protocol.py 93.7 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
5
# Adapted from
# https://github.com/lm-sys/FastChat/blob/168ccc29d3f7edc50823016105c024fe2282732a/fastchat/protocol/openai_api_protocol.py
6
import json
Zhuohan Li's avatar
Zhuohan Li committed
7
import time
8
from http import HTTPStatus
9
10
from typing import (Annotated, Any, ClassVar, Generic, Literal, Optional,
                    TypeVar, Union)
Zhuohan Li's avatar
Zhuohan Li committed
11

12
import regex as re
13
import torch
14
from fastapi import HTTPException, UploadFile
15
16
17
18
19
20
# yapf: disable
from openai.types.chat.chat_completion_audio import (
    ChatCompletionAudio as OpenAIChatCompletionAudio)
from openai.types.chat.chat_completion_message import (
    Annotation as OpenAIAnnotation)
# yapf: enable
21
22
23
24
25
26
27
28
29
30
31
32
33
from openai.types.responses import (
    ResponseCodeInterpreterCallCodeDeltaEvent,
    ResponseCodeInterpreterCallCodeDoneEvent,
    ResponseCodeInterpreterCallCompletedEvent,
    ResponseCodeInterpreterCallInProgressEvent,
    ResponseCodeInterpreterCallInterpretingEvent, ResponseCompletedEvent,
    ResponseContentPartAddedEvent, ResponseContentPartDoneEvent,
    ResponseCreatedEvent, ResponseFunctionToolCall, ResponseInProgressEvent,
    ResponseInputItemParam, ResponseOutputItem, ResponseOutputItemAddedEvent,
    ResponseOutputItemDoneEvent, ResponsePrompt, ResponseReasoningItem,
    ResponseReasoningTextDeltaEvent, ResponseReasoningTextDoneEvent,
    ResponseStatus, ResponseWebSearchCallCompletedEvent,
    ResponseWebSearchCallInProgressEvent, ResponseWebSearchCallSearchingEvent)
34
35
36
37
38
39
40
41

# Backward compatibility for OpenAI client versions
try:  # For older openai versions (< 1.100.0)
    from openai.types.responses import ResponseTextConfig
except ImportError:  # For newer openai versions (>= 1.100.0)
    from openai.types.responses import (ResponseFormatTextConfig as
                                        ResponseTextConfig)

42
from openai.types.responses.response import IncompleteDetails, ToolChoice
43
44
from openai.types.responses.tool import Tool
from openai.types.shared import Metadata, Reasoning
45
46
from pydantic import (BaseModel, ConfigDict, Field, TypeAdapter,
                      ValidationInfo, field_validator, model_validator)
47
from typing_extensions import TypeAlias
Zhuohan Li's avatar
Zhuohan Li committed
48

49
from vllm import envs
50
from vllm.entrypoints.chat_utils import (ChatCompletionMessageParam,
51
                                         make_tool_call_id)
52
53
from vllm.entrypoints.score_utils import (ScoreContentPartParam,
                                          ScoreMultiModalParam)
54
from vllm.logger import init_logger
55
from vllm.logprobs import Logprob
56
from vllm.pooling_params import PoolingParams
57
58
from vllm.sampling_params import (BeamSearchParams, RequestOutputKind,
                                  SamplingParams, StructuredOutputsParams)
59
from vllm.utils import random_uuid, resolve_obj_by_qualname
60

61
62
logger = init_logger(__name__)

63
_LONG_INFO = torch.iinfo(torch.long)
64

Zhuohan Li's avatar
Zhuohan Li committed
65

66
class OpenAIBaseModel(BaseModel):
67
68
69
    # OpenAI API does allow extra fields
    model_config = ConfigDict(extra="allow")

70
    # Cache class field names
71
    field_names: ClassVar[Optional[set[str]]] = None
72

73
    @model_validator(mode="wrap")
74
    @classmethod
75
76
77
78
    def __log_extra_fields__(cls, data, handler):
        result = handler(data)
        if not isinstance(data, dict):
            return result
79
80
        field_names = cls.field_names
        if field_names is None:
81
82
83
84
            # 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)
85
                if alias := getattr(field, "alias", None):
86
87
88
89
90
91
92
93
                    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",
94
95
                data.keys() - field_names,
            )
96
        return result
97
98


99
class ErrorInfo(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
100
101
102
    message: str
    type: str
    param: Optional[str] = None
103
    code: int
Zhuohan Li's avatar
Zhuohan Li committed
104
105


106
107
108
109
class ErrorResponse(OpenAIBaseModel):
    error: ErrorInfo


110
class ModelPermission(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
111
112
113
114
115
116
117
118
119
120
121
    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
122
    is_blocking: bool = False
Zhuohan Li's avatar
Zhuohan Li committed
123
124


125
class ModelCard(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
126
127
128
    id: str
    object: str = "model"
    created: int = Field(default_factory=lambda: int(time.time()))
Woosuk Kwon's avatar
Woosuk Kwon committed
129
    owned_by: str = "vllm"
Zhuohan Li's avatar
Zhuohan Li committed
130
131
    root: Optional[str] = None
    parent: Optional[str] = None
132
    max_model_len: Optional[int] = None
133
    permission: list[ModelPermission] = Field(default_factory=list)
Zhuohan Li's avatar
Zhuohan Li committed
134
135


136
class ModelList(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
137
    object: str = "list"
138
    data: list[ModelCard] = Field(default_factory=list)
Zhuohan Li's avatar
Zhuohan Li committed
139
140


141
142
143
144
class PromptTokenUsageInfo(OpenAIBaseModel):
    cached_tokens: Optional[int] = None


145
class UsageInfo(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
146
147
148
    prompt_tokens: int = 0
    total_tokens: int = 0
    completion_tokens: Optional[int] = 0
149
    prompt_tokens_details: Optional[PromptTokenUsageInfo] = None
Zhuohan Li's avatar
Zhuohan Li committed
150
151


152
153
154
155
156
class RequestResponseMetadata(BaseModel):
    request_id: str
    final_usage_info: Optional[UsageInfo] = None


157
158
159
160
161
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
162
    json_schema: Optional[dict[str, Any]] = Field(default=None, alias='schema')
163
164
165
    strict: Optional[bool] = None


166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
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]


181
class ResponseFormat(OpenAIBaseModel):
182
    # type must be "json_schema", "json_object", or "text"
183
184
    type: Literal["text", "json_object", "json_schema"]
    json_schema: Optional[JsonSchemaResponseFormat] = None
185
186


187
188
189
AnyResponseFormat = Union[ResponseFormat, StructuralTagResponseFormat]


190
class StreamOptions(OpenAIBaseModel):
191
    include_usage: Optional[bool] = True
192
    continuous_usage_stats: Optional[bool] = False
193
194


195
196
197
class FunctionDefinition(OpenAIBaseModel):
    name: str
    description: Optional[str] = None
198
    parameters: Optional[dict[str, Any]] = None
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214


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


class ChatCompletionNamedFunction(OpenAIBaseModel):
    name: str


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


215
216
# extra="forbid" is a workaround to have kwargs as a field,
# see https://github.com/pydantic/pydantic/issues/3125
217
218
class LogitsProcessorConstructor(BaseModel):
    qualname: str
219
220
    args: Optional[list[Any]] = None
    kwargs: Optional[dict[str, Any]] = None
221

222
223
    model_config = ConfigDict(extra="forbid")

224

225
LogitsProcessors = list[Union[str, LogitsProcessorConstructor]]
226
227
228


def get_logits_processors(processors: Optional[LogitsProcessors],
229
                          pattern: Optional[str]) -> Optional[list[Any]]:
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
    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 "
254
            "server. See --logits-processor-pattern engine argument "
255
256
257
258
            "for more information.")
    return None


259
ResponseInputOutputItem: TypeAlias = Union[ResponseInputItemParam,
260
                                           ResponseReasoningItem,
261
262
                                           ResponseFunctionToolCall]

263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
StreamingResponsesResponse: TypeAlias = Union[
    ResponseCreatedEvent,
    ResponseInProgressEvent,
    ResponseCompletedEvent,
    ResponseOutputItemAddedEvent,
    ResponseOutputItemDoneEvent,
    ResponseContentPartAddedEvent,
    ResponseContentPartDoneEvent,
    ResponseReasoningTextDeltaEvent,
    ResponseReasoningTextDoneEvent,
    ResponseCodeInterpreterCallInProgressEvent,
    ResponseCodeInterpreterCallCodeDeltaEvent,
    ResponseWebSearchCallInProgressEvent,
    ResponseWebSearchCallSearchingEvent,
    ResponseWebSearchCallCompletedEvent,
    ResponseCodeInterpreterCallCodeDoneEvent,
    ResponseCodeInterpreterCallInterpretingEvent,
    ResponseCodeInterpreterCallCompletedEvent,
]

283

284
285
286
287
288
289
290
291
292
293
294
295
296
297
class ResponsesRequest(OpenAIBaseModel):
    # Ordered by official OpenAI API documentation
    # https://platform.openai.com/docs/api-reference/responses/create
    background: Optional[bool] = False
    include: Optional[list[
        Literal[
            "code_interpreter_call.outputs",
            "computer_call_output.output.image_url",
            "file_search_call.results",
            "message.input_image.image_url",
            "message.output_text.logprobs",
            "reasoning.encrypted_content",
        ],
    ]] = None
298
    input: Union[str, list[ResponseInputOutputItem]]
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
    instructions: Optional[str] = None
    max_output_tokens: Optional[int] = None
    max_tool_calls: Optional[int] = None
    metadata: Optional[Metadata] = None
    model: Optional[str] = None
    parallel_tool_calls: Optional[bool] = True
    previous_response_id: Optional[str] = None
    prompt: Optional[ResponsePrompt] = None
    reasoning: Optional[Reasoning] = None
    service_tier: Literal["auto", "default", "flex", "scale",
                          "priority"] = "auto"
    store: Optional[bool] = True
    stream: Optional[bool] = False
    temperature: Optional[float] = None
    text: Optional[ResponseTextConfig] = None
    tool_choice: ToolChoice = "auto"
    tools: list[Tool] = Field(default_factory=list)
    top_logprobs: Optional[int] = 0
    top_p: Optional[float] = None
    truncation: Optional[Literal["auto", "disabled"]] = "disabled"
    user: Optional[str] = None

    # --8<-- [start:responses-extra-params]
    request_id: str = Field(
        default_factory=lambda: f"resp_{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 "
            "through out the inference process and return in response."),
    )
    mm_processor_kwargs: Optional[dict[str, Any]] = Field(
        default=None,
        description=("Additional kwargs to pass to the HF processor."),
    )
    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."),
    )
340
341
342
343
344
345
346
347
348
    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."))
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
    # --8<-- [end:responses-extra-params]

    _DEFAULT_SAMPLING_PARAMS = {
        "temperature": 1.0,
        "top_p": 1.0,
    }

    def to_sampling_params(
        self,
        default_max_tokens: int,
        default_sampling_params: Optional[dict] = None,
    ) -> SamplingParams:
        if self.max_output_tokens is None:
            max_tokens = default_max_tokens
        else:
            max_tokens = min(self.max_output_tokens, default_max_tokens)

        default_sampling_params = default_sampling_params or {}
        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"])
373
        stop_token_ids = default_sampling_params.get("stop_token_ids")
374
375

        # Structured output
376
        structured_outputs = None
377
378
        if self.text is not None and self.text.format is not None:
            response_format = self.text.format
379
380
381
            if (response_format.type == "json_schema"
                    and response_format.schema_ is not None):
                structured_outputs = StructuredOutputsParams(
382
383
384
385
386
387
388
389
390
                    json=response_format.schema_)
            elif response_format.type == "json_object":
                raise NotImplementedError("json_object is not supported")

        # TODO: add more parameters
        return SamplingParams.from_optional(
            temperature=temperature,
            top_p=top_p,
            max_tokens=max_tokens,
391
392
            logprobs=self.top_logprobs
            if self.is_include_output_logprobs() else None,
393
            stop_token_ids=stop_token_ids,
394
395
            output_kind=(RequestOutputKind.DELTA
                         if self.stream else RequestOutputKind.FINAL_ONLY),
396
            structured_outputs=structured_outputs,
397
398
        )

399
400
401
402
403
404
405
406
    def is_include_output_logprobs(self) -> bool:
        """Check if the request includes output logprobs."""
        if self.include is None:
            return False
        return isinstance(
            self.include,
            list) and "message.output_text.logprobs" in self.include

407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
    @model_validator(mode="before")
    def validate_background(cls, data):
        if not data.get("background"):
            return data
        if not data.get("store", True):
            raise ValueError(
                "background can only be used when `store` is true")
        return data

    @model_validator(mode="before")
    def validate_prompt(cls, data):
        if data.get("prompt") is not None:
            raise ValueError("prompt template is not supported")
        return data

422
423
424
425
426
427
428
429
430
431
432
433
434
    @model_validator(mode="before")
    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

435

436
class ChatCompletionRequest(OpenAIBaseModel):
437
438
    # Ordered by official OpenAI API documentation
    # https://platform.openai.com/docs/api-reference/chat/create
439
    messages: list[ChatCompletionMessageParam]
440
    model: Optional[str] = None
441
    frequency_penalty: Optional[float] = 0.0
442
    logit_bias: Optional[dict[str, float]] = None
443
    logprobs: Optional[bool] = False
444
    top_logprobs: Optional[int] = 0
445
446
447
448
449
    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
450
451
    n: Optional[int] = 1
    presence_penalty: Optional[float] = 0.0
452
    response_format: Optional[AnyResponseFormat] = None
453
    seed: Optional[int] = Field(None, ge=_LONG_INFO.min, le=_LONG_INFO.max)
454
    stop: Optional[Union[str, list[str]]] = []
Zhuohan Li's avatar
Zhuohan Li committed
455
    stream: Optional[bool] = False
456
    stream_options: Optional[StreamOptions] = None
457
458
    temperature: Optional[float] = None
    top_p: Optional[float] = None
459
    tools: Optional[list[ChatCompletionToolsParam]] = None
460
461
462
463
464
465
    tool_choice: Optional[Union[
        Literal["none"],
        Literal["auto"],
        Literal["required"],
        ChatCompletionNamedToolChoiceParam,
    ]] = "none"
466
467
    reasoning_effort: Optional[Literal["low", "medium", "high"]] = None
    include_reasoning: bool = True
468

469
    # NOTE this will be ignored by vLLM -- the model determines the behavior
470
    parallel_tool_calls: Optional[bool] = False
Zhuohan Li's avatar
Zhuohan Li committed
471
    user: Optional[str] = None
472

473
    # --8<-- [start:chat-completion-sampling-params]
474
    best_of: Optional[int] = None
475
    use_beam_search: bool = False
476
477
478
    top_k: Optional[int] = None
    min_p: Optional[float] = None
    repetition_penalty: Optional[float] = None
479
    length_penalty: float = 1.0
480
    stop_token_ids: Optional[list[int]] = []
481
482
483
484
485
    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
486
    truncate_prompt_tokens: Optional[Annotated[int, Field(ge=-1)]] = None
487
    prompt_logprobs: Optional[int] = None
488
    allowed_token_ids: Optional[list[int]] = None
489
    bad_words: list[str] = Field(default_factory=list)
490
    # --8<-- [end:chat-completion-sampling-params]
491

492
    # --8<-- [start:chat-completion-extra-params]
493
    echo: bool = Field(
494
495
496
497
498
        default=False,
        description=(
            "If true, the new message will be prepended with the last message "
            "if they belong to the same role."),
    )
499
    add_generation_prompt: bool = Field(
500
501
502
503
504
505
        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."),
    )
506
507
508
509
510
511
512
513
514
    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`."),
    )
515
    add_special_tokens: bool = Field(
516
517
518
519
520
        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 "
521
            "special tokens so this should be set to false (as is the "
522
523
            "default)."),
    )
524
    documents: Optional[list[dict[str, str]]] = Field(
525
526
527
528
529
530
531
532
533
534
535
536
        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. "
537
538
539
            "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."),
540
    )
541
    chat_template_kwargs: Optional[dict[str, Any]] = Field(
542
        default=None,
543
544
545
        description=(
            "Additional keyword args to pass to the template renderer. "
            "Will be accessible by the chat template."),
546
    )
547
    mm_processor_kwargs: Optional[dict[str, Any]] = Field(
548
549
550
        default=None,
        description=("Additional kwargs to pass to the HF processor."),
    )
551
    structured_outputs: Optional[StructuredOutputsParams] = Field(
552
        default=None,
553
        description="Additional kwargs for structured outputs",
554
    )
555
556
557
558
559
    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 "
560
561
            "if the served model does not use priority scheduling."),
    )
562
563
564
565
566
    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 "
567
568
            "through out the inference process and return in response."),
    )
569
570
571
572
573
574
575
576
577
578
579
    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'}}."))
580
581
582
583
584
585
    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."))
586
587
588
589
590
591
592
593
    return_token_ids: Optional[bool] = Field(
        default=None,
        description=(
            "If specified, the result will include token IDs alongside the "
            "generated text. In streaming mode, prompt_token_ids is included "
            "only in the first chunk, and token_ids contains the delta tokens "
            "for each chunk. This is useful for debugging or when you "
            "need to map generated text back to input tokens."))
594
595
596
597
598
599
600
601
602
    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
603
604
605
    kv_transfer_params: Optional[dict[str, Any]] = Field(
        default=None,
        description="KVTransfer parameters used for disaggregated serving.")
606

607
608
609
610
611
612
    vllm_xargs: Optional[dict[str, Union[str, int, float]]] = Field(
        default=None,
        description=("Additional request parameters with string or "
                     "numeric values, used by custom extensions."),
    )

613
    # --8<-- [end:chat-completion-extra-params]
Zhuohan Li's avatar
Zhuohan Li committed
614

615
616
617
618
619
    # Default sampling parameters for chat completion requests
    _DEFAULT_SAMPLING_PARAMS: dict = {
        "repetition_penalty": 1.0,
        "temperature": 1.0,
        "top_p": 1.0,
620
        "top_k": 0,
621
622
623
624
        "min_p": 0.0,
    }

    def to_beam_search_params(
625
626
            self, max_tokens: int,
            default_sampling_params: dict) -> BeamSearchParams:
627
628

        n = self.n if self.n is not None else 1
629
630
631
        if (temperature := self.temperature) is None:
            temperature = default_sampling_params.get(
                "temperature", self._DEFAULT_SAMPLING_PARAMS["temperature"])
632
633
634
635
636
637

        return BeamSearchParams(
            beam_width=n,
            max_tokens=max_tokens,
            ignore_eos=self.ignore_eos,
            temperature=temperature,
638
            length_penalty=self.length_penalty,
639
640
            include_stop_str_in_output=self.include_stop_str_in_output,
        )
641

642
    def to_sampling_params(
643
        self,
644
        max_tokens: int,
645
        logits_processor_pattern: Optional[str],
646
        default_sampling_params: dict,
647
    ) -> SamplingParams:
648

649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
        # 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"])

668
669
670
671
        prompt_logprobs = self.prompt_logprobs
        if prompt_logprobs is None and self.echo:
            prompt_logprobs = self.top_logprobs

672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
        response_format = self.response_format
        json_schema_from_tool = self._get_json_schema_from_tool()
        if response_format is not None or json_schema_from_tool is not None:
            # If structured outputs wasn't already enabled,
            # we must enable it for these features to work
            if self.structured_outputs is None:
                self.structured_outputs = StructuredOutputsParams()

            # Set structured output params for response format
            if response_format is not None:
                if response_format.type == "json_object":
                    self.structured_outputs.json_object = True
                elif response_format.type == "json_schema":
                    json_schema = response_format.json_schema
                    assert json_schema is not None
                    self.structured_outputs.json = json_schema.json_schema
                elif response_format.type == "structural_tag":
                    structural_tag = response_format
                    assert structural_tag is not None and isinstance(
                        structural_tag, StructuralTagResponseFormat)
                    s_tag_obj = structural_tag.model_dump(by_alias=True)
                    self.structured_outputs.structural_tag = json.dumps(
                        s_tag_obj)

            # Set structured output params for tool calling
            if json_schema_from_tool is not None:
                self.structured_outputs.json = json_schema_from_tool
699

700
701
702
703
        extra_args: dict[str, Any] = self.vllm_xargs if self.vllm_xargs else {}
        if self.kv_transfer_params:
            # Pass in kv_transfer_params via extra_args
            extra_args["kv_transfer_params"] = self.kv_transfer_params
704
        return SamplingParams.from_optional(
705
            n=self.n,
706
            best_of=self.best_of,
707
708
            presence_penalty=self.presence_penalty,
            frequency_penalty=self.frequency_penalty,
709
710
711
712
713
            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
714
            seed=self.seed,
715
716
            stop=self.stop,
            stop_token_ids=self.stop_token_ids,
717
            logprobs=self.top_logprobs if self.logprobs else None,
718
            prompt_logprobs=prompt_logprobs,
719
            ignore_eos=self.ignore_eos,
720
            max_tokens=max_tokens,
721
            min_tokens=self.min_tokens,
722
723
            skip_special_tokens=self.skip_special_tokens,
            spaces_between_special_tokens=self.spaces_between_special_tokens,
724
725
            logits_processors=get_logits_processors(self.logits_processors,
                                                    logits_processor_pattern),
726
            include_stop_str_in_output=self.include_stop_str_in_output,
727
            truncate_prompt_tokens=self.truncate_prompt_tokens,
728
729
            output_kind=RequestOutputKind.DELTA if self.stream \
                else RequestOutputKind.FINAL_ONLY,
730
            structured_outputs=self.structured_outputs,
Robert Shaw's avatar
Robert Shaw committed
731
            logit_bias=self.logit_bias,
732
            bad_words=self.bad_words,
733
            allowed_token_ids=self.allowed_token_ids,
734
735
            extra_args=extra_args or None,
        )
736

737
    def _get_json_schema_from_tool(self) -> Optional[Union[str, dict]]:
738
739
740
741
742
743
744
745
746
747
748
749
750
751
        # 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

752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
        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"]
                }

777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
            def get_tool_schema_defs(
                    tools: list[ChatCompletionToolsParam]) -> dict:
                all_defs = dict[str, dict[str, Any]]()
                for tool in tools:
                    if tool.function.parameters is None:
                        continue
                    defs = tool.function.parameters.pop("$defs", {})
                    for def_name, def_schema in defs.items():
                        if def_name in all_defs and all_defs[
                                def_name] != def_schema:
                            raise ValueError(
                                f"Tool definition '{def_name}' has "
                                "multiple schemas, which is not "
                                "supported.")
                        else:
                            all_defs[def_name] = def_schema
                return all_defs

795
796
797
798
799
800
801
802
            json_schema = {
                "type": "array",
                "minItems": 1,
                "items": {
                    "type": "object",
                    "anyOf": [get_tool_schema(tool) for tool in self.tools]
                }
            }
803
804
805
            json_schema_defs = get_tool_schema_defs(self.tools)
            if json_schema_defs:
                json_schema["$defs"] = json_schema_defs
806
807
            return json_schema

808
        return None
809

810
    @model_validator(mode="before")
811
    @classmethod
812
813
    def validate_stream_options(cls, data):
        if data.get("stream_options") and not data.get("stream"):
814
            raise ValueError(
815
816
817
818
819
820
821
822
                "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:
823
824
            if data.get("stream") and (prompt_logprobs > 0
                                       or prompt_logprobs == -1):
825
826
827
                raise ValueError(
                    "`prompt_logprobs` are not available when `stream=True`.")

828
829
830
831
832
833
            if prompt_logprobs < 0 and prompt_logprobs != -1:
                raise ValueError(
                    "`prompt_logprobs` must be a positive value or -1.")
            if prompt_logprobs == -1 and not envs.VLLM_USE_V1:
                raise ValueError("`prompt_logprobs=-1` is only supported with "
                                 "vLLM engine V1.")
834
835
836
837
        if (top_logprobs := data.get("top_logprobs")) is not None:
            if top_logprobs < 0:
                raise ValueError("`top_logprobs` must be a positive value.")

838
            if top_logprobs > 0 and not data.get("logprobs"):
839
840
841
842
843
                raise ValueError(
                    "when using `top_logprobs`, `logprobs` must be set to true."
                )

        return data
844

845
846
    @model_validator(mode="before")
    @classmethod
847
    def check_structured_outputs_count(cls, data):
848
849
850
        if isinstance(data, ValueError):
            raise data

851
852
853
854
855
856
857
858
859
        if "structured_outputs" not in data:
            return data

        structured_outputs_kwargs = data['structured_outputs']
        count = sum(
            structured_outputs_kwargs.get(k) is not None
            for k in ("json", "regex", "choice"))
        # you can only use one kind of constraints for structured outputs
        if count > 1:
860
            raise ValueError(
861
862
863
864
                "You can only use one kind of constraints for structured "
                "outputs ('json', 'regex' or 'choice').")
        # you can only either use structured outputs or tools, not both
        if count > 1 and data.get("tool_choice", "none") not in (
865
866
867
868
                "none",
                "auto",
                "required",
        ):
869
            raise ValueError(
870
871
                "You can only either use constraints for structured outputs "
                "or tools, not both.")
872
873
874
875
        return data

    @model_validator(mode="before")
    @classmethod
876
877
878
879
    def check_tool_usage(cls, data):

        # if "tool_choice" is not specified but tools are provided,
        # default to "auto" tool_choice
880
        if "tool_choice" not in data and data.get("tools"):
881
882
            data["tool_choice"] = "auto"

883
        # if "tool_choice" is "none" -- no validation is needed for tools
884
885
886
        if "tool_choice" in data and data["tool_choice"] == "none":
            return data

887
        # if "tool_choice" is specified -- validation
888
        if "tool_choice" in data and data["tool_choice"] is not None:
889
890

            # ensure that if "tool choice" is specified, tools are present
891
892
893
            if "tools" not in data or data["tools"] is None:
                raise ValueError(
                    "When using `tool_choice`, `tools` must be set.")
894
895

            # make sure that tool choice is either a named tool
896
897
898
899
            # OR that it's set to "auto" or "required"
            if data["tool_choice"] not in [
                    "auto", "required"
            ] and not isinstance(data["tool_choice"], dict):
900
                raise ValueError(
901
902
903
904
                    f'Invalid value for `tool_choice`: {data["tool_choice"]}! '\
                    'Only named tools, "none", "auto" or "required" '\
                    'are supported.'
                )
905

906
907
908
909
910
911
912
913
914
            # if tool_choice is "required" but the "tools" list is empty,
            # override the data to behave like "none" to align with
            # OpenAI’s behavior.
            if data["tool_choice"] == "required" and isinstance(
                    data["tools"], list) and len(data["tools"]) == 0:
                data["tool_choice"] = "none"
                del data["tools"]
                return data

915
916
            # ensure that if "tool_choice" is specified as an object,
            # it matches a valid tool
917
918
            correct_usage_message = 'Correct usage: `{"type": "function",' \
                ' "function": {"name": "my_function"}}`'
919
920
            if isinstance(data["tool_choice"], dict):
                valid_tool = False
921
922
                function = data["tool_choice"].get("function")
                if not isinstance(function, dict):
923
                    raise ValueError(
924
925
926
927
928
929
930
931
                        f"Invalid value for `function`: `{function}` in "
                        f"`tool_choice`! {correct_usage_message}")
                if "name" not in function:
                    raise ValueError(f"Expected field `name` in `function` in "
                                     f"`tool_choice`! {correct_usage_message}")
                function_name = function["name"]
                if not isinstance(function_name,
                                  str) or len(function_name) == 0:
932
                    raise ValueError(
933
934
                        f"Invalid `name` in `function`: `{function_name}`"
                        f" in `tool_choice`! {correct_usage_message}")
935
                for tool in data["tools"]:
936
                    if tool["function"]["name"] == function_name:
937
938
939
940
941
942
                        valid_tool = True
                        break
                if not valid_tool:
                    raise ValueError(
                        "The tool specified in `tool_choice` does not match any"
                        " of the specified `tools`")
943
944
        return data

945
946
947
948
949
950
951
952
953
    @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

954
955
956
957
958
959
960
961
962
963
964
965
966
967
    @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
968

969
class CompletionRequest(OpenAIBaseModel):
970
971
    # Ordered by official OpenAI API documentation
    # https://platform.openai.com/docs/api-reference/completions/create
972
    model: Optional[str] = None
973
974
    prompt: Optional[Union[list[int], list[list[int]], str, list[str]]] = None
    prompt_embeds: Optional[Union[bytes, list[bytes]]] = None
975
    best_of: Optional[int] = None
Zhuohan Li's avatar
Zhuohan Li committed
976
977
    echo: Optional[bool] = False
    frequency_penalty: Optional[float] = 0.0
978
    logit_bias: Optional[dict[str, float]] = None
979
980
    logprobs: Optional[int] = None
    max_tokens: Optional[int] = 16
981
    n: int = 1
982
    presence_penalty: Optional[float] = 0.0
983
    seed: Optional[int] = Field(None, ge=_LONG_INFO.min, le=_LONG_INFO.max)
984
    stop: Optional[Union[str, list[str]]] = []
985
    stream: Optional[bool] = False
986
    stream_options: Optional[StreamOptions] = None
987
    suffix: Optional[str] = None
988
989
    temperature: Optional[float] = None
    top_p: Optional[float] = None
Zhuohan Li's avatar
Zhuohan Li committed
990
    user: Optional[str] = None
991

992
    # --8<-- [start:completion-sampling-params]
993
    use_beam_search: bool = False
994
995
996
    top_k: Optional[int] = None
    min_p: Optional[float] = None
    repetition_penalty: Optional[float] = None
997
    length_penalty: float = 1.0
998
    stop_token_ids: Optional[list[int]] = []
999
1000
1001
1002
1003
    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
1004
    truncate_prompt_tokens: Optional[Annotated[int, Field(ge=-1)]] = None
1005
    allowed_token_ids: Optional[list[int]] = None
1006
    prompt_logprobs: Optional[int] = None
1007
    # --8<-- [end:completion-sampling-params]
1008

1009
    # --8<-- [start:completion-extra-params]
1010
1011
    add_special_tokens: bool = Field(
        default=True,
1012
        description=(
1013
1014
            "If true (the default), special tokens (e.g. BOS) will be added to "
            "the prompt."),
1015
    )
1016
    response_format: Optional[AnyResponseFormat] = Field(
1017
        default=None,
1018
1019
1020
1021
1022
        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."
        ),
1023
    )
1024
    structured_outputs: Optional[StructuredOutputsParams] = Field(
1025
        default=None,
1026
        description="Additional kwargs for structured outputs",
1027
    )
1028
1029
1030
1031
1032
    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 "
1033
1034
            "if the served model does not use priority scheduling."),
    )
1035
1036
1037
1038
1039
1040
1041
    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 "
            "through out the inference process and return in response."),
    )
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
    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'}}."))
1053

1054
1055
1056
1057
1058
1059
    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."))
1060
1061
1062
1063
1064
1065
1066
1067
    return_token_ids: Optional[bool] = Field(
        default=None,
        description=(
            "If specified, the result will include token IDs alongside the "
            "generated text. In streaming mode, prompt_token_ids is included "
            "only in the first chunk, and token_ids contains the delta tokens "
            "for each chunk. This is useful for debugging or when you "
            "need to map generated text back to input tokens."))
1068

1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
    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
1079
1080
1081
1082
    kv_transfer_params: Optional[dict[str, Any]] = Field(
        default=None,
        description="KVTransfer parameters used for disaggregated serving.")

1083
1084
1085
1086
1087
1088
    vllm_xargs: Optional[dict[str, Union[str, int, float]]] = Field(
        default=None,
        description=("Additional request parameters with string or "
                     "numeric values, used by custom extensions."),
    )

1089
    # --8<-- [end:completion-extra-params]
Zhuohan Li's avatar
Zhuohan Li committed
1090

1091
1092
1093
1094
1095
    # Default sampling parameters for completion requests
    _DEFAULT_SAMPLING_PARAMS: dict = {
        "repetition_penalty": 1.0,
        "temperature": 1.0,
        "top_p": 1.0,
1096
        "top_k": 0,
1097
1098
1099
1100
        "min_p": 0.0,
    }

    def to_beam_search_params(
1101
1102
1103
        self,
        max_tokens: int,
        default_sampling_params: Optional[dict] = None,
1104
    ) -> BeamSearchParams:
1105

1106
1107
        if default_sampling_params is None:
            default_sampling_params = {}
1108
        n = self.n if self.n is not None else 1
1109
1110
1111

        if (temperature := self.temperature) is None:
            temperature = default_sampling_params.get("temperature", 1.0)
1112
1113
1114
1115
1116
1117

        return BeamSearchParams(
            beam_width=n,
            max_tokens=max_tokens,
            ignore_eos=self.ignore_eos,
            temperature=temperature,
1118
            length_penalty=self.length_penalty,
1119
1120
            include_stop_str_in_output=self.include_stop_str_in_output,
        )
1121

1122
    def to_sampling_params(
1123
        self,
1124
        max_tokens: int,
1125
1126
1127
        logits_processor_pattern: Optional[str],
        default_sampling_params: Optional[dict] = None,
    ) -> SamplingParams:
1128

1129
1130
        if default_sampling_params is None:
            default_sampling_params = {}
1131

1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
        # 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"])

1151
1152
1153
1154
        prompt_logprobs = self.prompt_logprobs
        if prompt_logprobs is None and self.echo:
            prompt_logprobs = self.logprobs

1155
1156
        echo_without_generation = self.echo and self.max_tokens == 0

1157
1158
        if (self.structured_outputs is not None
                and self.response_format is not None
1159
                and self.response_format.type == "json_object"):
1160
            self.structured_outputs.json_object = True
1161

1162
1163
1164
1165
        extra_args: dict[str, Any] = self.vllm_xargs if self.vllm_xargs else {}
        if self.kv_transfer_params:
            # Pass in kv_transfer_params via extra_args
            extra_args["kv_transfer_params"] = self.kv_transfer_params
1166
        return SamplingParams.from_optional(
1167
            n=self.n,
1168
            best_of=self.best_of,
1169
1170
            presence_penalty=self.presence_penalty,
            frequency_penalty=self.frequency_penalty,
1171
1172
1173
1174
1175
            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
1176
            seed=self.seed,
1177
1178
            stop=self.stop,
            stop_token_ids=self.stop_token_ids,
1179
            logprobs=self.logprobs,
1180
            ignore_eos=self.ignore_eos,
1181
            max_tokens=max_tokens if not echo_without_generation else 1,
1182
            min_tokens=self.min_tokens,
1183
            prompt_logprobs=prompt_logprobs,
1184
            skip_special_tokens=self.skip_special_tokens,
1185
            spaces_between_special_tokens=self.spaces_between_special_tokens,
1186
            include_stop_str_in_output=self.include_stop_str_in_output,
1187
1188
            logits_processors=get_logits_processors(self.logits_processors,
                                                    logits_processor_pattern),
1189
            truncate_prompt_tokens=self.truncate_prompt_tokens,
1190
1191
            output_kind=RequestOutputKind.DELTA if self.stream \
                else RequestOutputKind.FINAL_ONLY,
1192
            structured_outputs=self.structured_outputs,
1193
            logit_bias=self.logit_bias,
Robert Shaw's avatar
Robert Shaw committed
1194
            allowed_token_ids=self.allowed_token_ids,
1195
1196
            extra_args=extra_args or None,
            )
1197

1198
1199
    @model_validator(mode="before")
    @classmethod
1200
1201
1202
1203
1204
1205
1206
1207
1208
    def check_structured_outputs_count(cls, data):
        if "structured_outputs" not in data:
            return data

        structured_outputs_kwargs = data['structured_outputs']
        count = sum(
            structured_outputs_kwargs.get(k) is not None
            for k in ("json", "regex", "choice"))
        if count > 1:
1209
            raise ValueError(
1210
1211
                "You can only use one kind of constraints for structured "
                "outputs ('json', 'regex' or 'choice').")
1212
1213
        return data

1214
1215
1216
    @model_validator(mode="before")
    @classmethod
    def check_logprobs(cls, data):
1217
        if (prompt_logprobs := data.get("prompt_logprobs")) is not None:
1218
1219
            if data.get("stream") and (prompt_logprobs > 0
                                       or prompt_logprobs == -1):
1220
1221
1222
                raise ValueError(
                    "`prompt_logprobs` are not available when `stream=True`.")

1223
1224
1225
1226
1227
1228
            if prompt_logprobs < 0 and prompt_logprobs != -1:
                raise ValueError(
                    "`prompt_logprobs` must be a positive value or -1.")
            if prompt_logprobs == -1 and not envs.VLLM_USE_V1:
                raise ValueError("`prompt_logprobs=-1` is only supported with "
                                 "vLLM engine V1.")
1229
1230
1231
        if (logprobs := data.get("logprobs")) is not None and logprobs < 0:
            raise ValueError("`logprobs` must be a positive value.")

1232
1233
        return data

1234
1235
1236
1237
1238
    @model_validator(mode="before")
    @classmethod
    def validate_stream_options(cls, data):
        if data.get("stream_options") and not data.get("stream"):
            raise ValueError(
1239
1240
                "Stream options can only be defined when `stream=True`.")

1241
1242
        return data

1243
1244
1245
    @model_validator(mode="before")
    @classmethod
    def validate_prompt_and_prompt_embeds(cls, data):
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
        prompt = data.get("prompt")
        prompt_embeds = data.get("prompt_embeds")

        prompt_is_empty = (prompt is None
                           or (isinstance(prompt, str) and prompt == ""))
        embeds_is_empty = (prompt_embeds is None
                           or (isinstance(prompt_embeds, list)
                               and len(prompt_embeds) == 0))

        if prompt_is_empty and embeds_is_empty:
1256
            raise ValueError(
1257
1258
1259
                "Either prompt or prompt_embeds must be provided and non-empty."
            )

1260
1261
        return data

1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
    @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
1276

1277
class EmbeddingCompletionRequest(OpenAIBaseModel):
1278
1279
    # Ordered by official OpenAI API documentation
    # https://platform.openai.com/docs/api-reference/embeddings
1280
    model: Optional[str] = None
1281
    input: Union[list[int], list[list[int]], str, list[str]]
1282
    encoding_format: Literal["float", "base64"] = "float"
1283
1284
    dimensions: Optional[int] = None
    user: Optional[str] = None
1285
    truncate_prompt_tokens: Optional[Annotated[int, Field(ge=-1)]] = None
1286

1287
    # --8<-- [start:embedding-extra-params]
1288
1289
1290
1291
1292
1293
    add_special_tokens: bool = Field(
        default=True,
        description=(
            "If true (the default), special tokens (e.g. BOS) will be added to "
            "the prompt."),
    )
1294
1295
1296
1297
1298
    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 "
1299
1300
            "if the served model does not use priority scheduling."),
    )
1301
1302
1303
1304
1305
1306
1307
    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 "
            "through out the inference process and return in response."),
    )
1308
    normalize: Optional[bool] = None
1309

1310
    # --8<-- [end:embedding-extra-params]
1311

1312
    def to_pooling_params(self):
1313
1314
1315
1316
        return PoolingParams(
            truncate_prompt_tokens=self.truncate_prompt_tokens,
            dimensions=self.dimensions,
            normalize=self.normalize)
1317
1318


1319
class EmbeddingChatRequest(OpenAIBaseModel):
1320
    model: Optional[str] = None
1321
    messages: list[ChatCompletionMessageParam]
1322
1323
1324
1325

    encoding_format: Literal["float", "base64"] = "float"
    dimensions: Optional[int] = None
    user: Optional[str] = None
1326
    truncate_prompt_tokens: Optional[Annotated[int, Field(ge=-1)]] = None
1327

1328
    # --8<-- [start:chat-embedding-extra-params]
1329
1330
1331
1332
1333
1334
1335
1336
    add_generation_prompt: bool = Field(
        default=False,
        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."),
    )

1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
    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."),
    )
1354
    chat_template_kwargs: Optional[dict[str, Any]] = Field(
1355
        default=None,
1356
1357
1358
        description=(
            "Additional keyword args to pass to the template renderer. "
            "Will be accessible by the chat template."),
1359
    )
1360
    mm_processor_kwargs: Optional[dict[str, Any]] = Field(
1361
1362
1363
        default=None,
        description=("Additional kwargs to pass to the HF processor."),
    )
1364
1365
1366
1367
1368
    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 "
1369
1370
            "if the served model does not use priority scheduling."),
    )
1371
1372
1373
1374
1375
1376
1377
    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 "
            "through out the inference process and return in response."),
    )
1378
    normalize: Optional[bool] = None
1379
    # --8<-- [end:chat-embedding-extra-params]
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390

    @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):
1391
1392
1393
1394
        return PoolingParams(
            truncate_prompt_tokens=self.truncate_prompt_tokens,
            dimensions=self.dimensions,
            normalize=self.normalize)
1395
1396
1397
1398


EmbeddingRequest = Union[EmbeddingCompletionRequest, EmbeddingChatRequest]

1399
1400
PoolingCompletionRequest = EmbeddingCompletionRequest
PoolingChatRequest = EmbeddingChatRequest
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418

T = TypeVar("T")


class IOProcessorRequest(OpenAIBaseModel, Generic[T]):
    model: Optional[str] = None

    priority: int = Field(default=0)
    """
    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.
    """
    data: T
    """
    When using plugins IOProcessor plugins, the actual input is processed
    by the plugin itself. Hence, we use a generic type for the request data
    """
1419
    softmax: bool = True
1420
1421

    def to_pooling_params(self):
1422
        return PoolingParams(task="encode", softmax=self.softmax)
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441


class IOProcessorResponse(OpenAIBaseModel, Generic[T]):

    request_id: Optional[str] = None
    """
    The request_id associated with this response
    """
    created_at: int = Field(default_factory=lambda: int(time.time()))

    data: T
    """
    When using plugins IOProcessor plugins, the actual output is generated
    by the plugin itself. Hence, we use a generic type for the response data
    """


PoolingRequest = Union[PoolingCompletionRequest, PoolingChatRequest,
                       IOProcessorRequest]
1442

1443

1444
class ScoreRequest(OpenAIBaseModel):
1445
    model: Optional[str] = None
1446
1447
    text_1: Union[list[str], str, ScoreMultiModalParam]
    text_2: Union[list[str], str, ScoreMultiModalParam]
1448
    truncate_prompt_tokens: Optional[Annotated[int, Field(ge=-1)]] = None
1449

1450
    # --8<-- [start:score-extra-params]
1451
1452
1453
1454
1455
1456

    mm_processor_kwargs: Optional[dict[str, Any]] = Field(
        default=None,
        description=("Additional kwargs to pass to the HF processor."),
    )

1457
1458
1459
1460
1461
    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 "
1462
1463
            "if the served model does not use priority scheduling."),
    )
1464

1465
1466
    activation: Optional[bool] = None

1467
    # --8<-- [end:score-extra-params]
1468

1469
    def to_pooling_params(self):
1470
1471
1472
        return PoolingParams(
            truncate_prompt_tokens=self.truncate_prompt_tokens,
            activation=self.activation)
1473
1474


1475
class RerankRequest(OpenAIBaseModel):
1476
    model: Optional[str] = None
1477
1478
    query: Union[str, ScoreMultiModalParam]
    documents: Union[list[str], ScoreMultiModalParam]
1479
    top_n: int = Field(default_factory=lambda: 0)
1480
    truncate_prompt_tokens: Optional[Annotated[int, Field(ge=-1)]] = None
1481

1482
    # --8<-- [start:rerank-extra-params]
1483
1484
1485
1486
1487
1488

    mm_processor_kwargs: Optional[dict[str, Any]] = Field(
        default=None,
        description=("Additional kwargs to pass to the HF processor."),
    )

1489
1490
1491
1492
1493
    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 "
1494
1495
            "if the served model does not use priority scheduling."),
    )
1496

1497
1498
    activation: Optional[bool] = None

1499
    # --8<-- [end:rerank-extra-params]
1500

1501
    def to_pooling_params(self):
1502
1503
1504
        return PoolingParams(
            truncate_prompt_tokens=self.truncate_prompt_tokens,
            activation=self.activation)
1505
1506
1507


class RerankDocument(BaseModel):
1508
    text: Optional[str] = None
1509
    multi_modal: Optional[ScoreContentPartParam] = None
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525


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
1526
    results: list[RerankResult]
1527
1528


1529
class CompletionLogProbs(OpenAIBaseModel):
1530
1531
1532
1533
    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,
1534
                                     float]]] = Field(default_factory=list)
Zhuohan Li's avatar
Zhuohan Li committed
1535
1536


1537
class CompletionResponseChoice(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
1538
1539
    index: int
    text: str
1540
    logprobs: Optional[CompletionLogProbs] = None
1541
1542
    finish_reason: Optional[str] = None
    stop_reason: Optional[Union[int, str]] = Field(
1543
1544
1545
1546
1547
1548
        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"),
    )
1549
    token_ids: Optional[list[int]] = None  # For response
1550
    prompt_logprobs: Optional[list[Optional[dict[int, Logprob]]]] = None
1551
    prompt_token_ids: Optional[list[int]] = None  # For prompt
Zhuohan Li's avatar
Zhuohan Li committed
1552
1553


1554
class CompletionResponse(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
1555
    id: str = Field(default_factory=lambda: f"cmpl-{random_uuid()}")
1556
    object: Literal["text_completion"] = "text_completion"
Zhuohan Li's avatar
Zhuohan Li committed
1557
1558
    created: int = Field(default_factory=lambda: int(time.time()))
    model: str
1559
    choices: list[CompletionResponseChoice]
1560
1561
1562
    service_tier: Optional[Literal["auto", "default", "flex", "scale",
                                   "priority"]] = None
    system_fingerprint: Optional[str] = None
Zhuohan Li's avatar
Zhuohan Li committed
1563
    usage: UsageInfo
1564
1565

    # vLLM-specific fields that are not in OpenAI spec
Robert Shaw's avatar
Robert Shaw committed
1566
1567
    kv_transfer_params: Optional[dict[str, Any]] = Field(
        default=None, description="KVTransfer parameters.")
Zhuohan Li's avatar
Zhuohan Li committed
1568
1569


1570
class CompletionResponseStreamChoice(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
1571
1572
    index: int
    text: str
1573
    logprobs: Optional[CompletionLogProbs] = None
1574
1575
    finish_reason: Optional[str] = None
    stop_reason: Optional[Union[int, str]] = Field(
1576
1577
1578
1579
1580
1581
        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"),
    )
1582
1583
1584
1585
    # not part of the OpenAI spec but for tracing the tokens
    # prompt tokens is put into choice to align with CompletionResponseChoice
    prompt_token_ids: Optional[list[int]] = None
    token_ids: Optional[list[int]] = None
Zhuohan Li's avatar
Zhuohan Li committed
1586
1587


1588
class CompletionStreamResponse(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
1589
1590
1591
1592
    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
1593
    choices: list[CompletionResponseStreamChoice]
1594
    usage: Optional[UsageInfo] = Field(default=None)
1595
1596


1597
class EmbeddingResponseData(OpenAIBaseModel):
1598
1599
    index: int
    object: str = "embedding"
1600
    embedding: Union[list[float], str]
1601
1602


1603
class EmbeddingResponse(OpenAIBaseModel):
1604
    id: str = Field(default_factory=lambda: f"embd-{random_uuid()}")
1605
1606
1607
    object: str = "list"
    created: int = Field(default_factory=lambda: int(time.time()))
    model: str
1608
    data: list[EmbeddingResponseData]
1609
1610
1611
    usage: UsageInfo


1612
1613
1614
class PoolingResponseData(OpenAIBaseModel):
    index: int
    object: str = "pooling"
1615
    data: Union[list[list[float]], list[float], str]
1616
1617
1618
1619
1620
1621
1622


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
1623
    data: list[PoolingResponseData]
1624
1625
1626
    usage: UsageInfo


1627
1628
1629
class ScoreResponseData(OpenAIBaseModel):
    index: int
    object: str = "score"
1630
    score: float
1631
1632
1633
1634
1635
1636
1637


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
1638
    data: list[ScoreResponseData]
1639
1640
1641
    usage: UsageInfo


1642
1643
1644
1645
1646
1647
class ClassificationRequest(OpenAIBaseModel):
    model: Optional[str] = None
    input: Union[list[str], str]
    truncate_prompt_tokens: Optional[int] = None
    user: Optional[str] = None

1648
    # --8<-- [start:classification-extra-params]
1649
1650
1651
1652
1653
1654
1655
1656
    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."),
    )

1657
1658
    activation: Optional[bool] = None

1659
    # --8<-- [end:classification-extra-params]
1660
1661

    def to_pooling_params(self):
1662
1663
1664
        return PoolingParams(
            truncate_prompt_tokens=self.truncate_prompt_tokens,
            activation=self.activation)
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682


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


1683
1684
1685
1686
1687
1688
class FunctionCall(OpenAIBaseModel):
    name: str
    arguments: str


class ToolCall(OpenAIBaseModel):
1689
    id: str = Field(default_factory=make_tool_call_id)
1690
1691
1692
1693
    type: Literal["function"] = "function"
    function: FunctionCall


1694
1695
1696
1697
1698
1699
1700
class DeltaFunctionCall(BaseModel):
    name: Optional[str] = None
    arguments: Optional[str] = None


# a tool call delta where everything is optional
class DeltaToolCall(OpenAIBaseModel):
1701
1702
    id: Optional[str] = None
    type: Optional[Literal["function"]] = None
1703
1704
1705
1706
1707
1708
1709
1710
1711
    index: int
    function: Optional[DeltaFunctionCall] = None


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

    # extracted tool calls
1712
    tool_calls: list[ToolCall]
1713
1714
1715
1716
1717
1718

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


1719
class ChatMessage(OpenAIBaseModel):
1720
    role: str
1721
    content: Optional[str] = None
1722
1723
1724
1725
    refusal: Optional[str] = None
    annotations: Optional[OpenAIAnnotation] = None
    audio: Optional[OpenAIChatCompletionAudio] = None
    function_call: Optional[FunctionCall] = None
1726
    tool_calls: list[ToolCall] = Field(default_factory=list)
1727

1728
1729
1730
    # vLLM-specific fields that are not in OpenAI spec
    reasoning_content: Optional[str] = None

1731

1732
1733
1734
class ChatCompletionLogProb(OpenAIBaseModel):
    token: str
    logprob: float = -9999.0
1735
    bytes: Optional[list[int]] = None
1736
1737
1738


class ChatCompletionLogProbsContent(ChatCompletionLogProb):
1739
1740
1741
    # Workaround: redefine fields name cache so that it's not
    # shared with the super class.
    field_names: ClassVar[Optional[set[str]]] = None
1742
    top_logprobs: list[ChatCompletionLogProb] = Field(default_factory=list)
1743
1744
1745


class ChatCompletionLogProbs(OpenAIBaseModel):
1746
    content: Optional[list[ChatCompletionLogProbsContent]] = None
1747
1748


1749
class ChatCompletionResponseChoice(OpenAIBaseModel):
1750
1751
    index: int
    message: ChatMessage
1752
    logprobs: Optional[ChatCompletionLogProbs] = None
1753
1754
1755
    # 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
1756
    stop_reason: Optional[Union[int, str]] = None
1757
1758
1759
    # not part of the OpenAI spec but is useful for tracing the tokens
    # in agent scenarios
    token_ids: Optional[list[int]] = None
1760
1761


1762
class ChatCompletionResponse(OpenAIBaseModel):
1763
    id: str = Field(default_factory=lambda: f"chatcmpl-{random_uuid()}")
1764
    object: Literal["chat.completion"] = "chat.completion"
1765
1766
    created: int = Field(default_factory=lambda: int(time.time()))
    model: str
1767
    choices: list[ChatCompletionResponseChoice]
1768
1769
1770
    service_tier: Optional[Literal["auto", "default", "flex", "scale",
                                   "priority"]] = None
    system_fingerprint: Optional[str] = None
1771
    usage: UsageInfo
1772
1773

    # vLLM-specific fields that are not in OpenAI spec
1774
    prompt_logprobs: Optional[list[Optional[dict[int, Logprob]]]] = None
1775
    prompt_token_ids: Optional[list[int]] = None
Robert Shaw's avatar
Robert Shaw committed
1776
1777
    kv_transfer_params: Optional[dict[str, Any]] = Field(
        default=None, description="KVTransfer parameters.")
1778
1779


1780
class DeltaMessage(OpenAIBaseModel):
1781
1782
    role: Optional[str] = None
    content: Optional[str] = None
1783
    reasoning_content: Optional[str] = None
1784
    tool_calls: list[DeltaToolCall] = Field(default_factory=list)
1785
1786


1787
class ChatCompletionResponseStreamChoice(OpenAIBaseModel):
1788
1789
    index: int
    delta: DeltaMessage
1790
    logprobs: Optional[ChatCompletionLogProbs] = None
1791
    finish_reason: Optional[str] = None
1792
    stop_reason: Optional[Union[int, str]] = None
1793
1794
    # not part of the OpenAI spec but for tracing the tokens
    token_ids: Optional[list[int]] = None
1795
1796


1797
class ChatCompletionStreamResponse(OpenAIBaseModel):
1798
    id: str = Field(default_factory=lambda: f"chatcmpl-{random_uuid()}")
1799
    object: Literal["chat.completion.chunk"] = "chat.completion.chunk"
1800
1801
    created: int = Field(default_factory=lambda: int(time.time()))
    model: str
1802
    choices: list[ChatCompletionResponseStreamChoice]
1803
    usage: Optional[UsageInfo] = Field(default=None)
1804
1805
    # not part of the OpenAI spec but for tracing the tokens
    prompt_token_ids: Optional[list[int]] = None
1806
1807


1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
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)


1823
1824
1825
1826
1827
class InputTokensDetails(OpenAIBaseModel):
    cached_tokens: int


class OutputTokensDetails(OpenAIBaseModel):
1828
1829
    reasoning_tokens: int = 0
    tool_output_tokens: int = 0
1830
1831
1832
1833
1834
1835
1836
1837


class ResponseUsage(OpenAIBaseModel):
    input_tokens: int
    input_tokens_details: InputTokensDetails
    output_tokens: int
    output_tokens_details: OutputTokensDetails
    total_tokens: int
1838
1839
1840
1841
1842
1843


class ResponsesResponse(OpenAIBaseModel):
    id: str = Field(default_factory=lambda: f"resp_{random_uuid()}")
    created_at: int = Field(default_factory=lambda: int(time.time()))
    # error: Optional[ResponseError] = None
1844
    incomplete_details: Optional[IncompleteDetails] = None
1845
1846
1847
1848
    instructions: Optional[str] = None
    metadata: Optional[Metadata] = None
    model: str
    object: Literal["response"] = "response"
1849
    output: list[ResponseOutputItem]
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
    parallel_tool_calls: bool
    temperature: float
    tool_choice: ToolChoice
    tools: list[Tool]
    top_p: float
    background: bool
    max_output_tokens: int
    max_tool_calls: Optional[int] = None
    previous_response_id: Optional[str] = None
    prompt: Optional[ResponsePrompt] = None
    reasoning: Optional[Reasoning] = None
    service_tier: Literal["auto", "default", "flex", "scale", "priority"]
    status: ResponseStatus
    text: Optional[ResponseTextConfig] = None
1864
    top_logprobs: Optional[int] = None
1865
    truncation: Literal["auto", "disabled"]
1866
    usage: Optional[ResponseUsage] = None
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
    user: Optional[str] = None

    @classmethod
    def from_request(
        cls,
        request: ResponsesRequest,
        sampling_params: SamplingParams,
        model_name: str,
        created_time: int,
        output: list[ResponseOutputItem],
        status: ResponseStatus,
1878
        usage: Optional[ResponseUsage] = None,
1879
    ) -> "ResponsesResponse":
1880
1881
1882
1883
1884
1885
1886
1887

        incomplete_details: Optional[IncompleteDetails] = None
        if status == 'incomplete':
            incomplete_details = IncompleteDetails(reason='max_output_tokens')
        # TODO: implement the other reason for incomplete_details,
        # which is content_filter
        # incomplete_details = IncompleteDetails(reason='content_filter')

1888
1889
1890
        return cls(
            id=request.request_id,
            created_at=created_time,
1891
            incomplete_details=incomplete_details,
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
            instructions=request.instructions,
            metadata=request.metadata,
            model=model_name,
            output=output,
            parallel_tool_calls=request.parallel_tool_calls,
            temperature=sampling_params.temperature,
            tool_choice=request.tool_choice,
            tools=request.tools,
            top_p=sampling_params.top_p,
            background=request.background,
            max_output_tokens=sampling_params.max_tokens,
            max_tool_calls=request.max_tool_calls,
            previous_response_id=request.previous_response_id,
            prompt=request.prompt,
            reasoning=request.reasoning,
            service_tier=request.service_tier,
            status=status,
            text=request.text,
            top_logprobs=sampling_params.logprobs,
            truncation=request.truncation,
            user=request.user,
            usage=usage,
        )


1917
1918
1919
1920
BatchRequestInputBody = Union[ChatCompletionRequest, EmbeddingRequest,
                              ScoreRequest, RerankRequest]


1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
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

1940
    # The parameters of the request.
1941
    body: BatchRequestInputBody
1942

1943
1944
1945
1946
    @field_validator('body', mode='plain')
    @classmethod
    def check_type_for_url(cls, value: Any, info: ValidationInfo):
        # Use url to disambiguate models
1947
        url: str = info.data["url"]
1948
1949
1950
1951
        if url == "/v1/chat/completions":
            return ChatCompletionRequest.model_validate(value)
        if url == "/v1/embeddings":
            return TypeAdapter(EmbeddingRequest).validate_python(value)
1952
        if url.endswith("/score"):
1953
            return ScoreRequest.model_validate(value)
1954
1955
1956
        if url.endswith("/rerank"):
            return RerankRequest.model_validate(value)
        return TypeAdapter(BatchRequestInputBody).validate_python(value)
1957

1958

1959
1960
1961
1962
1963
1964
1965
1966
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.
1967
    body: Optional[Union[ChatCompletionResponse, EmbeddingResponse,
1968
                         ScoreResponse, RerankResponse]] = None
1969
1970


1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
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

1982
    response: Optional[BatchResponseData]
1983
1984
1985
1986

    # For requests that failed with a non-HTTP error, this will contain more
    # information on the cause of the failure.
    error: Optional[Any]
1987
1988


1989
class TokenizeCompletionRequest(OpenAIBaseModel):
1990
    model: Optional[str] = None
1991
1992
    prompt: str

1993
1994
1995
1996
1997
1998
    add_special_tokens: bool = Field(
        default=True,
        description=(
            "If true (the default), special tokens (e.g. BOS) will be added to "
            "the prompt."),
    )
1999
2000
2001
2002
2003
    return_token_strs: Optional[bool] = Field(
        default=False,
        description=("If true, also return the token strings "
                     "corresponding to the token ids."),
    )
2004
2005
2006


class TokenizeChatRequest(OpenAIBaseModel):
2007
    model: Optional[str] = None
2008
    messages: list[ChatCompletionMessageParam]
2009

2010
2011
2012
2013
2014
2015
2016
    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."),
    )
2017
2018
2019
2020
2021
    return_token_strs: Optional[bool] = Field(
        default=False,
        description=("If true, also return the token strings "
                     "corresponding to the token ids."),
    )
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
    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."),
    )
2048
    chat_template_kwargs: Optional[dict[str, Any]] = Field(
2049
        default=None,
2050
2051
2052
        description=(
            "Additional keyword args to pass to the template renderer. "
            "Will be accessible by the chat template."),
2053
    )
2054
    mm_processor_kwargs: Optional[dict[str, Any]] = Field(
2055
2056
2057
        default=None,
        description=("Additional kwargs to pass to the HF processor."),
    )
2058
2059
2060
2061
    tools: Optional[list[ChatCompletionToolsParam]] = Field(
        default=None,
        description=("A list of tools the model may call."),
    )
2062

2063
2064
2065
2066
2067
2068
2069
2070
2071
    @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

2072
2073

TokenizeRequest = Union[TokenizeCompletionRequest, TokenizeChatRequest]
2074
2075
2076
2077
2078


class TokenizeResponse(OpenAIBaseModel):
    count: int
    max_model_len: int
2079
    tokens: list[int]
2080
    token_strs: Optional[list[str]] = None
2081
2082
2083


class DetokenizeRequest(OpenAIBaseModel):
2084
    model: Optional[str] = None
2085
    tokens: list[int]
2086
2087
2088
2089


class DetokenizeResponse(OpenAIBaseModel):
    prompt: str
2090
2091


2092
2093
class TokenizerInfoResponse(OpenAIBaseModel):
    """
2094
    Response containing tokenizer configuration
2095
2096
2097
2098
2099
2100
2101
    equivalent to tokenizer_config.json
    """

    model_config = ConfigDict(extra="allow")
    tokenizer_class: str


2102
class LoadLoRAAdapterRequest(BaseModel):
2103
2104
2105
2106
    lora_name: str
    lora_path: str


2107
class UnloadLoRAAdapterRequest(BaseModel):
2108
2109
    lora_name: str
    lora_int_id: Optional[int] = Field(default=None)
2110
2111
2112
2113
2114
2115
2116
2117
2118


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


class TranscriptionRequest(OpenAIBaseModel):
    # Ordered by official OpenAI API documentation
2119
    # https://platform.openai.com/docs/api-reference/audio/createTranscription
2120
2121
2122
2123
2124
2125
2126

    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.
    """

2127
    model: Optional[str] = None
2128
2129
2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
2153
2154
    """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 !!

2155
    timestamp_granularities: list[Literal["word", "segment"]] = Field(
2156
2157
2158
2159
2160
2161
2162
2163
2164
        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.
    """

2165
    stream: Optional[bool] = False
2166
    """When set, it will enable output to be streamed in a similar fashion
2167
    as the Chat Completion endpoint.
2168
    """
2169
    # --8<-- [start:transcription-extra-params]
2170
2171
2172
    # Flattened stream option to simplify form data.
    stream_include_usage: Optional[bool] = False
    stream_continuous_usage_stats: Optional[bool] = False
2173
2174
2175
2176
2177
2178

    vllm_xargs: Optional[dict[str, Union[str, int, float]]] = Field(
        default=None,
        description=("Additional request parameters with string or "
                     "numeric values, used by custom extensions."),
    )
2179
    # --8<-- [end:transcription-extra-params]
2180

2181
2182
2183
    to_language: Optional[str] = None
    """The language of the output audio we transcribe to.

2184
    Please note that this is not currently used by supported models at this
2185
2186
2187
    time, but it is a placeholder for future use, matching translation api.
    """

2188
    # --8<-- [start:transcription-sampling-params]
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
    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
2199
    """Enables nucleus (top-p) sampling, where tokens are selected from the
2200
2201
2202
2203
2204
2205
2206
    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
2207
    """Filters out tokens with a probability lower than `min_p`, ensuring a
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
    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."""
2222
    # --8<-- [end:transcription-sampling-params]
2223

2224
2225
    # Default sampling parameters for transcription requests.
    _DEFAULT_SAMPLING_PARAMS: dict = {
2226
2227
2228
        "repetition_penalty": 1.0,
        "temperature": 1.0,
        "top_p": 1.0,
2229
        "top_k": 0,
2230
        "min_p": 0.0,
2231
2232
2233
2234
2235
2236
    }

    def to_sampling_params(
            self,
            default_max_tokens: int,
            default_sampling_params: Optional[dict] = None) -> SamplingParams:
2237

2238
2239
2240
2241
        max_tokens = default_max_tokens

        if default_sampling_params is None:
            default_sampling_params = {}
2242

2243
2244
2245
2246
        # Default parameters
        if (temperature := self.temperature) is None:
            temperature = default_sampling_params.get(
                "temperature", self._DEFAULT_SAMPLING_PARAMS["temperature"])
2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
        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"])
2261
2262

        return SamplingParams.from_optional(temperature=temperature,
2263
                                            max_tokens=max_tokens,
2264
2265
2266
2267
2268
2269
2270
                                            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,
2271
2272
                                            output_kind=RequestOutputKind.DELTA
                                            if self.stream \
2273
2274
                                            else RequestOutputKind.FINAL_ONLY,
                                            extra_args=self.vllm_xargs)
2275
2276
2277

    @model_validator(mode="before")
    @classmethod
2278
2279
2280
2281
2282
2283
2284
    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'.",
            )

2285
2286
2287
2288
2289
2290
2291
        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
2292
2293
2294


# Transcription response objects
2295
2296
2297
2298
2299
class TranscriptionUsageAudio(OpenAIBaseModel):
    type: Literal["duration"] = "duration"
    seconds: int


2300
2301
2302
class TranscriptionResponse(OpenAIBaseModel):
    text: str
    """The transcribed text."""
2303
    usage: TranscriptionUsageAudio
2304
2305
2306
2307
2308
2309
2310
2311
2312
2313
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
2327
2328
2329
2330
2331
2332
2333
2334
2335
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
2350
2351
2352
2353
2354


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."""

2355
    tokens: list[int]
2356
2357
2358
2359
2360
2361
2362
2363
2364
2365
2366
2367
2368
    """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."""

2369
    segments: Optional[list[TranscriptionSegment]] = None
2370
2371
    """Segments of the transcribed text and their corresponding details."""

2372
    words: Optional[list[TranscriptionWord]] = None
2373
    """Extracted words and their corresponding timestamps."""
2374
2375
2376
2377
2378
2379
2380
2381
2382
2383
2384
2385
2386
2387
2388
2389
2390
2391
2392
2393
2394
2395
2396
2397
2398
2399
2400
2401
2402
2403
2404
2405
2406
2407
2408
2409
2410
2411
2412
2413
2414
2415
2416
2417
2418
2419
2420


class TranslationResponseStreamChoice(OpenAIBaseModel):
    delta: DeltaMessage
    finish_reason: Optional[str] = None
    stop_reason: Optional[Union[int, str]] = None


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


class TranslationRequest(OpenAIBaseModel):
    # Ordered by official OpenAI API documentation
    # https://platform.openai.com/docs/api-reference/audio/createTranslation

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

    model: Optional[str] = None
    """ID of the model to use.
    """

    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 support additional sampling parameters
    # --8<-- [start:translation-sampling-params]
2421
2422
2423
    seed: Optional[int] = Field(None, ge=_LONG_INFO.min, le=_LONG_INFO.max)
    """The seed to use for sampling."""

2424
2425
2426
2427
2428
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
2442
    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.
    """
    # --8<-- [end:translation-sampling-params]

    # --8<-- [start:translation-extra-params]
    language: Optional[str] = None
    """The language of the input audio we translate from.

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

2443
2444
2445
2446
2447
2448
2449
2450
    to_language: Optional[str] = None
    """The language of the input audio we translate to.

    Please note that this is not supported by all models, refer to the specific
    model documentation for more details.
    For instance, Whisper only supports `to_language=en`.
    """

2451
    stream: Optional[bool] = False
2452
    """Custom field not present in the original OpenAI definition. When set,
2453
    it will enable output to be streamed in a similar fashion as the Chat
2454
    Completion endpoint.
2455
2456
2457
2458
2459
2460
2461
2462
2463
2464
2465
2466
2467
2468
2469
    """
    # Flattened stream option to simplify form data.
    stream_include_usage: Optional[bool] = False
    stream_continuous_usage_stats: Optional[bool] = False
    # --8<-- [end:translation-extra-params]

    # Default sampling parameters for translation requests.
    _DEFAULT_SAMPLING_PARAMS: dict = {
        "temperature": 0,
    }

    def to_sampling_params(
            self,
            default_max_tokens: int,
            default_sampling_params: Optional[dict] = None) -> SamplingParams:
2470

2471
2472
2473
2474
2475
2476
2477
2478
2479
2480
2481
        max_tokens = default_max_tokens

        if default_sampling_params is None:
            default_sampling_params = {}
        # Default parameters
        if (temperature := self.temperature) is None:
            temperature = default_sampling_params.get(
                "temperature", self._DEFAULT_SAMPLING_PARAMS["temperature"])

        return SamplingParams.from_optional(temperature=temperature,
                                            max_tokens=max_tokens,
2482
                                            seed=self.seed,
2483
2484
2485
2486
2487
2488
2489
2490
2491
2492
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
2503
2504
2505
2506
2507
2508
2509
2510
2511
2512
2513
2514
2515
2516
2517
2518
2519
2520
2521
2522
2523
2524
2525
2526
2527
2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
2538
2539
2540
2541
2542
2543
2544
2545
2546
2547
2548
2549
2550
2551
2552
2553
2554
2555
2556
2557
2558
2559
2560
2561
2562
2563
2564
2565
2566
2567
2568
2569
2570
2571
2572
                                            output_kind=RequestOutputKind.DELTA
                                            if self.stream \
                                            else RequestOutputKind.FINAL_ONLY)

    @model_validator(mode="before")
    @classmethod
    def validate_stream_options(cls, data):
        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


# Translation response objects
class TranslationResponse(OpenAIBaseModel):
    text: str
    """The translated text."""


class TranslationWord(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 TranslationSegment(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."""

    tokens: list[int]
    """Array of token IDs for the text content."""


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

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

    text: str
    """The translated text."""

    segments: Optional[list[TranslationSegment]] = None
    """Segments of the translated text and their corresponding details."""

    words: Optional[list[TranslationWord]] = None
    """Extracted words and their corresponding timestamps."""