] = Field(default_factory=list)
1400
1401


1402
1403
1404
class ChatCompletionLogProb(OpenAIBaseModel):
    token: str
    logprob: float = -9999.0
1405
    bytes: Optional[list[int]] = None
1406
1407
1408


class ChatCompletionLogProbsContent(ChatCompletionLogProb):
1409
1410
1411
    # Workaround: redefine fields name cache so that it's not
    # shared with the super class.
    field_names: ClassVar[Optional[set[str]]] = None
1412
    top_logprobs: list[ChatCompletionLogProb] = Field(default_factory=list)
1413
1414
1415


class ChatCompletionLogProbs(OpenAIBaseModel):
1416
    content: Optional[list[ChatCompletionLogProbsContent]] = None
1417
1418


1419
class ChatCompletionResponseChoice(OpenAIBaseModel):
1420
1421
    index: int
    message: ChatMessage
1422
    logprobs: Optional[ChatCompletionLogProbs] = None
1423
1424
1425
    # per OpenAI spec this is the default
    finish_reason: Optional[str] = "stop"
    # not part of the OpenAI spec but included in vLLM for legacy reasons
1426
    stop_reason: Optional[Union[int, str]] = None
1427
1428


1429
class ChatCompletionResponse(OpenAIBaseModel):
1430
    id: str = Field(default_factory=lambda: f"chatcmpl-{random_uuid()}")
1431
    object: Literal["chat.completion"] = "chat.completion"
1432
1433
    created: int = Field(default_factory=lambda: int(time.time()))
    model: str
1434
    choices: list[ChatCompletionResponseChoice]
1435
    usage: UsageInfo
1436
    prompt_logprobs: Optional[list[Optional[dict[int, Logprob]]]] = None
Robert Shaw's avatar
Robert Shaw committed
1437
1438
    kv_transfer_params: Optional[dict[str, Any]] = Field(
        default=None, description="KVTransfer parameters.")
1439
1440


1441
class DeltaMessage(OpenAIBaseModel):
1442
1443
    role: Optional[str] = None
    content: Optional[str] = None
1444
    reasoning_content: Optional[str] = None
1445
    tool_calls: list[DeltaToolCall] = Field(default_factory=list)
1446
1447


1448
class ChatCompletionResponseStreamChoice(OpenAIBaseModel):
1449
1450
    index: int
    delta: DeltaMessage
1451
    logprobs: Optional[ChatCompletionLogProbs] = None
1452
    finish_reason: Optional[str] = None
1453
    stop_reason: Optional[Union[int, str]] = None
1454
1455


1456
class ChatCompletionStreamResponse(OpenAIBaseModel):
1457
    id: str = Field(default_factory=lambda: f"chatcmpl-{random_uuid()}")
1458
    object: Literal["chat.completion.chunk"] = "chat.completion.chunk"
1459
1460
    created: int = Field(default_factory=lambda: int(time.time()))
    model: str
1461
    choices: list[ChatCompletionResponseStreamChoice]
1462
    usage: Optional[UsageInfo] = Field(default=None)
1463
1464


1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
class TranscriptionResponseStreamChoice(OpenAIBaseModel):
    delta: DeltaMessage
    finish_reason: Optional[str] = None
    stop_reason: Optional[Union[int, str]] = None


class TranscriptionStreamResponse(OpenAIBaseModel):
    id: str = Field(default_factory=lambda: f"trsc-{random_uuid()}")
    object: Literal["transcription.chunk"] = "transcription.chunk"
    created: int = Field(default_factory=lambda: int(time.time()))
    model: str
    choices: list[TranscriptionResponseStreamChoice]
    usage: Optional[UsageInfo] = Field(default=None)


1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
class BatchRequestInput(OpenAIBaseModel):
    """
    The per-line object of the batch input file.

    NOTE: Currently only the `/v1/chat/completions` endpoint is supported.
    """

    # A developer-provided per-request id that will be used to match outputs to
    # inputs. Must be unique for each request in a batch.
    custom_id: str

    # The HTTP method to be used for the request. Currently only POST is
    # supported.
    method: str

    # The OpenAI API relative URL to be used for the request. Currently
    # /v1/chat/completions is supported.
    url: str

1499
    # The parameters of the request.
1500
    body: Union[ChatCompletionRequest, EmbeddingRequest, ScoreRequest]
1501

1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
    @field_validator('body', mode='plain')
    @classmethod
    def check_type_for_url(cls, value: Any, info: ValidationInfo):
        # Use url to disambiguate models
        url = info.data['url']
        if url == "/v1/chat/completions":
            return ChatCompletionRequest.model_validate(value)
        if url == "/v1/embeddings":
            return TypeAdapter(EmbeddingRequest).validate_python(value)
        if url == "/v1/score":
            return ScoreRequest.model_validate(value)
        return TypeAdapter(Union[ChatCompletionRequest, EmbeddingRequest,
                                 ScoreRequest]).validate_python(value)

1516

1517
1518
1519
1520
1521
1522
1523
1524
class BatchResponseData(OpenAIBaseModel):
    # HTTP status code of the response.
    status_code: int = 200

    # An unique identifier for the API request.
    request_id: str

    # The body of the response.
1525
1526
    body: Optional[Union[ChatCompletionResponse, EmbeddingResponse,
                         ScoreResponse]] = None
1527
1528


1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
class BatchRequestOutput(OpenAIBaseModel):
    """
    The per-line object of the batch output and error files
    """

    id: str

    # A developer-provided per-request id that will be used to match outputs to
    # inputs.
    custom_id: str

1540
    response: Optional[BatchResponseData]
1541
1542
1543
1544

    # For requests that failed with a non-HTTP error, this will contain more
    # information on the cause of the failure.
    error: Optional[Any]
1545
1546


1547
class TokenizeCompletionRequest(OpenAIBaseModel):
1548
    model: Optional[str] = None
1549
1550
    prompt: str

1551
1552
1553
1554
1555
1556
    add_special_tokens: bool = Field(
        default=True,
        description=(
            "If true (the default), special tokens (e.g. BOS) will be added to "
            "the prompt."),
    )
1557
1558
1559


class TokenizeChatRequest(OpenAIBaseModel):
1560
    model: Optional[str] = None
1561
    messages: list[ChatCompletionMessageParam]
1562

1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
    add_generation_prompt: bool = Field(
        default=True,
        description=
        ("If true, the generation prompt will be added to the chat template. "
         "This is a parameter used by chat template in tokenizer config of the "
         "model."),
    )
    continue_final_message: bool = Field(
        default=False,
        description=
        ("If this is set, the chat will be formatted so that the final "
         "message in the chat is open-ended, without any EOS tokens. The "
         "model will continue this message rather than starting a new one. "
         "This allows you to \"prefill\" part of the model's response for it. "
         "Cannot be used at the same time as `add_generation_prompt`."),
    )
    add_special_tokens: bool = Field(
        default=False,
        description=(
            "If true, special tokens (e.g. BOS) will be added to the prompt "
            "on top of what is added by the chat template. "
            "For most models, the chat template takes care of adding the "
            "special tokens so this should be set to false (as is the "
            "default)."),
    )
    chat_template: Optional[str] = Field(
        default=None,
        description=(
            "A Jinja template to use for this conversion. "
            "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."),
    )
1596
    chat_template_kwargs: Optional[dict[str, Any]] = Field(
1597
1598
1599
1600
        default=None,
        description=("Additional kwargs to pass to the template renderer. "
                     "Will be accessible by the chat template."),
    )
1601
    mm_processor_kwargs: Optional[dict[str, Any]] = Field(
1602
1603
1604
        default=None,
        description=("Additional kwargs to pass to the HF processor."),
    )
1605
1606
1607
1608
    tools: Optional[list[ChatCompletionToolsParam]] = Field(
        default=None,
        description=("A list of tools the model may call."),
    )
1609

1610
1611
1612
1613
1614
1615
1616
1617
1618
    @model_validator(mode="before")
    @classmethod
    def check_generation_prompt(cls, data):
        if data.get("continue_final_message") and data.get(
                "add_generation_prompt"):
            raise ValueError("Cannot set both `continue_final_message` and "
                             "`add_generation_prompt` to True.")
        return data

1619
1620

TokenizeRequest = Union[TokenizeCompletionRequest, TokenizeChatRequest]
1621
1622
1623
1624
1625


class TokenizeResponse(OpenAIBaseModel):
    count: int
    max_model_len: int
1626
    tokens: list[int]
1627
1628
1629


class DetokenizeRequest(OpenAIBaseModel):
1630
    model: Optional[str] = None
1631
    tokens: list[int]
1632
1633
1634
1635


class DetokenizeResponse(OpenAIBaseModel):
    prompt: str
1636
1637


1638
class LoadLoRAAdapterRequest(BaseModel):
1639
1640
1641
1642
    lora_name: str
    lora_path: str


1643
class UnloadLoRAAdapterRequest(BaseModel):
1644
1645
    lora_name: str
    lora_int_id: Optional[int] = Field(default=None)
1646
1647
1648
1649
1650
1651
1652
1653
1654


## Protocols for Audio
AudioResponseFormat: TypeAlias = Literal["json", "text", "srt", "verbose_json",
                                         "vtt"]


class TranscriptionRequest(OpenAIBaseModel):
    # Ordered by official OpenAI API documentation
1655
    # https://platform.openai.com/docs/api-reference/audio/createTranscription
1656
1657
1658
1659
1660
1661
1662

    file: UploadFile
    """
    The audio file object (not file name) to transcribe, in one of these
    formats: flac, mp3, mp4, mpeg, mpga, m4a, ogg, wav, or webm.
    """

1663
    model: Optional[str] = None
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
    """ID of the model to use.
    """

    language: Optional[str] = None
    """The language of the input audio.

    Supplying the input language in
    [ISO-639-1](https://en.wikipedia.org/wiki/List_of_ISO_639-1_codes) format
    will improve accuracy and latency.
    """

    prompt: str = Field(default="")
    """An optional text to guide the model's style or continue a previous audio
    segment.

    The [prompt](https://platform.openai.com/docs/guides/speech-to-text#prompting)
    should match the audio language.
    """

    response_format: AudioResponseFormat = Field(default="json")
    """
    The format of the output, in one of these options: `json`, `text`, `srt`,
    `verbose_json`, or `vtt`.
    """

    ## TODO (varun) : Support if set to 0, certain thresholds are met !!

1691
    timestamp_granularities: list[Literal["word", "segment"]] = Field(
1692
1693
1694
1695
1696
1697
1698
1699
1700
        alias="timestamp_granularities[]", default=[])
    """The timestamp granularities to populate for this transcription.

    `response_format` must be set `verbose_json` to use timestamp granularities.
    Either or both of these options are supported: `word`, or `segment`. Note:
    There is no additional latency for segment timestamps, but generating word
    timestamps incurs additional latency.
    """

1701
    # --8<-- [start:transcription-extra-params]
1702
    stream: Optional[bool] = False
1703
    """Custom field not present in the original OpenAI definition. When set,
1704
    it will enable output to be streamed in a similar fashion as the Chat
1705
    Completion endpoint.
1706
1707
1708
1709
    """
    # Flattened stream option to simplify form data.
    stream_include_usage: Optional[bool] = False
    stream_continuous_usage_stats: Optional[bool] = False
1710
    # --8<-- [end:transcription-extra-params]
1711

1712
    # --8<-- [start:transcription-sampling-params]
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
    temperature: float = Field(default=0.0)
    """The sampling temperature, between 0 and 1.

    Higher values like 0.8 will make the output more random, while lower values
    like 0.2 will make it more focused / deterministic. If set to 0, the model
    will use [log probability](https://en.wikipedia.org/wiki/Log_probability)
    to automatically increase the temperature until certain thresholds are hit.
    """

    top_p: Optional[float] = None
1723
    """Enables nucleus (top-p) sampling, where tokens are selected from the
1724
1725
1726
1727
1728
1729
1730
    smallest possible set whose cumulative probability exceeds `p`.
    """

    top_k: Optional[int] = None
    """Limits sampling to the `k` most probable tokens at each step."""

    min_p: Optional[float] = None
1731
    """Filters out tokens with a probability lower than `min_p`, ensuring a
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
    minimum likelihood threshold during sampling.
    """

    seed: Optional[int] = Field(None, ge=_LONG_INFO.min, le=_LONG_INFO.max)
    """The seed to use for sampling."""

    frequency_penalty: Optional[float] = 0.0
    """The frequency penalty to use for sampling."""

    repetition_penalty: Optional[float] = None
    """The repetition penalty to use for sampling."""

    presence_penalty: Optional[float] = 0.0
    """The presence penalty to use for sampling."""
1746
    # --8<-- [end:transcription-sampling-params]
1747

1748
1749
    # Default sampling parameters for transcription requests.
    _DEFAULT_SAMPLING_PARAMS: dict = {
1750
1751
1752
        "repetition_penalty": 1.0,
        "temperature": 1.0,
        "top_p": 1.0,
1753
        "top_k": 0,
1754
        "min_p": 0.0,
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
    }

    def to_sampling_params(
            self,
            default_max_tokens: int,
            default_sampling_params: Optional[dict] = None) -> SamplingParams:
        # TODO(#9845): remove max_tokens when field is removed from OpenAI API
        max_tokens = default_max_tokens

        if default_sampling_params is None:
            default_sampling_params = {}
1766

1767
1768
1769
1770
        # Default parameters
        if (temperature := self.temperature) is None:
            temperature = default_sampling_params.get(
                "temperature", self._DEFAULT_SAMPLING_PARAMS["temperature"])
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
        if (top_p := self.top_p) is None:
            top_p = default_sampling_params.get(
                "top_p", self._DEFAULT_SAMPLING_PARAMS["top_p"])
        if (top_k := self.top_k) is None:
            top_k = default_sampling_params.get(
                "top_k", self._DEFAULT_SAMPLING_PARAMS["top_k"])
        if (min_p := self.min_p) is None:
            min_p = default_sampling_params.get(
                "min_p", self._DEFAULT_SAMPLING_PARAMS["min_p"])

        if (repetition_penalty := self.repetition_penalty) is None:
            repetition_penalty = default_sampling_params.get(
                "repetition_penalty",
                self._DEFAULT_SAMPLING_PARAMS["repetition_penalty"])
1785
1786

        return SamplingParams.from_optional(temperature=temperature,
1787
                                            max_tokens=max_tokens,
1788
1789
1790
1791
1792
1793
1794
                                            seed=self.seed,
                                            top_p=top_p,
                                            top_k=top_k,
                                            min_p=min_p,
                                            frequency_penalty=self.frequency_penalty,
                                            repetition_penalty=repetition_penalty,
                                            presence_penalty=self.presence_penalty,
1795
1796
1797
1798
1799
1800
                                            output_kind=RequestOutputKind.DELTA
                                            if self.stream \
                                            else RequestOutputKind.FINAL_ONLY)

    @model_validator(mode="before")
    @classmethod
1801
1802
1803
1804
1805
1806
1807
    def validate_transcription_request(cls, data):
        if isinstance(data.get("file"), str):
            raise HTTPException(
                status_code=HTTPStatus.UNPROCESSABLE_ENTITY,
                detail="Expected 'file' to be a file-like object, not 'str'.",
            )

1808
1809
1810
1811
1812
1813
1814
        stream_opts = ["stream_include_usage", "stream_continuous_usage_stats"]
        stream = data.get("stream", False)
        if any(bool(data.get(so, False)) for so in stream_opts) and not stream:
            raise ValueError(
                "Stream options can only be defined when `stream=True`.")

        return data
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871


# Transcription response objects
class TranscriptionResponse(OpenAIBaseModel):
    text: str
    """The transcribed text."""


class TranscriptionWord(OpenAIBaseModel):
    end: float
    """End time of the word in seconds."""

    start: float
    """Start time of the word in seconds."""

    word: str
    """The text content of the word."""


class TranscriptionSegment(OpenAIBaseModel):
    id: int
    """Unique identifier of the segment."""

    avg_logprob: float
    """Average logprob of the segment.

    If the value is lower than -1, consider the logprobs failed.
    """

    compression_ratio: float
    """Compression ratio of the segment.

    If the value is greater than 2.4, consider the compression failed.
    """

    end: float
    """End time of the segment in seconds."""

    no_speech_prob: float
    """Probability of no speech in the segment.

    If the value is higher than 1.0 and the `avg_logprob` is below -1, consider
    this segment silent.
    """

    seek: int
    """Seek offset of the segment."""

    start: float
    """Start time of the segment in seconds."""

    temperature: float
    """Temperature parameter used for generating the segment."""

    text: str
    """Text content of the segment."""

1872
    tokens: list[int]
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
    """Array of token IDs for the text content."""


class TranscriptionResponseVerbose(OpenAIBaseModel):
    duration: str
    """The duration of the input audio."""

    language: str
    """The language of the input audio."""

    text: str
    """The transcribed text."""

1886
    segments: Optional[list[TranscriptionSegment]] = None
1887
1888
    """Segments of the transcribed text and their corresponding details."""

1889
    words: Optional[list[TranscriptionWord]] = None
1890
    """Extracted words and their corresponding timestamps."""