protocol.py 86.2 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
from typing import Annotated, Any, ClassVar, Literal, Optional, Union
Zhuohan Li's avatar
Zhuohan Li committed
10

11
import regex as re
12
import torch
13
from fastapi import HTTPException, UploadFile
14
15
16
17
18
19
# 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
20
21
22
23
24
25
from openai.types.responses import (ResponseInputParam, ResponseOutputItem,
                                    ResponseOutputMessage, ResponsePrompt,
                                    ResponseStatus, ResponseTextConfig)
from openai.types.responses.response import ToolChoice
from openai.types.responses.tool import Tool
from openai.types.shared import Metadata, Reasoning
26
27
from pydantic import (BaseModel, ConfigDict, Field, TypeAdapter,
                      ValidationInfo, field_validator, model_validator)
28
from typing_extensions import TypeAlias
Zhuohan Li's avatar
Zhuohan Li committed
29

30
from vllm import envs
31
32
from vllm.entrypoints.chat_utils import (ChatCompletionMessageParam,
                                         random_tool_call_id)
33
34
from vllm.entrypoints.score_utils import (ScoreContentPartParam,
                                          ScoreMultiModalParam)
35
from vllm.logger import init_logger
36
from vllm.pooling_params import PoolingParams
37
38
from vllm.sampling_params import (BeamSearchParams, GuidedDecodingParams,
                                  RequestOutputKind, SamplingParams)
39
from vllm.sequence import Logprob
40
from vllm.utils import random_uuid, resolve_obj_by_qualname
41

42
43
logger = init_logger(__name__)

44
_LONG_INFO = torch.iinfo(torch.long)
45

Zhuohan Li's avatar
Zhuohan Li committed
46

47
class OpenAIBaseModel(BaseModel):
48
49
50
    # OpenAI API does allow extra fields
    model_config = ConfigDict(extra="allow")

51
    # Cache class field names
52
    field_names: ClassVar[Optional[set[str]]] = None
53

54
    @model_validator(mode="wrap")
55
    @classmethod
56
57
58
59
    def __log_extra_fields__(cls, data, handler):
        result = handler(data)
        if not isinstance(data, dict):
            return result
60
61
        field_names = cls.field_names
        if field_names is None:
62
63
64
65
            # 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)
66
                if alias := getattr(field, "alias", None):
67
68
69
70
71
72
73
74
                    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",
75
76
                data.keys() - field_names,
            )
77
        return result
78
79
80


class ErrorResponse(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
81
82
83
84
    object: str = "error"
    message: str
    type: str
    param: Optional[str] = None
85
    code: int
Zhuohan Li's avatar
Zhuohan Li committed
86
87


88
class ModelPermission(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
89
90
91
92
93
94
95
96
97
98
99
    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
100
    is_blocking: bool = False
Zhuohan Li's avatar
Zhuohan Li committed
101
102


103
class ModelCard(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
104
105
106
    id: str
    object: str = "model"
    created: int = Field(default_factory=lambda: int(time.time()))
Woosuk Kwon's avatar
Woosuk Kwon committed
107
    owned_by: str = "vllm"
Zhuohan Li's avatar
Zhuohan Li committed
108
109
    root: Optional[str] = None
    parent: Optional[str] = None
110
    max_model_len: Optional[int] = None
111
    permission: list[ModelPermission] = Field(default_factory=list)
Zhuohan Li's avatar
Zhuohan Li committed
112
113


114
class ModelList(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
115
    object: str = "list"
116
    data: list[ModelCard] = Field(default_factory=list)
Zhuohan Li's avatar
Zhuohan Li committed
117
118


119
120
121
122
class PromptTokenUsageInfo(OpenAIBaseModel):
    cached_tokens: Optional[int] = None


123
class UsageInfo(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
124
125
126
    prompt_tokens: int = 0
    total_tokens: int = 0
    completion_tokens: Optional[int] = 0
127
    prompt_tokens_details: Optional[PromptTokenUsageInfo] = None
Zhuohan Li's avatar
Zhuohan Li committed
128
129


130
131
132
133
134
class RequestResponseMetadata(BaseModel):
    request_id: str
    final_usage_info: Optional[UsageInfo] = None


135
136
137
138
139
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
140
    json_schema: Optional[dict[str, Any]] = Field(default=None, alias='schema')
141
142
143
    strict: Optional[bool] = None


144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
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]


159
class ResponseFormat(OpenAIBaseModel):
160
    # type must be "json_schema", "json_object", or "text"
161
162
    type: Literal["text", "json_object", "json_schema"]
    json_schema: Optional[JsonSchemaResponseFormat] = None
163
164


165
166
167
AnyResponseFormat = Union[ResponseFormat, StructuralTagResponseFormat]


168
class StreamOptions(OpenAIBaseModel):
169
    include_usage: Optional[bool] = True
170
    continuous_usage_stats: Optional[bool] = False
171
172


173
174
175
class FunctionDefinition(OpenAIBaseModel):
    name: str
    description: Optional[str] = None
176
    parameters: Optional[dict[str, Any]] = None
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192


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


class ChatCompletionNamedFunction(OpenAIBaseModel):
    name: str


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


193
194
# extra="forbid" is a workaround to have kwargs as a field,
# see https://github.com/pydantic/pydantic/issues/3125
195
196
class LogitsProcessorConstructor(BaseModel):
    qualname: str
197
198
    args: Optional[list[Any]] = None
    kwargs: Optional[dict[str, Any]] = None
199

200
201
    model_config = ConfigDict(extra="forbid")

202

203
LogitsProcessors = list[Union[str, LogitsProcessorConstructor]]
204
205
206


def get_logits_processors(processors: Optional[LogitsProcessors],
207
                          pattern: Optional[str]) -> Optional[list[Any]]:
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
    if processors and pattern:
        logits_processors = []
        for processor in processors:
            qualname = processor if isinstance(processor,
                                               str) else processor.qualname
            if not re.match(pattern, qualname):
                raise ValueError(
                    f"Logits processor '{qualname}' is not allowed by this "
                    "server. See --logits-processor-pattern engine argument "
                    "for more information.")
            try:
                logits_processor = resolve_obj_by_qualname(qualname)
            except Exception as e:
                raise ValueError(
                    f"Logits processor '{qualname}' could not be resolved: {e}"
                ) from e
            if isinstance(processor, LogitsProcessorConstructor):
                logits_processor = logits_processor(*processor.args or [],
                                                    **processor.kwargs or {})
            logits_processors.append(logits_processor)
        return logits_processors
    elif processors:
        raise ValueError(
            "The `logits_processors` argument is not supported by this "
            "server. See --logits-processor-pattern engine argugment "
            "for more information.")
    return None


237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
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
    input: Union[str, ResponseInputParam]
    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."),
    )
293
294
295
296
297
298
299
300
301
    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."))
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
    # --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"])

        # Structured output
        guided_decoding = None
        if self.text is not None and self.text.format is not None:
            response_format = self.text.format
            if response_format.type == "json_schema":
                guided_decoding = GuidedDecodingParams.from_optional(
                    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,
            logprobs=self.top_logprobs,
            output_kind=(RequestOutputKind.DELTA
                         if self.stream else RequestOutputKind.FINAL_ONLY),
            guided_decoding=guided_decoding,
        )

    @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

363
364
365
366
367
368
369
370
371
372
373
374
375
    @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

376

377
class ChatCompletionRequest(OpenAIBaseModel):
378
379
    # Ordered by official OpenAI API documentation
    # https://platform.openai.com/docs/api-reference/chat/create
380
    messages: list[ChatCompletionMessageParam]
381
    model: Optional[str] = None
382
    frequency_penalty: Optional[float] = 0.0
383
    logit_bias: Optional[dict[str, float]] = None
384
    logprobs: Optional[bool] = False
385
    top_logprobs: Optional[int] = 0
386
387
388
389
390
    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
391
392
    n: Optional[int] = 1
    presence_penalty: Optional[float] = 0.0
393
    response_format: Optional[AnyResponseFormat] = None
394
    seed: Optional[int] = Field(None, ge=_LONG_INFO.min, le=_LONG_INFO.max)
395
    stop: Optional[Union[str, list[str]]] = []
Zhuohan Li's avatar
Zhuohan Li committed
396
    stream: Optional[bool] = False
397
    stream_options: Optional[StreamOptions] = None
398
399
    temperature: Optional[float] = None
    top_p: Optional[float] = None
400
    tools: Optional[list[ChatCompletionToolsParam]] = None
401
402
403
404
405
406
    tool_choice: Optional[Union[
        Literal["none"],
        Literal["auto"],
        Literal["required"],
        ChatCompletionNamedToolChoiceParam,
    ]] = "none"
407

408
    # NOTE this will be ignored by vLLM -- the model determines the behavior
409
    parallel_tool_calls: Optional[bool] = False
Zhuohan Li's avatar
Zhuohan Li committed
410
    user: Optional[str] = None
411

412
    # --8<-- [start:chat-completion-sampling-params]
413
    best_of: Optional[int] = None
414
    use_beam_search: bool = False
415
416
417
    top_k: Optional[int] = None
    min_p: Optional[float] = None
    repetition_penalty: Optional[float] = None
418
    length_penalty: float = 1.0
419
    stop_token_ids: Optional[list[int]] = []
420
421
422
423
424
425
    include_stop_str_in_output: bool = False
    ignore_eos: bool = False
    min_tokens: int = 0
    skip_special_tokens: bool = True
    spaces_between_special_tokens: bool = True
    truncate_prompt_tokens: Optional[Annotated[int, Field(ge=1)]] = None
426
    prompt_logprobs: Optional[int] = None
427
    allowed_token_ids: Optional[list[int]] = None
428
    bad_words: list[str] = Field(default_factory=list)
429
    # --8<-- [end:chat-completion-sampling-params]
430

431
    # --8<-- [start:chat-completion-extra-params]
432
    echo: bool = Field(
433
434
435
436
437
        default=False,
        description=(
            "If true, the new message will be prepended with the last message "
            "if they belong to the same role."),
    )
438
    add_generation_prompt: bool = Field(
439
440
441
442
443
444
        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."),
    )
445
446
447
448
449
450
451
452
453
    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`."),
    )
454
    add_special_tokens: bool = Field(
455
456
457
458
459
        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 "
460
            "special tokens so this should be set to false (as is the "
461
462
            "default)."),
    )
463
    documents: Optional[list[dict[str, str]]] = Field(
464
465
466
467
468
469
470
471
472
473
474
475
        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. "
476
477
478
            "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."),
479
    )
480
    chat_template_kwargs: Optional[dict[str, Any]] = Field(
481
        default=None,
482
483
484
        description=(
            "Additional keyword args to pass to the template renderer. "
            "Will be accessible by the chat template."),
485
    )
486
    mm_processor_kwargs: Optional[dict[str, Any]] = Field(
487
488
489
        default=None,
        description=("Additional kwargs to pass to the HF processor."),
    )
490
491
492
493
494
495
496
497
498
    guided_json: Optional[Union[str, dict, BaseModel]] = Field(
        default=None,
        description=("If specified, the output will follow the JSON schema."),
    )
    guided_regex: Optional[str] = Field(
        default=None,
        description=(
            "If specified, the output will follow the regex pattern."),
    )
499
    guided_choice: Optional[list[str]] = Field(
500
501
502
503
504
505
506
507
508
        default=None,
        description=(
            "If specified, the output will be exactly one of the choices."),
    )
    guided_grammar: Optional[str] = Field(
        default=None,
        description=(
            "If specified, the output will follow the context free grammar."),
    )
509
510
511
512
513
    structural_tag: Optional[str] = Field(
        default=None,
        description=(
            "If specified, the output will follow the structural tag schema."),
    )
514
515
516
517
518
    guided_decoding_backend: Optional[str] = Field(
        default=None,
        description=(
            "If specified, will override the default guided decoding backend "
            "of the server for this specific request. If set, must be either "
519
520
            "'outlines' / 'lm-format-enforcer'"),
    )
521
522
523
524
    guided_whitespace_pattern: Optional[str] = Field(
        default=None,
        description=(
            "If specified, will override the default whitespace pattern "
525
526
            "for guided json decoding."),
    )
527
528
529
530
531
    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 "
532
533
            "if the served model does not use priority scheduling."),
    )
534
535
536
537
538
    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 "
539
540
            "through out the inference process and return in response."),
    )
541
542
543
544
545
546
547
548
549
550
551
    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'}}."))
552
553
554
555
556
557
    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."))
558
559
560
561
562
563
564
565
566
    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
567
568
569
    kv_transfer_params: Optional[dict[str, Any]] = Field(
        default=None,
        description="KVTransfer parameters used for disaggregated serving.")
570

571
572
573
574
575
576
    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."),
    )

577
    # --8<-- [end:chat-completion-extra-params]
Zhuohan Li's avatar
Zhuohan Li committed
578

579
580
581
582
583
    # Default sampling parameters for chat completion requests
    _DEFAULT_SAMPLING_PARAMS: dict = {
        "repetition_penalty": 1.0,
        "temperature": 1.0,
        "top_p": 1.0,
584
        "top_k": 0,
585
586
587
588
        "min_p": 0.0,
    }

    def to_beam_search_params(
589
590
            self, max_tokens: int,
            default_sampling_params: dict) -> BeamSearchParams:
591
592

        n = self.n if self.n is not None else 1
593
594
595
        if (temperature := self.temperature) is None:
            temperature = default_sampling_params.get(
                "temperature", self._DEFAULT_SAMPLING_PARAMS["temperature"])
596
597
598
599
600
601

        return BeamSearchParams(
            beam_width=n,
            max_tokens=max_tokens,
            ignore_eos=self.ignore_eos,
            temperature=temperature,
602
            length_penalty=self.length_penalty,
603
604
            include_stop_str_in_output=self.include_stop_str_in_output,
        )
605

606
    def to_sampling_params(
607
        self,
608
        max_tokens: int,
609
        logits_processor_pattern: Optional[str],
610
        default_sampling_params: dict,
611
    ) -> SamplingParams:
612

613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
        # 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"])

632
633
634
635
        prompt_logprobs = self.prompt_logprobs
        if prompt_logprobs is None and self.echo:
            prompt_logprobs = self.top_logprobs

636
        guided_json_object = None
637
638
639
640
641
642
643
        if self.response_format is not None:
            if self.response_format.type == "json_object":
                guided_json_object = True
            elif self.response_format.type == "json_schema":
                json_schema = self.response_format.json_schema
                assert json_schema is not None
                self.guided_json = json_schema.json_schema
644
645
646
647
648
649
            elif self.response_format.type == "structural_tag":
                structural_tag = self.response_format
                assert structural_tag is not None and isinstance(
                    structural_tag, StructuralTagResponseFormat)
                s_tag_obj = structural_tag.model_dump(by_alias=True)
                self.structural_tag = json.dumps(s_tag_obj)
650
651
652
653
654
655
656
657

        guided_decoding = GuidedDecodingParams.from_optional(
            json=self._get_guided_json_from_tool() or self.guided_json,
            regex=self.guided_regex,
            choice=self.guided_choice,
            grammar=self.guided_grammar,
            json_object=guided_json_object,
            backend=self.guided_decoding_backend,
658
            whitespace_pattern=self.guided_whitespace_pattern,
659
            structural_tag=self.structural_tag,
660
        )
661

662
663
664
665
        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
666
        return SamplingParams.from_optional(
667
            n=self.n,
668
            best_of=self.best_of,
669
670
            presence_penalty=self.presence_penalty,
            frequency_penalty=self.frequency_penalty,
671
672
673
674
675
            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
676
            seed=self.seed,
677
678
            stop=self.stop,
            stop_token_ids=self.stop_token_ids,
679
            logprobs=self.top_logprobs if self.logprobs else None,
680
            prompt_logprobs=prompt_logprobs,
681
            ignore_eos=self.ignore_eos,
682
            max_tokens=max_tokens,
683
            min_tokens=self.min_tokens,
684
685
            skip_special_tokens=self.skip_special_tokens,
            spaces_between_special_tokens=self.spaces_between_special_tokens,
686
687
            logits_processors=get_logits_processors(self.logits_processors,
                                                    logits_processor_pattern),
688
            include_stop_str_in_output=self.include_stop_str_in_output,
689
            truncate_prompt_tokens=self.truncate_prompt_tokens,
690
691
            output_kind=RequestOutputKind.DELTA if self.stream \
                else RequestOutputKind.FINAL_ONLY,
692
            guided_decoding=guided_decoding,
Robert Shaw's avatar
Robert Shaw committed
693
            logit_bias=self.logit_bias,
694
            bad_words= self.bad_words,
695
            allowed_token_ids=self.allowed_token_ids,
696
697
            extra_args=extra_args or None,
        )
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714

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

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

715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
        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"]
                }

740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
            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

758
759
760
761
762
763
764
765
            json_schema = {
                "type": "array",
                "minItems": 1,
                "items": {
                    "type": "object",
                    "anyOf": [get_tool_schema(tool) for tool in self.tools]
                }
            }
766
767
768
            json_schema_defs = get_tool_schema_defs(self.tools)
            if json_schema_defs:
                json_schema["$defs"] = json_schema_defs
769
770
            return json_schema

771
        return None
772

773
    @model_validator(mode="before")
774
    @classmethod
775
776
    def validate_stream_options(cls, data):
        if data.get("stream_options") and not data.get("stream"):
777
            raise ValueError(
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
                "Stream options can only be defined when `stream=True`.")

        return data

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

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

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

797
            if top_logprobs > 0 and not data.get("logprobs"):
798
799
800
801
802
                raise ValueError(
                    "when using `top_logprobs`, `logprobs` must be set to true."
                )

        return data
803

804
805
806
    @model_validator(mode="before")
    @classmethod
    def check_guided_decoding_count(cls, data):
807
808
809
        if isinstance(data, ValueError):
            raise data

810
811
812
813
814
        guide_count = sum([
            "guided_json" in data and data["guided_json"] is not None,
            "guided_regex" in data and data["guided_regex"] is not None,
            "guided_choice" in data and data["guided_choice"] is not None
        ])
815
        # you can only use one kind of guided decoding
816
817
818
819
        if guide_count > 1:
            raise ValueError(
                "You can only use one kind of guided decoding "
                "('guided_json', 'guided_regex' or 'guided_choice').")
820
        # you can only either use guided decoding or tools, not both
821
822
823
824
825
        if guide_count > 1 and data.get("tool_choice", "none") not in (
                "none",
                "auto",
                "required",
        ):
826
827
828
829
830
831
            raise ValueError(
                "You can only either use guided decoding or tools, not both.")
        return data

    @model_validator(mode="before")
    @classmethod
832
833
834
835
    def check_tool_usage(cls, data):

        # if "tool_choice" is not specified but tools are provided,
        # default to "auto" tool_choice
836
        if "tool_choice" not in data and data.get("tools"):
837
838
            data["tool_choice"] = "auto"

839
        # if "tool_choice" is "none" -- no validation is needed for tools
840
841
842
        if "tool_choice" in data and data["tool_choice"] == "none":
            return data

843
844
845
846
        # if "tool_choice" is specified -- validation
        if "tool_choice" in data:

            # ensure that if "tool choice" is specified, tools are present
847
848
849
            if "tools" not in data or data["tools"] is None:
                raise ValueError(
                    "When using `tool_choice`, `tools` must be set.")
850
851

            # make sure that tool choice is either a named tool
852
853
854
855
856
857
858
859
860
            # OR that it's set to "auto" or "required"
            if data["tool_choice"] not in [
                    "auto", "required"
            ] and not isinstance(data["tool_choice"], dict):
                raise NotImplementedError(
                    f'Invalid value for `tool_choice`: {data["tool_choice"]}! '\
                    'Only named tools, "none", "auto" or "required" '\
                    'are supported.'
                )
861
862
863

            # ensure that if "tool_choice" is specified as an object,
            # it matches a valid tool
864
865
            correct_usage_message = 'Correct usage: `{"type": "function",' \
                ' "function": {"name": "my_function"}}`'
866
867
            if isinstance(data["tool_choice"], dict):
                valid_tool = False
868
869
                function = data["tool_choice"].get("function")
                if not isinstance(function, dict):
870
                    raise ValueError(
871
872
873
874
875
876
877
878
                        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:
879
                    raise ValueError(
880
881
                        f"Invalid `name` in `function`: `{function_name}`"
                        f" in `tool_choice`! {correct_usage_message}")
882
                for tool in data["tools"]:
883
                    if tool["function"]["name"] == function_name:
884
885
886
887
888
889
                        valid_tool = True
                        break
                if not valid_tool:
                    raise ValueError(
                        "The tool specified in `tool_choice` does not match any"
                        " of the specified `tools`")
890
891
        return data

892
893
894
895
896
897
898
899
900
    @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

901
902
903
904
905
906
907
908
909
910
911
912
913
914
    @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
915

916
class CompletionRequest(OpenAIBaseModel):
917
918
    # Ordered by official OpenAI API documentation
    # https://platform.openai.com/docs/api-reference/completions/create
919
    model: Optional[str] = None
920
921
    prompt: Optional[Union[list[int], list[list[int]], str, list[str]]] = None
    prompt_embeds: Optional[Union[bytes, list[bytes]]] = None
922
    best_of: Optional[int] = None
Zhuohan Li's avatar
Zhuohan Li committed
923
924
    echo: Optional[bool] = False
    frequency_penalty: Optional[float] = 0.0
925
    logit_bias: Optional[dict[str, float]] = None
926
927
    logprobs: Optional[int] = None
    max_tokens: Optional[int] = 16
928
    n: int = 1
929
    presence_penalty: Optional[float] = 0.0
930
    seed: Optional[int] = Field(None, ge=_LONG_INFO.min, le=_LONG_INFO.max)
931
    stop: Optional[Union[str, list[str]]] = []
932
    stream: Optional[bool] = False
933
    stream_options: Optional[StreamOptions] = None
934
    suffix: Optional[str] = None
935
936
    temperature: Optional[float] = None
    top_p: Optional[float] = None
Zhuohan Li's avatar
Zhuohan Li committed
937
    user: Optional[str] = None
938

939
    # --8<-- [start:completion-sampling-params]
940
    use_beam_search: bool = False
941
942
943
    top_k: Optional[int] = None
    min_p: Optional[float] = None
    repetition_penalty: Optional[float] = None
944
    length_penalty: float = 1.0
945
    stop_token_ids: Optional[list[int]] = []
946
947
948
949
950
    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
951
    truncate_prompt_tokens: Optional[Annotated[int, Field(ge=1)]] = None
952
    allowed_token_ids: Optional[list[int]] = None
953
    prompt_logprobs: Optional[int] = None
954
    # --8<-- [end:completion-sampling-params]
955

956
    # --8<-- [start:completion-extra-params]
957
958
    add_special_tokens: bool = Field(
        default=True,
959
        description=(
960
961
            "If true (the default), special tokens (e.g. BOS) will be added to "
            "the prompt."),
962
    )
963
    response_format: Optional[AnyResponseFormat] = Field(
964
        default=None,
965
966
967
968
969
        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."
        ),
970
971
972
    )
    guided_json: Optional[Union[str, dict, BaseModel]] = Field(
        default=None,
973
        description="If specified, the output will follow the JSON schema.",
974
975
976
977
978
979
    )
    guided_regex: Optional[str] = Field(
        default=None,
        description=(
            "If specified, the output will follow the regex pattern."),
    )
980
    guided_choice: Optional[list[str]] = Field(
981
982
983
984
985
986
987
988
989
        default=None,
        description=(
            "If specified, the output will be exactly one of the choices."),
    )
    guided_grammar: Optional[str] = Field(
        default=None,
        description=(
            "If specified, the output will follow the context free grammar."),
    )
990
991
992
993
994
    guided_decoding_backend: Optional[str] = Field(
        default=None,
        description=(
            "If specified, will override the default guided decoding backend "
            "of the server for this specific request. If set, must be one of "
995
996
            "'outlines' / 'lm-format-enforcer'"),
    )
997
998
999
1000
    guided_whitespace_pattern: Optional[str] = Field(
        default=None,
        description=(
            "If specified, will override the default whitespace pattern "
1001
1002
            "for guided json decoding."),
    )
1003
1004
1005
1006
1007
    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 "
1008
1009
            "if the served model does not use priority scheduling."),
    )
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
    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'}}."))
1021

1022
1023
1024
1025
1026
1027
    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."))
1028

1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
    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
1039
1040
1041
1042
    kv_transfer_params: Optional[dict[str, Any]] = Field(
        default=None,
        description="KVTransfer parameters used for disaggregated serving.")

1043
1044
1045
1046
1047
1048
    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."),
    )

1049
    # --8<-- [end:completion-extra-params]
Zhuohan Li's avatar
Zhuohan Li committed
1050

1051
1052
1053
1054
1055
    # Default sampling parameters for completion requests
    _DEFAULT_SAMPLING_PARAMS: dict = {
        "repetition_penalty": 1.0,
        "temperature": 1.0,
        "top_p": 1.0,
1056
        "top_k": 0,
1057
1058
1059
1060
        "min_p": 0.0,
    }

    def to_beam_search_params(
1061
1062
1063
        self,
        max_tokens: int,
        default_sampling_params: Optional[dict] = None,
1064
    ) -> BeamSearchParams:
1065

1066
1067
        if default_sampling_params is None:
            default_sampling_params = {}
1068
        n = self.n if self.n is not None else 1
1069
1070
1071

        if (temperature := self.temperature) is None:
            temperature = default_sampling_params.get("temperature", 1.0)
1072
1073
1074
1075
1076
1077

        return BeamSearchParams(
            beam_width=n,
            max_tokens=max_tokens,
            ignore_eos=self.ignore_eos,
            temperature=temperature,
1078
            length_penalty=self.length_penalty,
1079
1080
            include_stop_str_in_output=self.include_stop_str_in_output,
        )
1081

1082
    def to_sampling_params(
1083
        self,
1084
        max_tokens: int,
1085
1086
1087
        logits_processor_pattern: Optional[str],
        default_sampling_params: Optional[dict] = None,
    ) -> SamplingParams:
1088

1089
1090
        if default_sampling_params is None:
            default_sampling_params = {}
1091

1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
        # 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"])

1111
1112
1113
1114
        prompt_logprobs = self.prompt_logprobs
        if prompt_logprobs is None and self.echo:
            prompt_logprobs = self.logprobs

1115
1116
        echo_without_generation = self.echo and self.max_tokens == 0

1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
        guided_json_object = None
        if (self.response_format is not None
                and self.response_format.type == "json_object"):
            guided_json_object = True

        guided_decoding = GuidedDecodingParams.from_optional(
            json=self.guided_json,
            regex=self.guided_regex,
            choice=self.guided_choice,
            grammar=self.guided_grammar,
            json_object=guided_json_object,
            backend=self.guided_decoding_backend,
1129
1130
            whitespace_pattern=self.guided_whitespace_pattern,
        )
1131

1132
1133
1134
1135
        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
1136
        return SamplingParams.from_optional(
1137
            n=self.n,
1138
            best_of=self.best_of,
1139
1140
            presence_penalty=self.presence_penalty,
            frequency_penalty=self.frequency_penalty,
1141
1142
1143
1144
1145
            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
1146
            seed=self.seed,
1147
1148
            stop=self.stop,
            stop_token_ids=self.stop_token_ids,
1149
            logprobs=self.logprobs,
1150
            ignore_eos=self.ignore_eos,
1151
            max_tokens=max_tokens if not echo_without_generation else 1,
1152
            min_tokens=self.min_tokens,
1153
            prompt_logprobs=prompt_logprobs,
1154
            skip_special_tokens=self.skip_special_tokens,
1155
            spaces_between_special_tokens=self.spaces_between_special_tokens,
1156
            include_stop_str_in_output=self.include_stop_str_in_output,
1157
1158
            logits_processors=get_logits_processors(self.logits_processors,
                                                    logits_processor_pattern),
1159
            truncate_prompt_tokens=self.truncate_prompt_tokens,
1160
1161
            output_kind=RequestOutputKind.DELTA if self.stream \
                else RequestOutputKind.FINAL_ONLY,
1162
1163
            guided_decoding=guided_decoding,
            logit_bias=self.logit_bias,
Robert Shaw's avatar
Robert Shaw committed
1164
            allowed_token_ids=self.allowed_token_ids,
1165
1166
            extra_args=extra_args or None,
            )
1167

1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
    @model_validator(mode="before")
    @classmethod
    def check_guided_decoding_count(cls, data):
        guide_count = sum([
            "guided_json" in data and data["guided_json"] is not None,
            "guided_regex" in data and data["guided_regex"] is not None,
            "guided_choice" in data and data["guided_choice"] is not None
        ])
        if guide_count > 1:
            raise ValueError(
                "You can only use one kind of guided decoding "
                "('guided_json', 'guided_regex' or 'guided_choice').")
        return data

1182
1183
1184
    @model_validator(mode="before")
    @classmethod
    def check_logprobs(cls, data):
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
        if (prompt_logprobs := data.get("prompt_logprobs")) is not None:
            if data.get("stream") and prompt_logprobs > 0:
                raise ValueError(
                    "`prompt_logprobs` are not available when `stream=True`.")

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

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

1196
1197
        return data

1198
1199
1200
1201
1202
    @model_validator(mode="before")
    @classmethod
    def validate_stream_options(cls, data):
        if data.get("stream_options") and not data.get("stream"):
            raise ValueError(
1203
1204
                "Stream options can only be defined when `stream=True`.")

1205
1206
        return data

1207
1208
1209
1210
1211
1212
1213
1214
    @model_validator(mode="before")
    @classmethod
    def validate_prompt_and_prompt_embeds(cls, data):
        if data.get("prompt") is None and data.get("prompt_embeds") is None:
            raise ValueError(
                "At least one of `prompt` or `prompt_embeds` must be set.")
        return data

1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
    @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
1229

1230
class EmbeddingCompletionRequest(OpenAIBaseModel):
1231
1232
    # Ordered by official OpenAI API documentation
    # https://platform.openai.com/docs/api-reference/embeddings
1233
    model: Optional[str] = None
1234
    input: Union[list[int], list[list[int]], str, list[str]]
1235
    encoding_format: Literal["float", "base64"] = "float"
1236
1237
    dimensions: Optional[int] = None
    user: Optional[str] = None
1238
    truncate_prompt_tokens: Optional[Annotated[int, Field(ge=-1)]] = None
1239

1240
    # --8<-- [start:embedding-extra-params]
1241
1242
1243
1244
1245
1246
    add_special_tokens: bool = Field(
        default=True,
        description=(
            "If true (the default), special tokens (e.g. BOS) will be added to "
            "the prompt."),
    )
1247
1248
1249
1250
1251
    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 "
1252
1253
            "if the served model does not use priority scheduling."),
    )
1254

1255
    # --8<-- [end:embedding-extra-params]
1256

1257
    def to_pooling_params(self):
1258
        return PoolingParams(dimensions=self.dimensions)
1259
1260


1261
class EmbeddingChatRequest(OpenAIBaseModel):
1262
    model: Optional[str] = None
1263
    messages: list[ChatCompletionMessageParam]
1264
1265
1266
1267

    encoding_format: Literal["float", "base64"] = "float"
    dimensions: Optional[int] = None
    user: Optional[str] = None
1268
    truncate_prompt_tokens: Optional[Annotated[int, Field(ge=-1)]] = None
1269

1270
    # --8<-- [start:chat-embedding-extra-params]
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
    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."),
    )
1288
    chat_template_kwargs: Optional[dict[str, Any]] = Field(
1289
        default=None,
1290
1291
1292
        description=(
            "Additional keyword args to pass to the template renderer. "
            "Will be accessible by the chat template."),
1293
    )
1294
    mm_processor_kwargs: Optional[dict[str, Any]] = Field(
1295
1296
1297
        default=None,
        description=("Additional kwargs to pass to the HF processor."),
    )
1298
1299
1300
1301
1302
    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 "
1303
1304
            "if the served model does not use priority scheduling."),
    )
1305
    # --8<-- [end:chat-embedding-extra-params]
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316

    @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):
1317
        return PoolingParams(dimensions=self.dimensions)
1318
1319
1320
1321


EmbeddingRequest = Union[EmbeddingCompletionRequest, EmbeddingChatRequest]

1322
1323
1324
1325
PoolingCompletionRequest = EmbeddingCompletionRequest
PoolingChatRequest = EmbeddingChatRequest
PoolingRequest = Union[PoolingCompletionRequest, PoolingChatRequest]

1326

1327
class ScoreRequest(OpenAIBaseModel):
1328
    model: Optional[str] = None
1329
1330
    text_1: Union[list[str], str, ScoreMultiModalParam]
    text_2: Union[list[str], str, ScoreMultiModalParam]
1331
    truncate_prompt_tokens: Optional[Annotated[int, Field(ge=-1)]] = None
1332

1333
    # --8<-- [start:score-extra-params]
1334
1335
1336
1337
1338
1339

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

1340
1341
1342
1343
1344
    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 "
1345
1346
            "if the served model does not use priority scheduling."),
    )
1347

1348
    # --8<-- [end:score-extra-params]
1349

1350
    def to_pooling_params(self, *, use_cross_encoder: bool = False):
1351
        return PoolingParams(use_cross_encoder=use_cross_encoder)
1352
1353


1354
class RerankRequest(OpenAIBaseModel):
1355
    model: Optional[str] = None
1356
1357
    query: Union[str, ScoreMultiModalParam]
    documents: Union[list[str], ScoreMultiModalParam]
1358
    top_n: int = Field(default_factory=lambda: 0)
1359
    truncate_prompt_tokens: Optional[Annotated[int, Field(ge=-1)]] = None
1360

1361
    # --8<-- [start:rerank-extra-params]
1362
1363
1364
1365
1366
1367

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

1368
1369
1370
1371
1372
    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 "
1373
1374
            "if the served model does not use priority scheduling."),
    )
1375

1376
    # --8<-- [end:rerank-extra-params]
1377

1378
    def to_pooling_params(self, *, use_cross_encoder: bool = False):
1379
        return PoolingParams(use_cross_encoder=use_cross_encoder)
1380
1381
1382


class RerankDocument(BaseModel):
1383
    text: Optional[str] = None
1384
    multi_modal: Optional[ScoreContentPartParam] = None
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400


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
1401
    results: list[RerankResult]
1402
1403


1404
class CompletionLogProbs(OpenAIBaseModel):
1405
1406
1407
1408
    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,
1409
                                     float]]] = Field(default_factory=list)
Zhuohan Li's avatar
Zhuohan Li committed
1410
1411


1412
class CompletionResponseChoice(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
1413
1414
    index: int
    text: str
1415
    logprobs: Optional[CompletionLogProbs] = None
1416
1417
    finish_reason: Optional[str] = None
    stop_reason: Optional[Union[int, str]] = Field(
1418
1419
1420
1421
1422
1423
        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"),
    )
1424
    prompt_logprobs: Optional[list[Optional[dict[int, Logprob]]]] = None
Zhuohan Li's avatar
Zhuohan Li committed
1425
1426


1427
class CompletionResponse(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
1428
    id: str = Field(default_factory=lambda: f"cmpl-{random_uuid()}")
1429
    object: Literal["text_completion"] = "text_completion"
Zhuohan Li's avatar
Zhuohan Li committed
1430
1431
    created: int = Field(default_factory=lambda: int(time.time()))
    model: str
1432
    choices: list[CompletionResponseChoice]
1433
1434
1435
    service_tier: Optional[Literal["auto", "default", "flex", "scale",
                                   "priority"]] = None
    system_fingerprint: Optional[str] = None
Zhuohan Li's avatar
Zhuohan Li committed
1436
    usage: UsageInfo
1437
1438

    # vLLM-specific fields that are not in OpenAI spec
Robert Shaw's avatar
Robert Shaw committed
1439
1440
    kv_transfer_params: Optional[dict[str, Any]] = Field(
        default=None, description="KVTransfer parameters.")
Zhuohan Li's avatar
Zhuohan Li committed
1441
1442


1443
class CompletionResponseStreamChoice(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
1444
1445
    index: int
    text: str
1446
    logprobs: Optional[CompletionLogProbs] = None
1447
1448
    finish_reason: Optional[str] = None
    stop_reason: Optional[Union[int, str]] = Field(
1449
1450
1451
1452
1453
1454
        default=None,
        description=(
            "The stop string or token id that caused the completion "
            "to stop, None if the completion finished for some other reason "
            "including encountering the EOS token"),
    )
Zhuohan Li's avatar
Zhuohan Li committed
1455
1456


1457
class CompletionStreamResponse(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
1458
1459
1460
1461
    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
1462
    choices: list[CompletionResponseStreamChoice]
1463
    usage: Optional[UsageInfo] = Field(default=None)
1464
1465


1466
class EmbeddingResponseData(OpenAIBaseModel):
1467
1468
    index: int
    object: str = "embedding"
1469
    embedding: Union[list[float], str]
1470
1471


1472
class EmbeddingResponse(OpenAIBaseModel):
1473
    id: str = Field(default_factory=lambda: f"embd-{random_uuid()}")
1474
1475
1476
    object: str = "list"
    created: int = Field(default_factory=lambda: int(time.time()))
    model: str
1477
    data: list[EmbeddingResponseData]
1478
1479
1480
    usage: UsageInfo


1481
1482
1483
class PoolingResponseData(OpenAIBaseModel):
    index: int
    object: str = "pooling"
1484
    data: Union[list[list[float]], list[float], str]
1485
1486
1487
1488
1489
1490
1491


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
1492
    data: list[PoolingResponseData]
1493
1494
1495
    usage: UsageInfo


1496
1497
1498
class ScoreResponseData(OpenAIBaseModel):
    index: int
    object: str = "score"
1499
    score: float
1500
1501
1502
1503
1504
1505
1506


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
1507
    data: list[ScoreResponseData]
1508
1509
1510
    usage: UsageInfo


1511
1512
1513
1514
1515
1516
class ClassificationRequest(OpenAIBaseModel):
    model: Optional[str] = None
    input: Union[list[str], str]
    truncate_prompt_tokens: Optional[int] = None
    user: Optional[str] = None

1517
    # --8<-- [start:classification-extra-params]
1518
1519
1520
1521
1522
1523
1524
1525
    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."),
    )

1526
    # --8<-- [end:classification-extra-params]
1527
1528

    def to_pooling_params(self):
1529
        return PoolingParams()
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547


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


1548
1549
1550
1551
1552
1553
class FunctionCall(OpenAIBaseModel):
    name: str
    arguments: str


class ToolCall(OpenAIBaseModel):
1554
    id: str = Field(default_factory=random_tool_call_id)
1555
1556
1557
1558
    type: Literal["function"] = "function"
    function: FunctionCall


1559
1560
1561
1562
1563
1564
1565
class DeltaFunctionCall(BaseModel):
    name: Optional[str] = None
    arguments: Optional[str] = None


# a tool call delta where everything is optional
class DeltaToolCall(OpenAIBaseModel):
1566
1567
    id: Optional[str] = None
    type: Optional[Literal["function"]] = None
1568
1569
1570
1571
1572
1573
1574
1575
1576
    index: int
    function: Optional[DeltaFunctionCall] = None


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

    # extracted tool calls
1577
    tool_calls: list[ToolCall]
1578
1579
1580
1581
1582
1583

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


1584
class ChatMessage(OpenAIBaseModel):
1585
    role: str
1586
    content: Optional[str] = None
1587
1588
1589
1590
    refusal: Optional[str] = None
    annotations: Optional[OpenAIAnnotation] = None
    audio: Optional[OpenAIChatCompletionAudio] = None
    function_call: Optional[FunctionCall] = None
1591
    tool_calls: list[ToolCall] = Field(default_factory=list)
1592

1593
1594
1595
    # vLLM-specific fields that are not in OpenAI spec
    reasoning_content: Optional[str] = None

1596

1597
1598
1599
class ChatCompletionLogProb(OpenAIBaseModel):
    token: str
    logprob: float = -9999.0
1600
    bytes: Optional[list[int]] = None
1601
1602
1603


class ChatCompletionLogProbsContent(ChatCompletionLogProb):
1604
1605
1606
    # Workaround: redefine fields name cache so that it's not
    # shared with the super class.
    field_names: ClassVar[Optional[set[str]]] = None
1607
    top_logprobs: list[ChatCompletionLogProb] = Field(default_factory=list)
1608
1609
1610


class ChatCompletionLogProbs(OpenAIBaseModel):
1611
    content: Optional[list[ChatCompletionLogProbsContent]] = None
1612
1613


1614
class ChatCompletionResponseChoice(OpenAIBaseModel):
1615
1616
    index: int
    message: ChatMessage
1617
    logprobs: Optional[ChatCompletionLogProbs] = None
1618
1619
1620
    # 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
1621
    stop_reason: Optional[Union[int, str]] = None
1622
1623


1624
class ChatCompletionResponse(OpenAIBaseModel):
1625
    id: str = Field(default_factory=lambda: f"chatcmpl-{random_uuid()}")
1626
    object: Literal["chat.completion"] = "chat.completion"
1627
1628
    created: int = Field(default_factory=lambda: int(time.time()))
    model: str
1629
    choices: list[ChatCompletionResponseChoice]
1630
1631
1632
    service_tier: Optional[Literal["auto", "default", "flex", "scale",
                                   "priority"]] = None
    system_fingerprint: Optional[str] = None
1633
    usage: UsageInfo
1634
1635

    # vLLM-specific fields that are not in OpenAI spec
1636
    prompt_logprobs: Optional[list[Optional[dict[int, Logprob]]]] = None
Robert Shaw's avatar
Robert Shaw committed
1637
1638
    kv_transfer_params: Optional[dict[str, Any]] = Field(
        default=None, description="KVTransfer parameters.")
1639
1640


1641
class DeltaMessage(OpenAIBaseModel):
1642
1643
    role: Optional[str] = None
    content: Optional[str] = None
1644
    reasoning_content: Optional[str] = None
1645
    tool_calls: list[DeltaToolCall] = Field(default_factory=list)
1646
1647


1648
class ChatCompletionResponseStreamChoice(OpenAIBaseModel):
1649
1650
    index: int
    delta: DeltaMessage
1651
    logprobs: Optional[ChatCompletionLogProbs] = None
1652
    finish_reason: Optional[str] = None
1653
    stop_reason: Optional[Union[int, str]] = None
1654
1655


1656
class ChatCompletionStreamResponse(OpenAIBaseModel):
1657
    id: str = Field(default_factory=lambda: f"chatcmpl-{random_uuid()}")
1658
    object: Literal["chat.completion.chunk"] = "chat.completion.chunk"
1659
1660
    created: int = Field(default_factory=lambda: int(time.time()))
    model: str
1661
    choices: list[ChatCompletionResponseStreamChoice]
1662
    usage: Optional[UsageInfo] = Field(default=None)
1663
1664


1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
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)


1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
class ResponseReasoningItem(OpenAIBaseModel):
    id: str = Field(default_factory=lambda: f"rs_{random_uuid()}")
    text: str
    summary: list = Field(default_factory=list)
    type: Literal["reasoning"] = "reasoning"
    encrypted_content: Optional[str] = None
    status: Optional[Literal["in_progress", "completed", "incomplete"]]


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
    # incomplete_details: Optional[IncompleteDetails] = None
    instructions: Optional[str] = None
    metadata: Optional[Metadata] = None
    model: str
    object: Literal["response"] = "response"
    output: list[Union[ResponseOutputMessage, ResponseReasoningItem]]
    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
    top_logprobs: int
    truncation: Literal["auto", "disabled"]
    usage: Optional[UsageInfo] = None
    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,
        usage: Optional[UsageInfo] = None,
    ) -> "ResponsesResponse":
        return cls(
            id=request.request_id,
            created_at=created_time,
            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,
        )


1757
1758
1759
1760
BatchRequestInputBody = Union[ChatCompletionRequest, EmbeddingRequest,
                              ScoreRequest, RerankRequest]


1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
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

1780
    # The parameters of the request.
1781
    body: BatchRequestInputBody
1782

1783
1784
1785
1786
    @field_validator('body', mode='plain')
    @classmethod
    def check_type_for_url(cls, value: Any, info: ValidationInfo):
        # Use url to disambiguate models
1787
        url: str = info.data["url"]
1788
1789
1790
1791
        if url == "/v1/chat/completions":
            return ChatCompletionRequest.model_validate(value)
        if url == "/v1/embeddings":
            return TypeAdapter(EmbeddingRequest).validate_python(value)
1792
        if url.endswith("/score"):
1793
            return ScoreRequest.model_validate(value)
1794
1795
1796
        if url.endswith("/rerank"):
            return RerankRequest.model_validate(value)
        return TypeAdapter(BatchRequestInputBody).validate_python(value)
1797

1798

1799
1800
1801
1802
1803
1804
1805
1806
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.
1807
    body: Optional[Union[ChatCompletionResponse, EmbeddingResponse,
1808
                         ScoreResponse, RerankResponse]] = None
1809
1810


1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
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

1822
    response: Optional[BatchResponseData]
1823
1824
1825
1826

    # For requests that failed with a non-HTTP error, this will contain more
    # information on the cause of the failure.
    error: Optional[Any]
1827
1828


1829
class TokenizeCompletionRequest(OpenAIBaseModel):
1830
    model: Optional[str] = None
1831
1832
    prompt: str

1833
1834
1835
1836
1837
1838
    add_special_tokens: bool = Field(
        default=True,
        description=(
            "If true (the default), special tokens (e.g. BOS) will be added to "
            "the prompt."),
    )
1839
1840
1841
1842
1843
    return_token_strs: Optional[bool] = Field(
        default=False,
        description=("If true, also return the token strings "
                     "corresponding to the token ids."),
    )
1844
1845
1846


class TokenizeChatRequest(OpenAIBaseModel):
1847
    model: Optional[str] = None
1848
    messages: list[ChatCompletionMessageParam]
1849

1850
1851
1852
1853
1854
1855
1856
    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."),
    )
1857
1858
1859
1860
1861
    return_token_strs: Optional[bool] = Field(
        default=False,
        description=("If true, also return the token strings "
                     "corresponding to the token ids."),
    )
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
    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."),
    )
1888
    chat_template_kwargs: Optional[dict[str, Any]] = Field(
1889
        default=None,
1890
1891
1892
        description=(
            "Additional keyword args to pass to the template renderer. "
            "Will be accessible by the chat template."),
1893
    )
1894
    mm_processor_kwargs: Optional[dict[str, Any]] = Field(
1895
1896
1897
        default=None,
        description=("Additional kwargs to pass to the HF processor."),
    )
1898
1899
1900
1901
    tools: Optional[list[ChatCompletionToolsParam]] = Field(
        default=None,
        description=("A list of tools the model may call."),
    )
1902

1903
1904
1905
1906
1907
1908
1909
1910
1911
    @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

1912
1913

TokenizeRequest = Union[TokenizeCompletionRequest, TokenizeChatRequest]
1914
1915
1916
1917
1918


class TokenizeResponse(OpenAIBaseModel):
    count: int
    max_model_len: int
1919
    tokens: list[int]
1920
    token_strs: Optional[list[str]] = None
1921
1922
1923


class DetokenizeRequest(OpenAIBaseModel):
1924
    model: Optional[str] = None
1925
    tokens: list[int]
1926
1927
1928
1929


class DetokenizeResponse(OpenAIBaseModel):
    prompt: str
1930
1931


1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
class TokenizerInfoResponse(OpenAIBaseModel):
    """
    Response containing tokenizer configuration 
    equivalent to tokenizer_config.json
    """

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


1942
class LoadLoRAAdapterRequest(BaseModel):
1943
1944
1945
1946
    lora_name: str
    lora_path: str


1947
class UnloadLoRAAdapterRequest(BaseModel):
1948
1949
    lora_name: str
    lora_int_id: Optional[int] = Field(default=None)
1950
1951
1952
1953
1954
1955
1956
1957
1958


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


class TranscriptionRequest(OpenAIBaseModel):
    # Ordered by official OpenAI API documentation
1959
    # https://platform.openai.com/docs/api-reference/audio/createTranscription
1960
1961
1962
1963
1964
1965
1966

    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.
    """

1967
    model: Optional[str] = None
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
    """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 !!

1995
    timestamp_granularities: list[Literal["word", "segment"]] = Field(
1996
1997
1998
1999
2000
2001
2002
2003
2004
        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.
    """

2005
    stream: Optional[bool] = False
2006
    """When set, it will enable output to be streamed in a similar fashion
2007
    as the Chat Completion endpoint.
2008
    """
2009
    # --8<-- [start:transcription-extra-params]
2010
2011
2012
    # Flattened stream option to simplify form data.
    stream_include_usage: Optional[bool] = False
    stream_continuous_usage_stats: Optional[bool] = False
2013
2014
2015
2016
2017
2018

    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."),
    )
2019
    # --8<-- [end:transcription-extra-params]
2020

2021
    # --8<-- [start:transcription-sampling-params]
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
    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
2032
    """Enables nucleus (top-p) sampling, where tokens are selected from the
2033
2034
2035
2036
2037
2038
2039
    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
2040
    """Filters out tokens with a probability lower than `min_p`, ensuring a
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
2054
    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."""
2055
    # --8<-- [end:transcription-sampling-params]
2056

2057
2058
    # Default sampling parameters for transcription requests.
    _DEFAULT_SAMPLING_PARAMS: dict = {
2059
2060
2061
        "repetition_penalty": 1.0,
        "temperature": 1.0,
        "top_p": 1.0,
2062
        "top_k": 0,
2063
        "min_p": 0.0,
2064
2065
2066
2067
2068
2069
    }

    def to_sampling_params(
            self,
            default_max_tokens: int,
            default_sampling_params: Optional[dict] = None) -> SamplingParams:
2070

2071
2072
2073
2074
        max_tokens = default_max_tokens

        if default_sampling_params is None:
            default_sampling_params = {}
2075

2076
2077
2078
2079
        # Default parameters
        if (temperature := self.temperature) is None:
            temperature = default_sampling_params.get(
                "temperature", self._DEFAULT_SAMPLING_PARAMS["temperature"])
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
        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"])
2094
2095

        return SamplingParams.from_optional(temperature=temperature,
2096
                                            max_tokens=max_tokens,
2097
2098
2099
2100
2101
2102
2103
                                            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,
2104
2105
                                            output_kind=RequestOutputKind.DELTA
                                            if self.stream \
2106
2107
                                            else RequestOutputKind.FINAL_ONLY,
                                            extra_args=self.vllm_xargs)
2108
2109
2110

    @model_validator(mode="before")
    @classmethod
2111
2112
2113
2114
2115
2116
2117
    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'.",
            )

2118
2119
2120
2121
2122
2123
2124
        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
2125
2126
2127
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
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171
2172
2173
2174
2175
2176
2177
2178
2179
2180
2181


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


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

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

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


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

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

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

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

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

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

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

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

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

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

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

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

2182
    tokens: list[int]
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
    """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."""

2196
    segments: Optional[list[TranscriptionSegment]] = None
2197
2198
    """Segments of the transcribed text and their corresponding details."""

2199
    words: Optional[list[TranscriptionWord]] = None
2200
    """Extracted words and their corresponding timestamps."""
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
2232
2233
2234
2235
2236
2237
2238
2239
2240
2241
2242
2243
2244
2245
2246
2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
2261
2262
2263
2264
2265
2266
2267


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]
    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.
    """

    stream: Optional[bool] = False
2268
    """Custom field not present in the original OpenAI definition. When set,
2269
    it will enable output to be streamed in a similar fashion as the Chat
2270
    Completion endpoint.
2271
2272
2273
2274
2275
2276
2277
2278
2279
2280
2281
2282
2283
2284
2285
    """
    # 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:
2286

2287
2288
2289
2290
2291
2292
2293
2294
2295
2296
2297
2298
2299
2300
2301
2302
2303
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
2355
2356
2357
2358
2359
2360
2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
2373
2374
2375
2376
2377
2378
2379
2380
2381
2382
2383
2384
2385
2386
2387
        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,
                                            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."""