protocol.py 107 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, Generic, Literal, TypeAlias, TypeVar
Zhuohan Li's avatar
Zhuohan Li committed
10

11
import regex as re
12
import torch
13
from fastapi import HTTPException, UploadFile
14
from openai.types.chat.chat_completion_audio import (
15
16
17
    ChatCompletionAudio as OpenAIChatCompletionAudio,
)
from openai.types.chat.chat_completion_message import Annotation as OpenAIAnnotation
18
19
20
21
22
from openai.types.responses import (
    ResponseCodeInterpreterCallCodeDeltaEvent,
    ResponseCodeInterpreterCallCodeDoneEvent,
    ResponseCodeInterpreterCallCompletedEvent,
    ResponseCodeInterpreterCallInProgressEvent,
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
    ResponseCodeInterpreterCallInterpretingEvent,
    ResponseContentPartAddedEvent,
    ResponseContentPartDoneEvent,
    ResponseFunctionToolCall,
    ResponseInputItemParam,
    ResponseOutputItem,
    ResponseOutputItemAddedEvent,
    ResponseOutputItemDoneEvent,
    ResponsePrompt,
    ResponseReasoningItem,
    ResponseReasoningTextDeltaEvent,
    ResponseReasoningTextDoneEvent,
    ResponseStatus,
    ResponseWebSearchCallCompletedEvent,
    ResponseWebSearchCallInProgressEvent,
    ResponseWebSearchCallSearchingEvent,
)
40
from openai.types.responses import (
41
42
43
    ResponseCompletedEvent as OpenAIResponseCompletedEvent,
)
from openai.types.responses import ResponseCreatedEvent as OpenAIResponseCreatedEvent
44
from openai.types.responses import (
45
46
    ResponseInProgressEvent as OpenAIResponseInProgressEvent,
)
47
from openai.types.responses.response_reasoning_item import (
48
49
    Content as ResponseReasoningTextContent,
)
50
from openai_harmony import Message as OpenAIHarmonyMessage
51

52
53
from vllm.config.pooler import get_use_activation
from vllm.tasks import PoolingTask
54
55
56
57
58
59
from vllm.utils.serial_utils import (
    EmbedDType,
    EncodingFormat,
    Endianness,
)

60
61
62
63
# Backward compatibility for OpenAI client versions
try:  # For older openai versions (< 1.100.0)
    from openai.types.responses import ResponseTextConfig
except ImportError:  # For newer openai versions (>= 1.100.0)
64
    from openai.types.responses import ResponseFormatTextConfig as ResponseTextConfig
65

66

67
from openai.types.responses.response import IncompleteDetails, ToolChoice
68
69
from openai.types.responses.tool import Tool
from openai.types.shared import Metadata, Reasoning
70
71
72
73
74
from pydantic import (
    BaseModel,
    ConfigDict,
    Field,
    TypeAdapter,
75
    ValidationError,
76
    ValidationInfo,
77
    field_serializer,
78
79
80
    field_validator,
    model_validator,
)
Zhuohan Li's avatar
Zhuohan Li committed
81

82
from vllm import envs
83
84
from vllm.entrypoints.chat_utils import ChatCompletionMessageParam, make_tool_call_id
from vllm.entrypoints.score_utils import ScoreContentPartParam, ScoreMultiModalParam
85
from vllm.logger import init_logger
86
from vllm.logprobs import Logprob
87
from vllm.pooling_params import PoolingParams
88
89
90
91
92
93
from vllm.sampling_params import (
    BeamSearchParams,
    RequestOutputKind,
    SamplingParams,
    StructuredOutputsParams,
)
94
95
from vllm.utils import random_uuid
from vllm.utils.import_utils import resolve_obj_by_qualname
96

97
98
logger = init_logger(__name__)

99
_LONG_INFO = torch.iinfo(torch.long)
100

Zhuohan Li's avatar
Zhuohan Li committed
101

102
class OpenAIBaseModel(BaseModel):
103
104
105
    # OpenAI API does allow extra fields
    model_config = ConfigDict(extra="allow")

106
    # Cache class field names
107
    field_names: ClassVar[set[str] | None] = None
108

109
    @model_validator(mode="wrap")
110
    @classmethod
111
112
113
114
    def __log_extra_fields__(cls, data, handler):
        result = handler(data)
        if not isinstance(data, dict):
            return result
115
116
        field_names = cls.field_names
        if field_names is None:
117
118
119
120
            # 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)
121
                if alias := getattr(field, "alias", None):
122
123
124
125
126
127
                    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(
128
                "The following fields were present in the request but ignored: %s",
129
130
                data.keys() - field_names,
            )
131
        return result
132
133


134
class ErrorInfo(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
135
136
    message: str
    type: str
137
    param: str | None = None
138
    code: int
Zhuohan Li's avatar
Zhuohan Li committed
139
140


141
142
143
144
class ErrorResponse(OpenAIBaseModel):
    error: ErrorInfo


145
class ModelPermission(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
146
147
148
149
150
151
152
153
154
155
    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 = "*"
156
    group: str | None = None
157
    is_blocking: bool = False
Zhuohan Li's avatar
Zhuohan Li committed
158
159


160
class ModelCard(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
161
162
163
    id: str
    object: str = "model"
    created: int = Field(default_factory=lambda: int(time.time()))
Woosuk Kwon's avatar
Woosuk Kwon committed
164
    owned_by: str = "vllm"
165
166
167
    root: str | None = None
    parent: str | None = None
    max_model_len: int | None = None
168
    permission: list[ModelPermission] = Field(default_factory=list)
Zhuohan Li's avatar
Zhuohan Li committed
169
170


171
class ModelList(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
172
    object: str = "list"
173
    data: list[ModelCard] = Field(default_factory=list)
Zhuohan Li's avatar
Zhuohan Li committed
174
175


176
class PromptTokenUsageInfo(OpenAIBaseModel):
177
    cached_tokens: int | None = None
178
179


180
class UsageInfo(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
181
182
    prompt_tokens: int = 0
    total_tokens: int = 0
183
184
    completion_tokens: int | None = 0
    prompt_tokens_details: PromptTokenUsageInfo | None = None
Zhuohan Li's avatar
Zhuohan Li committed
185
186


187
188
class RequestResponseMetadata(BaseModel):
    request_id: str
189
    final_usage_info: UsageInfo | None = None
190
191


192
193
class JsonSchemaResponseFormat(OpenAIBaseModel):
    name: str
194
    description: str | None = None
195
196
    # schema is the field in openai but that causes conflicts with pydantic so
    # instead use json_schema with an alias
197
198
    json_schema: dict[str, Any] | None = Field(default=None, alias="schema")
    strict: bool | None = None
199
200


201
class LegacyStructuralTag(OpenAIBaseModel):
202
203
204
    begin: str
    # schema is the field, but that causes conflicts with pydantic so
    # instead use structural_tag_schema with an alias
205
    structural_tag_schema: dict[str, Any] | None = Field(default=None, alias="schema")
206
207
208
    end: str


209
class LegacyStructuralTagResponseFormat(OpenAIBaseModel):
210
    type: Literal["structural_tag"]
211
    structures: list[LegacyStructuralTag]
212
213
214
    triggers: list[str]


215
216
217
218
219
220
221
222
223
224
class StructuralTagResponseFormat(OpenAIBaseModel):
    type: Literal["structural_tag"]
    format: Any


AnyStructuralTagResponseFormat: TypeAlias = (
    LegacyStructuralTagResponseFormat | StructuralTagResponseFormat
)


225
class ResponseFormat(OpenAIBaseModel):
226
    # type must be "json_schema", "json_object", or "text"
227
    type: Literal["text", "json_object", "json_schema"]
228
    json_schema: JsonSchemaResponseFormat | None = None
229
230


231
232
233
AnyResponseFormat: TypeAlias = (
    ResponseFormat | StructuralTagResponseFormat | LegacyStructuralTagResponseFormat
)
234
235


236
class StreamOptions(OpenAIBaseModel):
237
238
    include_usage: bool | None = True
    continuous_usage_stats: bool | None = False
239
240


241
242
class FunctionDefinition(OpenAIBaseModel):
    name: str
243
244
    description: str | None = None
    parameters: dict[str, Any] | None = None
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260


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


class ChatCompletionNamedFunction(OpenAIBaseModel):
    name: str


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


261
262
# extra="forbid" is a workaround to have kwargs as a field,
# see https://github.com/pydantic/pydantic/issues/3125
263
264
class LogitsProcessorConstructor(BaseModel):
    qualname: str
265
266
    args: list[Any] | None = None
    kwargs: dict[str, Any] | None = None
267

268
269
    model_config = ConfigDict(extra="forbid")

270

271
LogitsProcessors = list[str | LogitsProcessorConstructor]
272
273


274
def get_logits_processors(
275
276
    processors: LogitsProcessors | None, pattern: str | None
) -> list[Any] | None:
277
278
279
    if processors and pattern:
        logits_processors = []
        for processor in processors:
280
            qualname = processor if isinstance(processor, str) else processor.qualname
281
282
283
284
            if not re.match(pattern, qualname):
                raise ValueError(
                    f"Logits processor '{qualname}' is not allowed by this "
                    "server. See --logits-processor-pattern engine argument "
285
286
                    "for more information."
                )
287
288
289
290
291
292
293
            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):
294
295
296
                logits_processor = logits_processor(
                    *processor.args or [], **processor.kwargs or {}
                )
297
298
299
300
301
            logits_processors.append(logits_processor)
        return logits_processors
    elif processors:
        raise ValueError(
            "The `logits_processors` argument is not supported by this "
302
            "server. See --logits-processor-pattern engine argument "
303
304
            "for more information."
        )
305
306
307
    return None


308
309
310
ResponseInputOutputItem: TypeAlias = (
    ResponseInputItemParam | ResponseReasoningItem | ResponseFunctionToolCall
)
311
312


313
314
315
class ResponsesRequest(OpenAIBaseModel):
    # Ordered by official OpenAI API documentation
    # https://platform.openai.com/docs/api-reference/responses/create
316
317
    background: bool | None = False
    include: (
318
319
320
321
322
323
324
325
326
327
        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",
            ],
        ]
328
329
330
331
332
333
334
335
336
337
338
339
        | None
    ) = None
    input: str | list[ResponseInputOutputItem]
    instructions: str | None = None
    max_output_tokens: int | None = None
    max_tool_calls: int | None = None
    metadata: Metadata | None = None
    model: str | None = None
    parallel_tool_calls: bool | None = True
    previous_response_id: str | None = None
    prompt: ResponsePrompt | None = None
    reasoning: Reasoning | None = None
340
    service_tier: Literal["auto", "default", "flex", "scale", "priority"] = "auto"
341
342
343
344
    store: bool | None = True
    stream: bool | None = False
    temperature: float | None = None
    text: ResponseTextConfig | None = None
345
346
    tool_choice: ToolChoice = "auto"
    tools: list[Tool] = Field(default_factory=list)
347
348
349
350
    top_logprobs: int | None = 0
    top_p: float | None = None
    truncation: Literal["auto", "disabled"] | None = "disabled"
    user: str | None = None
351
352
353
354
355
356
357

    # --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 "
358
359
            "through out the inference process and return in response."
        ),
360
    )
361
    mm_processor_kwargs: dict[str, Any] | None = Field(
362
363
364
365
366
367
368
369
        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 "
370
371
            "if the served model does not use priority scheduling."
        ),
372
    )
373
    cache_salt: str | None = Field(
374
375
376
377
378
379
380
        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 "
381
382
383
            "to 256 bit). Not supported by vLLM engine V0."
        ),
    )
384
385
386
387
388

    enable_response_messages: bool = Field(
        default=False,
        description=(
            "Dictates whether or not to return messages as part of the "
389
            "response object. Currently only supported for"
390
391
392
            "non-background and gpt-oss only. "
        ),
    )
393
394
395
396
397
    # similar to input_messages / output_messages in ResponsesResponse
    # we take in previous_input_messages (ie in harmony format)
    # this cannot be used in conjunction with previous_response_id
    # TODO: consider supporting non harmony messages as well
    previous_input_messages: list[OpenAIHarmonyMessage | dict] | None = None
398
399
400
401
402
403
404
405
406
407
    # --8<-- [end:responses-extra-params]

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

    def to_sampling_params(
        self,
        default_max_tokens: int,
408
        default_sampling_params: dict | None = None,
409
410
411
412
413
414
415
416
417
    ) -> 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(
418
419
                "temperature", self._DEFAULT_SAMPLING_PARAMS["temperature"]
            )
420
421
        if (top_p := self.top_p) is None:
            top_p = default_sampling_params.get(
422
423
                "top_p", self._DEFAULT_SAMPLING_PARAMS["top_p"]
            )
424
        stop_token_ids = default_sampling_params.get("stop_token_ids")
425
426

        # Structured output
427
        structured_outputs = None
428
429
        if self.text is not None and self.text.format is not None:
            response_format = self.text.format
430
431
432
433
            if (
                response_format.type == "json_schema"
                and response_format.schema_ is not None
            ):
434
                structured_outputs = StructuredOutputsParams(
435
436
                    json=response_format.schema_
                )
437
438
439
440
441
442
443
444
            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,
445
            logprobs=self.top_logprobs if self.is_include_output_logprobs() else None,
446
            stop_token_ids=stop_token_ids,
447
448
449
            output_kind=(
                RequestOutputKind.DELTA if self.stream else RequestOutputKind.FINAL_ONLY
            ),
450
            structured_outputs=structured_outputs,
451
452
        )

453
454
455
456
    def is_include_output_logprobs(self) -> bool:
        """Check if the request includes output logprobs."""
        if self.include is None:
            return False
457
458
459
460
        return (
            isinstance(self.include, list)
            and "message.output_text.logprobs" in self.include
        )
461

462
463
464
465
466
    @model_validator(mode="before")
    def validate_background(cls, data):
        if not data.get("background"):
            return data
        if not data.get("store", True):
467
            raise ValueError("background can only be used when `store` is true")
468
469
470
471
472
473
474
475
        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

476
477
478
479
480
481
    @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 "
482
483
484
485
486
487
                    "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."
                )
488
489
        return data

490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
    @model_validator(mode="before")
    def function_call_parsing(cls, data):
        """Parse function_call dictionaries into ResponseFunctionToolCall objects.
        This ensures Pydantic can properly resolve union types in the input field.
        Function calls provided as dicts are converted to ResponseFunctionToolCall
        objects before validation, while invalid structures are left for Pydantic
        to reject with appropriate error messages.
        """

        input_data = data.get("input")

        # Early return for None, strings, or bytes
        # (strings are iterable but shouldn't be processed)
        if input_data is None or isinstance(input_data, (str, bytes)):
            return data

        # Convert iterators (like ValidatorIterator) to list
        if not isinstance(input_data, list):
            try:
                input_data = list(input_data)
            except TypeError:
                # Not iterable, leave as-is for Pydantic to handle
                return data

        processed_input = []
        for item in input_data:
            if isinstance(item, dict) and item.get("type") == "function_call":
                try:
                    processed_input.append(ResponseFunctionToolCall(**item))
                except ValidationError:
                    # Let Pydantic handle validation for malformed function calls
                    logger.debug(
                        "Failed to parse function_call to ResponseFunctionToolCall, "
                        "leaving for Pydantic validation"
                    )
                    processed_input.append(item)
            else:
                processed_input.append(item)

        data["input"] = processed_input
        return data

532

533
class ChatCompletionRequest(OpenAIBaseModel):
534
535
    # Ordered by official OpenAI API documentation
    # https://platform.openai.com/docs/api-reference/chat/create
536
    messages: list[ChatCompletionMessageParam]
537
538
539
540
541
542
    model: str | None = None
    frequency_penalty: float | None = 0.0
    logit_bias: dict[str, float] | None = None
    logprobs: bool | None = False
    top_logprobs: int | None = 0
    max_tokens: int | None = Field(
543
        default=None,
544
545
        deprecated="max_tokens is deprecated in favor of "
        "the max_completion_tokens field",
546
    )
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
    max_completion_tokens: int | None = None
    n: int | None = 1
    presence_penalty: float | None = 0.0
    response_format: AnyResponseFormat | None = None
    seed: int | None = Field(None, ge=_LONG_INFO.min, le=_LONG_INFO.max)
    stop: str | list[str] | None = []
    stream: bool | None = False
    stream_options: StreamOptions | None = None
    temperature: float | None = None
    top_p: float | None = None
    tools: list[ChatCompletionToolsParam] | None = None
    tool_choice: (
        Literal["none"]
        | Literal["auto"]
        | Literal["required"]
        | ChatCompletionNamedToolChoiceParam
        | None
    ) = "none"
    reasoning_effort: Literal["low", "medium", "high"] | None = None
566
    include_reasoning: bool = True
567

568
    # NOTE this will be ignored by vLLM -- the model determines the behavior
569
570
    parallel_tool_calls: bool | None = False
    user: str | None = None
571

572
    # --8<-- [start:chat-completion-sampling-params]
573
    best_of: int | None = None
574
    use_beam_search: bool = False
575
576
577
    top_k: int | None = None
    min_p: float | None = None
    repetition_penalty: float | None = None
578
    length_penalty: float = 1.0
579
    stop_token_ids: list[int] | None = []
580
581
582
583
584
    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
585
586
587
    truncate_prompt_tokens: Annotated[int, Field(ge=-1)] | None = None
    prompt_logprobs: int | None = None
    allowed_token_ids: list[int] | None = None
588
    bad_words: list[str] = Field(default_factory=list)
589
    # --8<-- [end:chat-completion-sampling-params]
590

591
    # --8<-- [start:chat-completion-extra-params]
592
    echo: bool = Field(
593
594
595
        default=False,
        description=(
            "If true, the new message will be prepended with the last message "
596
597
            "if they belong to the same role."
        ),
598
    )
599
    add_generation_prompt: bool = Field(
600
        default=True,
601
602
603
604
605
        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."
        ),
606
    )
607
608
    continue_final_message: bool = Field(
        default=False,
609
610
611
612
613
614
615
        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`."
        ),
616
    )
617
    add_special_tokens: bool = Field(
618
619
620
621
622
        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 "
623
            "special tokens so this should be set to false (as is the "
624
625
            "default)."
        ),
626
    )
627
    documents: list[dict[str, str]] | None = Field(
628
        default=None,
629
630
631
632
633
634
635
        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.'
        ),
636
    )
637
    chat_template: str | None = Field(
638
639
640
        default=None,
        description=(
            "A Jinja template to use for this conversion. "
641
642
            "As of transformers v4.44, default chat template is no longer "
            "allowed, so you must provide a chat template if the tokenizer "
643
644
            "does not define one."
        ),
645
    )
646
    chat_template_kwargs: dict[str, Any] | None = Field(
647
        default=None,
648
649
        description=(
            "Additional keyword args to pass to the template renderer. "
650
651
            "Will be accessible by the chat template."
        ),
652
    )
653
    mm_processor_kwargs: dict[str, Any] | None = Field(
654
655
656
        default=None,
        description=("Additional kwargs to pass to the HF processor."),
    )
657
    structured_outputs: StructuredOutputsParams | None = Field(
658
        default=None,
659
        description="Additional kwargs for structured outputs",
660
    )
661
    guided_json: str | dict | BaseModel | None = Field(
662
663
664
665
        default=None,
        description=(
            "`guided_json` is deprecated. "
            "This will be removed in v0.12.0 or v1.0.0, whichever is soonest. "
666
667
            "Please pass `json` to `structured_outputs` instead."
        ),
668
    )
669
    guided_regex: str | None = Field(
670
671
672
673
        default=None,
        description=(
            "`guided_regex` is deprecated. "
            "This will be removed in v0.12.0 or v1.0.0, whichever is soonest. "
674
675
            "Please pass `regex` to `structured_outputs` instead."
        ),
676
    )
677
    guided_choice: list[str] | None = Field(
678
679
680
681
        default=None,
        description=(
            "`guided_choice` is deprecated. "
            "This will be removed in v0.12.0 or v1.0.0, whichever is soonest. "
682
683
            "Please pass `choice` to `structured_outputs` instead."
        ),
684
    )
685
    guided_grammar: str | None = Field(
686
687
688
689
        default=None,
        description=(
            "`guided_grammar` is deprecated. "
            "This will be removed in v0.12.0 or v1.0.0, whichever is soonest. "
690
691
            "Please pass `grammar` to `structured_outputs` instead."
        ),
692
    )
693
    structural_tag: str | None = Field(
694
695
696
697
        default=None,
        description=(
            "`structural_tag` is deprecated. "
            "This will be removed in v0.12.0 or v1.0.0, whichever is soonest. "
698
699
            "Please pass `structural_tag` to `structured_outputs` instead."
        ),
700
    )
701
    guided_decoding_backend: str | None = Field(
702
703
704
705
        default=None,
        description=(
            "`guided_decoding_backend` is deprecated. "
            "This will be removed in v0.12.0 or v1.0.0, whichever is soonest. "
706
707
            "Please remove it from your request."
        ),
708
    )
709
    guided_whitespace_pattern: str | None = Field(
710
711
712
713
714
715
716
        default=None,
        description=(
            "`guided_whitespace_pattern` is deprecated. "
            "This will be removed in v0.12.0 or v1.0.0, whichever is soonest. "
            "Please pass `whitespace_pattern` to `structured_outputs` instead."
        ),
    )
717
718
719
720
721
    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 "
722
723
            "if the served model does not use priority scheduling."
        ),
724
    )
725
726
727
728
729
    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 "
730
731
            "through out the inference process and return in response."
        ),
732
    )
733
    logits_processors: LogitsProcessors | None = Field(
734
735
736
737
738
739
740
741
742
        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': "
743
744
745
            "{'param': 'value'}}."
        ),
    )
746
    return_tokens_as_token_ids: bool | None = Field(
747
748
749
750
        default=None,
        description=(
            "If specified with 'logprobs', tokens are represented "
            " as strings of the form 'token_id:{token_id}' so that tokens "
751
752
753
            "that are not JSON-encodable can be identified."
        ),
    )
754
    return_token_ids: bool | None = Field(
755
756
757
758
759
760
        default=None,
        description=(
            "If specified, the result will include token IDs alongside the "
            "generated text. In streaming mode, prompt_token_ids is included "
            "only in the first chunk, and token_ids contains the delta tokens "
            "for each chunk. This is useful for debugging or when you "
761
762
763
            "need to map generated text back to input tokens."
        ),
    )
764
    cache_salt: str | None = Field(
765
766
767
768
769
770
771
        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 "
772
773
774
            "to 256 bit). Not supported by vLLM engine V0."
        ),
    )
775
    kv_transfer_params: dict[str, Any] | None = Field(
Robert Shaw's avatar
Robert Shaw committed
776
        default=None,
777
778
        description="KVTransfer parameters used for disaggregated serving.",
    )
779

780
    vllm_xargs: dict[str, str | int | float] | None = Field(
781
        default=None,
782
783
784
785
        description=(
            "Additional request parameters with string or "
            "numeric values, used by custom extensions."
        ),
786
787
    )

788
    # --8<-- [end:chat-completion-extra-params]
Zhuohan Li's avatar
Zhuohan Li committed
789

790
791
792
793
794
    # Default sampling parameters for chat completion requests
    _DEFAULT_SAMPLING_PARAMS: dict = {
        "repetition_penalty": 1.0,
        "temperature": 1.0,
        "top_p": 1.0,
795
        "top_k": 0,
796
797
798
799
        "min_p": 0.0,
    }

    def to_beam_search_params(
800
801
        self, max_tokens: int, default_sampling_params: dict
    ) -> BeamSearchParams:
802
        n = self.n if self.n is not None else 1
803
804
        if (temperature := self.temperature) is None:
            temperature = default_sampling_params.get(
805
806
                "temperature", self._DEFAULT_SAMPLING_PARAMS["temperature"]
            )
807
808
809
810
811
812

        return BeamSearchParams(
            beam_width=n,
            max_tokens=max_tokens,
            ignore_eos=self.ignore_eos,
            temperature=temperature,
813
            length_penalty=self.length_penalty,
814
815
            include_stop_str_in_output=self.include_stop_str_in_output,
        )
816

817
    def to_sampling_params(
818
        self,
819
        max_tokens: int,
820
        logits_processor_pattern: str | None,
821
        default_sampling_params: dict,
822
    ) -> SamplingParams:
823
824
825
826
827
828
829
830
        # 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(
831
832
                "temperature", self._DEFAULT_SAMPLING_PARAMS["temperature"]
            )
833
834
        if (top_p := self.top_p) is None:
            top_p = default_sampling_params.get(
835
836
                "top_p", self._DEFAULT_SAMPLING_PARAMS["top_p"]
            )
837
838
        if (top_k := self.top_k) is None:
            top_k = default_sampling_params.get(
839
840
                "top_k", self._DEFAULT_SAMPLING_PARAMS["top_k"]
            )
841
842
        if (min_p := self.min_p) is None:
            min_p = default_sampling_params.get(
843
844
                "min_p", self._DEFAULT_SAMPLING_PARAMS["min_p"]
            )
845

846
847
848
849
        prompt_logprobs = self.prompt_logprobs
        if prompt_logprobs is None and self.echo:
            prompt_logprobs = self.top_logprobs

850
851
852
853
854
855
856
857
858
859
860
861
862
863
        # Forward deprecated guided_* parameters to structured_outputs
        if self.structured_outputs is None:
            kwargs = dict[str, Any](
                json=self.guided_json,
                regex=self.guided_regex,
                choice=self.guided_choice,
                grammar=self.guided_grammar,
                whitespace_pattern=self.guided_whitespace_pattern,
                structural_tag=self.structural_tag,
            )
            kwargs = {k: v for k, v in kwargs.items() if v is not None}
            if len(kwargs) > 0:
                self.structured_outputs = StructuredOutputsParams(**kwargs)

864
        response_format = self.response_format
865
        if response_format is not None:
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
            # If structured outputs wasn't already enabled,
            # we must enable it for these features to work
            if self.structured_outputs is None:
                self.structured_outputs = StructuredOutputsParams()

            # Set structured output params for response format
            if response_format is not None:
                if response_format.type == "json_object":
                    self.structured_outputs.json_object = True
                elif response_format.type == "json_schema":
                    json_schema = response_format.json_schema
                    assert json_schema is not None
                    self.structured_outputs.json = json_schema.json_schema
                elif response_format.type == "structural_tag":
                    structural_tag = response_format
                    assert structural_tag is not None and isinstance(
882
883
884
885
886
                        structural_tag,
                        (
                            LegacyStructuralTagResponseFormat,
                            StructuralTagResponseFormat,
                        ),
887
                    )
888
                    s_tag_obj = structural_tag.model_dump(by_alias=True)
889
                    self.structured_outputs.structural_tag = json.dumps(s_tag_obj)
890

891
892
893
894
        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
895
        return SamplingParams.from_optional(
896
            n=self.n,
897
            best_of=self.best_of,
898
899
            presence_penalty=self.presence_penalty,
            frequency_penalty=self.frequency_penalty,
900
901
902
903
904
            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
905
            seed=self.seed,
906
907
            stop=self.stop,
            stop_token_ids=self.stop_token_ids,
908
            logprobs=self.top_logprobs if self.logprobs else None,
909
            prompt_logprobs=prompt_logprobs,
910
            ignore_eos=self.ignore_eos,
911
            max_tokens=max_tokens,
912
            min_tokens=self.min_tokens,
913
914
            skip_special_tokens=self.skip_special_tokens,
            spaces_between_special_tokens=self.spaces_between_special_tokens,
915
916
917
            logits_processors=get_logits_processors(
                self.logits_processors, logits_processor_pattern
            ),
918
            include_stop_str_in_output=self.include_stop_str_in_output,
919
            truncate_prompt_tokens=self.truncate_prompt_tokens,
920
921
922
            output_kind=RequestOutputKind.DELTA
            if self.stream
            else RequestOutputKind.FINAL_ONLY,
923
            structured_outputs=self.structured_outputs,
Robert Shaw's avatar
Robert Shaw committed
924
            logit_bias=self.logit_bias,
925
            bad_words=self.bad_words,
926
            allowed_token_ids=self.allowed_token_ids,
927
928
            extra_args=extra_args or None,
        )
929

930
    @model_validator(mode="before")
931
    @classmethod
932
933
    def validate_stream_options(cls, data):
        if data.get("stream_options") and not data.get("stream"):
934
            raise ValueError("Stream options can only be defined when `stream=True`.")
935
936
937
938
939
940
941

        return data

    @model_validator(mode="before")
    @classmethod
    def check_logprobs(cls, data):
        if (prompt_logprobs := data.get("prompt_logprobs")) is not None:
942
            if data.get("stream") and (prompt_logprobs > 0 or prompt_logprobs == -1):
943
                raise ValueError(
944
945
                    "`prompt_logprobs` are not available when `stream=True`."
                )
946

947
            if prompt_logprobs < 0 and prompt_logprobs != -1:
948
                raise ValueError("`prompt_logprobs` must be a positive value or -1.")
949
            if prompt_logprobs == -1 and not envs.VLLM_USE_V1:
950
951
952
                raise ValueError(
                    "`prompt_logprobs=-1` is only supported with vLLM engine V1."
                )
953
        if (top_logprobs := data.get("top_logprobs")) is not None:
954
            if top_logprobs < 0 and top_logprobs != -1:
955
                raise ValueError("`top_logprobs` must be a positive value or -1.")
956

957
            if (top_logprobs == -1 or top_logprobs > 0) and not data.get("logprobs"):
958
959
960
961
962
                raise ValueError(
                    "when using `top_logprobs`, `logprobs` must be set to true."
                )

        return data
963

964
965
    @model_validator(mode="before")
    @classmethod
966
    def check_structured_outputs_count(cls, data):
967
968
969
        if isinstance(data, ValueError):
            raise data

970
        if data.get("structured_outputs", None) is None:
971
972
            return data

973
        structured_outputs_kwargs = data["structured_outputs"]
974
975
        count = sum(
            structured_outputs_kwargs.get(k) is not None
976
977
            for k in ("json", "regex", "choice")
        )
978
979
        # you can only use one kind of constraints for structured outputs
        if count > 1:
980
            raise ValueError(
981
                "You can only use one kind of constraints for structured "
982
983
                "outputs ('json', 'regex' or 'choice')."
            )
984
985
        # you can only either use structured outputs or tools, not both
        if count > 1 and data.get("tool_choice", "none") not in (
986
987
988
            "none",
            "auto",
            "required",
989
        ):
990
            raise ValueError(
991
                "You can only either use constraints for structured outputs "
992
993
                "or tools, not both."
            )
994
995
996
997
        return data

    @model_validator(mode="before")
    @classmethod
998
999
1000
    def check_tool_usage(cls, data):
        # if "tool_choice" is not specified but tools are provided,
        # default to "auto" tool_choice
1001
        if "tool_choice" not in data and data.get("tools"):
1002
1003
            data["tool_choice"] = "auto"

1004
        # if "tool_choice" is "none" -- no validation is needed for tools
1005
1006
1007
        if "tool_choice" in data and data["tool_choice"] == "none":
            return data

1008
        # if "tool_choice" is specified -- validation
1009
        if "tool_choice" in data and data["tool_choice"] is not None:
1010
            # ensure that if "tool choice" is specified, tools are present
1011
            if "tools" not in data or data["tools"] is None:
1012
                raise ValueError("When using `tool_choice`, `tools` must be set.")
1013
1014

            # make sure that tool choice is either a named tool
1015
            # OR that it's set to "auto" or "required"
1016
1017
1018
            if data["tool_choice"] not in ["auto", "required"] and not isinstance(
                data["tool_choice"], dict
            ):
1019
                raise ValueError(
1020
1021
1022
                    f"Invalid value for `tool_choice`: {data['tool_choice']}! "
                    'Only named tools, "none", "auto" or "required" '
                    "are supported."
1023
                )
1024

1025
1026
1027
            # if tool_choice is "required" but the "tools" list is empty,
            # override the data to behave like "none" to align with
            # OpenAI’s behavior.
1028
1029
1030
1031
1032
            if (
                data["tool_choice"] == "required"
                and isinstance(data["tools"], list)
                and len(data["tools"]) == 0
            ):
1033
1034
1035
1036
                data["tool_choice"] = "none"
                del data["tools"]
                return data

1037
1038
            # ensure that if "tool_choice" is specified as an object,
            # it matches a valid tool
1039
1040
            correct_usage_message = (
                'Correct usage: `{"type": "function",'
1041
                ' "function": {"name": "my_function"}}`'
1042
            )
1043
1044
            if isinstance(data["tool_choice"], dict):
                valid_tool = False
1045
1046
                function = data["tool_choice"].get("function")
                if not isinstance(function, dict):
1047
                    raise ValueError(
1048
                        f"Invalid value for `function`: `{function}` in "
1049
1050
                        f"`tool_choice`! {correct_usage_message}"
                    )
1051
                if "name" not in function:
1052
1053
1054
1055
                    raise ValueError(
                        f"Expected field `name` in `function` in "
                        f"`tool_choice`! {correct_usage_message}"
                    )
1056
                function_name = function["name"]
1057
                if not isinstance(function_name, str) or len(function_name) == 0:
1058
                    raise ValueError(
1059
                        f"Invalid `name` in `function`: `{function_name}`"
1060
1061
                        f" in `tool_choice`! {correct_usage_message}"
                    )
1062
                for tool in data["tools"]:
1063
                    if tool["function"]["name"] == function_name:
1064
1065
1066
1067
1068
                        valid_tool = True
                        break
                if not valid_tool:
                    raise ValueError(
                        "The tool specified in `tool_choice` does not match any"
1069
1070
                        " of the specified `tools`"
                    )
1071
1072
        return data

1073
1074
1075
    @model_validator(mode="before")
    @classmethod
    def check_generation_prompt(cls, data):
1076
1077
1078
1079
1080
        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."
            )
1081
1082
        return data

1083
1084
1085
1086
1087
1088
1089
    @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 "
1090
1091
1092
1093
1094
1095
                    "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."
                )
1096
1097
        return data

Zhuohan Li's avatar
Zhuohan Li committed
1098

1099
class CompletionRequest(OpenAIBaseModel):
1100
1101
    # Ordered by official OpenAI API documentation
    # https://platform.openai.com/docs/api-reference/completions/create
1102
1103
1104
1105
1106
1107
1108
1109
    model: str | None = None
    prompt: list[int] | list[list[int]] | str | list[str] | None = None
    best_of: int | None = None
    echo: bool | None = False
    frequency_penalty: float | None = 0.0
    logit_bias: dict[str, float] | None = None
    logprobs: int | None = None
    max_tokens: int | None = 16
1110
    n: int = 1
1111
1112
1113
1114
1115
1116
1117
1118
1119
    presence_penalty: float | None = 0.0
    seed: int | None = Field(None, ge=_LONG_INFO.min, le=_LONG_INFO.max)
    stop: str | list[str] | None = []
    stream: bool | None = False
    stream_options: StreamOptions | None = None
    suffix: str | None = None
    temperature: float | None = None
    top_p: float | None = None
    user: str | None = None
1120

1121
    # --8<-- [start:completion-sampling-params]
1122
    use_beam_search: bool = False
1123
1124
1125
    top_k: int | None = None
    min_p: float | None = None
    repetition_penalty: float | None = None
1126
    length_penalty: float = 1.0
1127
    stop_token_ids: list[int] | None = []
1128
1129
1130
1131
1132
    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
1133
1134
1135
    truncate_prompt_tokens: Annotated[int, Field(ge=-1)] | None = None
    allowed_token_ids: list[int] | None = None
    prompt_logprobs: int | None = None
1136
    # --8<-- [end:completion-sampling-params]
1137

1138
    # --8<-- [start:completion-extra-params]
1139
    prompt_embeds: bytes | list[bytes] | None = None
1140
1141
    add_special_tokens: bool = Field(
        default=True,
1142
        description=(
1143
            "If true (the default), special tokens (e.g. BOS) will be added to "
1144
1145
            "the prompt."
        ),
1146
    )
1147
    response_format: AnyResponseFormat | None = Field(
1148
        default=None,
1149
1150
1151
1152
1153
        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."
        ),
1154
    )
1155
    structured_outputs: StructuredOutputsParams | None = Field(
1156
        default=None,
1157
        description="Additional kwargs for structured outputs",
1158
    )
1159
    guided_json: str | dict | BaseModel | None = Field(
1160
1161
1162
1163
        default=None,
        description=(
            "`guided_json` is deprecated. "
            "This will be removed in v0.12.0 or v1.0.0, whichever is soonest. "
1164
1165
            "Please pass `json` to `structured_outputs` instead."
        ),
1166
    )
1167
    guided_regex: str | None = Field(
1168
1169
1170
1171
        default=None,
        description=(
            "`guided_regex` is deprecated. "
            "This will be removed in v0.12.0 or v1.0.0, whichever is soonest. "
1172
1173
            "Please pass `regex` to `structured_outputs` instead."
        ),
1174
    )
1175
    guided_choice: list[str] | None = Field(
1176
1177
1178
1179
        default=None,
        description=(
            "`guided_choice` is deprecated. "
            "This will be removed in v0.12.0 or v1.0.0, whichever is soonest. "
1180
1181
            "Please pass `choice` to `structured_outputs` instead."
        ),
1182
    )
1183
    guided_grammar: str | None = Field(
1184
1185
1186
1187
        default=None,
        description=(
            "`guided_grammar` is deprecated. "
            "This will be removed in v0.12.0 or v1.0.0, whichever is soonest. "
1188
1189
            "Please pass `grammar` to `structured_outputs` instead."
        ),
1190
    )
1191
1192
1193
1194
    structural_tag: str | None = Field(
        default=None,
        description=("If specified, the output will follow the structural tag schema."),
    )
1195
    guided_decoding_backend: str | None = Field(
1196
1197
1198
1199
        default=None,
        description=(
            "`guided_decoding_backend` is deprecated. "
            "This will be removed in v0.12.0 or v1.0.0, whichever is soonest. "
1200
1201
            "Please remove it from your request."
        ),
1202
    )
1203
    guided_whitespace_pattern: str | None = Field(
1204
1205
1206
1207
1208
1209
1210
        default=None,
        description=(
            "`guided_whitespace_pattern` is deprecated. "
            "This will be removed in v0.12.0 or v1.0.0, whichever is soonest. "
            "Please pass `whitespace_pattern` to `structured_outputs` instead."
        ),
    )
1211
1212
1213
1214
1215
    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 "
1216
1217
            "if the served model does not use priority scheduling."
        ),
1218
    )
1219
1220
1221
1222
1223
    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 "
1224
1225
            "through out the inference process and return in response."
        ),
1226
    )
1227
    logits_processors: LogitsProcessors | None = Field(
1228
1229
1230
1231
1232
1233
1234
1235
1236
        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': "
1237
1238
1239
            "{'param': 'value'}}."
        ),
    )
1240

1241
    return_tokens_as_token_ids: bool | None = Field(
1242
1243
1244
1245
        default=None,
        description=(
            "If specified with 'logprobs', tokens are represented "
            " as strings of the form 'token_id:{token_id}' so that tokens "
1246
1247
1248
            "that are not JSON-encodable can be identified."
        ),
    )
1249
    return_token_ids: bool | None = Field(
1250
1251
1252
1253
1254
1255
        default=None,
        description=(
            "If specified, the result will include token IDs alongside the "
            "generated text. In streaming mode, prompt_token_ids is included "
            "only in the first chunk, and token_ids contains the delta tokens "
            "for each chunk. This is useful for debugging or when you "
1256
1257
1258
            "need to map generated text back to input tokens."
        ),
    )
1259

1260
    cache_salt: str | None = Field(
1261
1262
1263
1264
1265
1266
1267
        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 "
1268
1269
1270
            "to 256 bit). Not supported by vLLM engine V0."
        ),
    )
1271

1272
    kv_transfer_params: dict[str, Any] | None = Field(
Robert Shaw's avatar
Robert Shaw committed
1273
        default=None,
1274
1275
        description="KVTransfer parameters used for disaggregated serving.",
    )
Robert Shaw's avatar
Robert Shaw committed
1276

1277
    vllm_xargs: dict[str, str | int | float] | None = Field(
1278
        default=None,
1279
1280
1281
1282
        description=(
            "Additional request parameters with string or "
            "numeric values, used by custom extensions."
        ),
1283
1284
    )

1285
    # --8<-- [end:completion-extra-params]
Zhuohan Li's avatar
Zhuohan Li committed
1286

1287
1288
1289
1290
1291
    # Default sampling parameters for completion requests
    _DEFAULT_SAMPLING_PARAMS: dict = {
        "repetition_penalty": 1.0,
        "temperature": 1.0,
        "top_p": 1.0,
1292
        "top_k": 0,
1293
1294
1295
1296
        "min_p": 0.0,
    }

    def to_beam_search_params(
1297
1298
        self,
        max_tokens: int,
1299
        default_sampling_params: dict | None = None,
1300
1301
1302
    ) -> BeamSearchParams:
        if default_sampling_params is None:
            default_sampling_params = {}
1303
        n = self.n if self.n is not None else 1
1304
1305
1306

        if (temperature := self.temperature) is None:
            temperature = default_sampling_params.get("temperature", 1.0)
1307
1308
1309
1310
1311
1312

        return BeamSearchParams(
            beam_width=n,
            max_tokens=max_tokens,
            ignore_eos=self.ignore_eos,
            temperature=temperature,
1313
            length_penalty=self.length_penalty,
1314
1315
            include_stop_str_in_output=self.include_stop_str_in_output,
        )
1316

1317
    def to_sampling_params(
1318
        self,
1319
        max_tokens: int,
1320
1321
        logits_processor_pattern: str | None,
        default_sampling_params: dict | None = None,
1322
    ) -> SamplingParams:
1323
1324
        if default_sampling_params is None:
            default_sampling_params = {}
1325

1326
1327
1328
1329
1330
1331
1332
1333
        # 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(
1334
1335
                "temperature", self._DEFAULT_SAMPLING_PARAMS["temperature"]
            )
1336
1337
        if (top_p := self.top_p) is None:
            top_p = default_sampling_params.get(
1338
1339
                "top_p", self._DEFAULT_SAMPLING_PARAMS["top_p"]
            )
1340
1341
        if (top_k := self.top_k) is None:
            top_k = default_sampling_params.get(
1342
1343
                "top_k", self._DEFAULT_SAMPLING_PARAMS["top_k"]
            )
1344
1345
        if (min_p := self.min_p) is None:
            min_p = default_sampling_params.get(
1346
1347
                "min_p", self._DEFAULT_SAMPLING_PARAMS["min_p"]
            )
1348

1349
1350
1351
1352
        prompt_logprobs = self.prompt_logprobs
        if prompt_logprobs is None and self.echo:
            prompt_logprobs = self.logprobs

1353
1354
        echo_without_generation = self.echo and self.max_tokens == 0

1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
        guided_json_object = None
        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
            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)

1371
1372
1373
1374
        # Forward deprecated guided_* parameters to structured_outputs
        if self.structured_outputs is None:
            kwargs = dict[str, Any](
                json=self.guided_json,
1375
                json_object=guided_json_object,
1376
1377
1378
1379
1380
1381
1382
1383
1384
                regex=self.guided_regex,
                choice=self.guided_choice,
                grammar=self.guided_grammar,
                whitespace_pattern=self.guided_whitespace_pattern,
            )
            kwargs = {k: v for k, v in kwargs.items() if v is not None}
            if len(kwargs) > 0:
                self.structured_outputs = StructuredOutputsParams(**kwargs)

1385
1386
1387
1388
        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
1389
        return SamplingParams.from_optional(
1390
            n=self.n,
1391
            best_of=self.best_of,
1392
1393
            presence_penalty=self.presence_penalty,
            frequency_penalty=self.frequency_penalty,
1394
1395
1396
1397
1398
            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
1399
            seed=self.seed,
1400
1401
            stop=self.stop,
            stop_token_ids=self.stop_token_ids,
1402
            logprobs=self.logprobs,
1403
            ignore_eos=self.ignore_eos,
1404
            max_tokens=max_tokens if not echo_without_generation else 1,
1405
            min_tokens=self.min_tokens,
1406
            prompt_logprobs=prompt_logprobs,
1407
            skip_special_tokens=self.skip_special_tokens,
1408
            spaces_between_special_tokens=self.spaces_between_special_tokens,
1409
            include_stop_str_in_output=self.include_stop_str_in_output,
1410
1411
1412
            logits_processors=get_logits_processors(
                self.logits_processors, logits_processor_pattern
            ),
1413
            truncate_prompt_tokens=self.truncate_prompt_tokens,
1414
1415
1416
            output_kind=RequestOutputKind.DELTA
            if self.stream
            else RequestOutputKind.FINAL_ONLY,
1417
            structured_outputs=self.structured_outputs,
1418
            logit_bias=self.logit_bias,
Robert Shaw's avatar
Robert Shaw committed
1419
            allowed_token_ids=self.allowed_token_ids,
1420
            extra_args=extra_args or None,
1421
        )
1422

1423
1424
    @model_validator(mode="before")
    @classmethod
1425
    def check_structured_outputs_count(cls, data):
1426
        if data.get("structured_outputs", None) is None:
1427
1428
            return data

1429
        structured_outputs_kwargs = data["structured_outputs"]
1430
1431
        count = sum(
            structured_outputs_kwargs.get(k) is not None
1432
1433
            for k in ("json", "regex", "choice")
        )
1434
        if count > 1:
1435
            raise ValueError(
1436
                "You can only use one kind of constraints for structured "
1437
1438
                "outputs ('json', 'regex' or 'choice')."
            )
1439
1440
        return data

1441
1442
1443
    @model_validator(mode="before")
    @classmethod
    def check_logprobs(cls, data):
1444
        if (prompt_logprobs := data.get("prompt_logprobs")) is not None:
1445
            if data.get("stream") and (prompt_logprobs > 0 or prompt_logprobs == -1):
1446
                raise ValueError(
1447
1448
                    "`prompt_logprobs` are not available when `stream=True`."
                )
1449

1450
            if prompt_logprobs < 0 and prompt_logprobs != -1:
1451
                raise ValueError("`prompt_logprobs` must be a positive value or -1.")
1452
            if prompt_logprobs == -1 and not envs.VLLM_USE_V1:
1453
1454
1455
                raise ValueError(
                    "`prompt_logprobs=-1` is only supported with vLLM engine V1."
                )
1456
1457
1458
        if (logprobs := data.get("logprobs")) is not None and logprobs < 0:
            raise ValueError("`logprobs` must be a positive value.")

1459
1460
        return data

1461
1462
1463
1464
    @model_validator(mode="before")
    @classmethod
    def validate_stream_options(cls, data):
        if data.get("stream_options") and not data.get("stream"):
1465
            raise ValueError("Stream options can only be defined when `stream=True`.")
1466

1467
1468
        return data

1469
1470
1471
    @model_validator(mode="before")
    @classmethod
    def validate_prompt_and_prompt_embeds(cls, data):
1472
1473
1474
        prompt = data.get("prompt")
        prompt_embeds = data.get("prompt_embeds")

1475
1476
1477
1478
        prompt_is_empty = prompt is None or (isinstance(prompt, str) and prompt == "")
        embeds_is_empty = prompt_embeds is None or (
            isinstance(prompt_embeds, list) and len(prompt_embeds) == 0
        )
1479
1480

        if prompt_is_empty and embeds_is_empty:
1481
            raise ValueError(
1482
1483
1484
                "Either prompt or prompt_embeds must be provided and non-empty."
            )

1485
1486
        return data

1487
1488
1489
1490
1491
1492
1493
    @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 "
1494
1495
1496
1497
1498
1499
                    "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."
                )
1500
1501
        return data

Zhuohan Li's avatar
Zhuohan Li committed
1502

1503
class EmbeddingCompletionRequest(OpenAIBaseModel):
1504
1505
    # Ordered by official OpenAI API documentation
    # https://platform.openai.com/docs/api-reference/embeddings
1506
1507
    model: str | None = None
    input: list[int] | list[list[int]] | str | list[str]
1508
    encoding_format: EncodingFormat = "float"
1509
1510
1511
    dimensions: int | None = None
    user: str | None = None
    truncate_prompt_tokens: Annotated[int, Field(ge=-1)] | None = None
1512

1513
    # --8<-- [start:embedding-extra-params]
1514
1515
1516
1517
    add_special_tokens: bool = Field(
        default=True,
        description=(
            "If true (the default), special tokens (e.g. BOS) will be added to "
1518
1519
            "the prompt."
        ),
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 "
1526
1527
            "if the served model does not use priority scheduling."
        ),
1528
    )
1529
1530
1531
1532
1533
    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 "
1534
1535
            "through out the inference process and return in response."
        ),
1536
    )
1537
1538
1539
1540
    normalize: bool | None = Field(
        default=None,
        description="Whether to normalize the embeddings outputs. Default is True.",
    )
1541
    embed_dtype: EmbedDType = Field(
1542
1543
        default="float32",
        description=(
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
            "What dtype to use for encoding. Default to using float32 for base64 "
            "encoding to match the OpenAI python client behavior. "
            "This parameter will affect base64 and binary_response."
        ),
    )
    endianness: Endianness = Field(
        default="native",
        description=(
            "What endianness to use for encoding. Default to using native for "
            "base64 encoding to match the OpenAI python client behavior."
            "This parameter will affect base64 and binary_response."
1555
1556
        ),
    )
1557
    # --8<-- [end:embedding-extra-params]
1558

1559
    def to_pooling_params(self):
1560
1561
1562
        return PoolingParams(
            truncate_prompt_tokens=self.truncate_prompt_tokens,
            dimensions=self.dimensions,
1563
1564
            normalize=self.normalize,
        )
1565
1566


1567
class EmbeddingChatRequest(OpenAIBaseModel):
1568
    model: str | None = None
1569
    messages: list[ChatCompletionMessageParam]
1570

1571
    encoding_format: EncodingFormat = "float"
1572
1573
1574
    dimensions: int | None = None
    user: str | None = None
    truncate_prompt_tokens: Annotated[int, Field(ge=-1)] | None = None
1575

1576
    # --8<-- [start:chat-embedding-extra-params]
1577
1578
    add_generation_prompt: bool = Field(
        default=False,
1579
1580
1581
1582
1583
        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."
        ),
1584
1585
    )

1586
1587
1588
1589
1590
1591
1592
    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 "
1593
1594
            "default)."
        ),
1595
    )
1596
    chat_template: str | None = Field(
1597
1598
1599
1600
1601
        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 "
1602
1603
            "does not define one."
        ),
1604
    )
1605
    chat_template_kwargs: dict[str, Any] | None = Field(
1606
        default=None,
1607
1608
        description=(
            "Additional keyword args to pass to the template renderer. "
1609
1610
            "Will be accessible by the chat template."
        ),
1611
    )
1612
    mm_processor_kwargs: dict[str, Any] | None = Field(
1613
1614
1615
        default=None,
        description=("Additional kwargs to pass to the HF processor."),
    )
1616
1617
1618
1619
1620
    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 "
1621
1622
            "if the served model does not use priority scheduling."
        ),
1623
    )
1624
1625
1626
1627
1628
    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 "
1629
1630
            "through out the inference process and return in response."
        ),
1631
    )
1632
1633
1634
1635
    normalize: bool | None = Field(
        default=None,
        description="Whether to normalize the embeddings outputs. Default is True.",
    )
1636
    embed_dtype: EmbedDType = Field(
1637
1638
        default="float32",
        description=(
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
            "What dtype to use for encoding. Default to using float32 for base64 "
            "encoding to match the OpenAI python client behavior. "
            "This parameter will affect base64 and binary_response."
        ),
    )
    endianness: Endianness = Field(
        default="native",
        description=(
            "What endianness to use for encoding. Default to using native for "
            "base64 encoding to match the OpenAI python client behavior."
            "This parameter will affect base64 and binary_response."
1650
1651
        ),
    )
1652
    # --8<-- [end:chat-embedding-extra-params]
1653
1654
1655
1656

    @model_validator(mode="before")
    @classmethod
    def check_generation_prompt(cls, data):
1657
1658
1659
1660
1661
        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."
            )
1662
1663
1664
        return data

    def to_pooling_params(self):
1665
1666
1667
        return PoolingParams(
            truncate_prompt_tokens=self.truncate_prompt_tokens,
            dimensions=self.dimensions,
1668
1669
            normalize=self.normalize,
        )
1670
1671


1672
EmbeddingRequest: TypeAlias = EmbeddingCompletionRequest | EmbeddingChatRequest
1673

1674
1675
1676
1677
1678
1679
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

class PoolingCompletionRequest(EmbeddingCompletionRequest):
    task: PoolingTask | None = None
    softmax: bool | None = Field(
        default=None,
        description="softmax will be deprecated, please use use_activation instead.",
    )
    activation: bool | None = Field(
        default=None,
        description="activation will be deprecated, please use use_activation instead.",
    )
    use_activation: bool | None = Field(
        default=None,
        description="Whether to use activation for classification outputs. "
        "If it is a classify or token_classify task, the default is True; "
        "for other tasks, this value should be None.",
    )

    def to_pooling_params(self):
        return PoolingParams(
            truncate_prompt_tokens=self.truncate_prompt_tokens,
            dimensions=self.dimensions,
            normalize=self.normalize,
            use_activation=get_use_activation(self),
        )


class PoolingChatRequest(EmbeddingChatRequest):
    task: PoolingTask | None = None
    softmax: bool | None = Field(
        default=None,
        description="softmax will be deprecated, please use use_activation instead.",
    )
    activation: bool | None = Field(
        default=None,
        description="activation will be deprecated, please use use_activation instead.",
    )
    use_activation: bool | None = Field(
        default=None,
        description="Whether to use activation for classification outputs. "
        "If it is a classify or token_classify task, the default is True; "
        "for other tasks, this value should be None.",
    )

    def to_pooling_params(self):
        return PoolingParams(
            truncate_prompt_tokens=self.truncate_prompt_tokens,
            dimensions=self.dimensions,
            normalize=self.normalize,
            use_activation=get_use_activation(self),
        )

1726
1727
1728
1729
1730

T = TypeVar("T")


class IOProcessorRequest(OpenAIBaseModel, Generic[T]):
1731
    model: str | None = None
1732
1733
1734
1735
1736
1737
1738
1739
1740

    priority: int = Field(default=0)
    """
    The priority of the request (lower means earlier handling;
    default: 0). Any priority other than 0 will raise an error
    if the served model does not use priority scheduling.
    """
    data: T

1741
    task: PoolingTask = "plugin"
1742
1743
    encoding_format: EncodingFormat = "float"
    embed_dtype: EmbedDType = Field(
1744
1745
        default="float32",
        description=(
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
            "What dtype to use for encoding. Default to using float32 for base64 "
            "encoding to match the OpenAI python client behavior. "
            "This parameter will affect base64 and binary_response."
        ),
    )
    endianness: Endianness = Field(
        default="native",
        description=(
            "What endianness to use for encoding. Default to using native for "
            "base64 encoding to match the OpenAI python client behavior."
            "This parameter will affect base64 and binary_response."
1757
1758
1759
        ),
    )

1760
    def to_pooling_params(self):
1761
        return PoolingParams()
1762
1763
1764


class IOProcessorResponse(OpenAIBaseModel, Generic[T]):
1765
    request_id: str | None = None
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
    """
    The request_id associated with this response
    """
    created_at: int = Field(default_factory=lambda: int(time.time()))

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


1778
1779
1780
PoolingRequest: TypeAlias = (
    PoolingCompletionRequest | PoolingChatRequest | IOProcessorRequest
)
1781

1782

1783
class ScoreRequest(OpenAIBaseModel):
1784
1785
1786
1787
    model: str | None = None
    text_1: list[str] | str | ScoreMultiModalParam
    text_2: list[str] | str | ScoreMultiModalParam
    truncate_prompt_tokens: Annotated[int, Field(ge=-1)] | None = None
1788

1789
    # --8<-- [start:score-extra-params]
1790

1791
    mm_processor_kwargs: dict[str, Any] | None = Field(
1792
1793
1794
1795
        default=None,
        description=("Additional kwargs to pass to the HF processor."),
    )

1796
1797
1798
1799
1800
    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 "
1801
1802
            "if the served model does not use priority scheduling."
        ),
1803
    )
1804

1805
1806
1807
1808
    softmax: bool | None = Field(
        default=None,
        description="softmax will be deprecated, please use use_activation instead.",
    )
1809

1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
    activation: bool | None = Field(
        default=None,
        description="activation will be deprecated, please use use_activation instead.",
    )

    use_activation: bool | None = Field(
        default=None,
        description="Whether to use activation for classification outputs. "
        "Default is True.",
    )
1820
    # --8<-- [end:score-extra-params]
1821

1822
    def to_pooling_params(self):
1823
1824
        return PoolingParams(
            truncate_prompt_tokens=self.truncate_prompt_tokens,
1825
            use_activation=get_use_activation(self),
1826
        )
1827
1828


1829
class RerankRequest(OpenAIBaseModel):
1830
1831
1832
    model: str | None = None
    query: str | ScoreMultiModalParam
    documents: list[str] | ScoreMultiModalParam
1833
    top_n: int = Field(default_factory=lambda: 0)
1834
    truncate_prompt_tokens: Annotated[int, Field(ge=-1)] | None = None
1835

1836
    # --8<-- [start:rerank-extra-params]
1837

1838
    mm_processor_kwargs: dict[str, Any] | None = Field(
1839
1840
1841
1842
        default=None,
        description=("Additional kwargs to pass to the HF processor."),
    )

1843
1844
1845
1846
1847
    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 "
1848
1849
            "if the served model does not use priority scheduling."
        ),
1850
    )
1851

1852
1853
1854
1855
    softmax: bool | None = Field(
        default=None,
        description="softmax will be deprecated, please use use_activation instead.",
    )
1856

1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
    activation: bool | None = Field(
        default=None,
        description="activation will be deprecated, please use use_activation instead.",
    )

    use_activation: bool | None = Field(
        default=None,
        description="Whether to use activation for classification outputs. "
        "Default is True.",
    )
1867
    # --8<-- [end:rerank-extra-params]
1868

1869
    def to_pooling_params(self):
1870
1871
        return PoolingParams(
            truncate_prompt_tokens=self.truncate_prompt_tokens,
1872
            use_activation=get_use_activation(self),
1873
        )
1874
1875
1876


class RerankDocument(BaseModel):
1877
1878
    text: str | None = None
    multi_modal: ScoreContentPartParam | None = None
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894


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
1895
    results: list[RerankResult]
1896
1897


1898
class CompletionLogProbs(OpenAIBaseModel):
1899
    text_offset: list[int] = Field(default_factory=list)
1900
    token_logprobs: list[float | None] = Field(default_factory=list)
1901
    tokens: list[str] = Field(default_factory=list)
1902
    top_logprobs: list[dict[str, float] | None] = Field(default_factory=list)
Zhuohan Li's avatar
Zhuohan Li committed
1903
1904


1905
class CompletionResponseChoice(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
1906
1907
    index: int
    text: str
1908
1909
1910
    logprobs: CompletionLogProbs | None = None
    finish_reason: str | None = None
    stop_reason: int | str | None = Field(
1911
1912
1913
1914
        default=None,
        description=(
            "The stop string or token id that caused the completion "
            "to stop, None if the completion finished for some other reason "
1915
1916
            "including encountering the EOS token"
        ),
1917
    )
1918
1919
1920
    token_ids: list[int] | None = None  # For response
    prompt_logprobs: list[dict[int, Logprob] | None] | None = None
    prompt_token_ids: list[int] | None = None  # For prompt
Zhuohan Li's avatar
Zhuohan Li committed
1921
1922


1923
class CompletionResponse(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
1924
    id: str = Field(default_factory=lambda: f"cmpl-{random_uuid()}")
1925
    object: Literal["text_completion"] = "text_completion"
Zhuohan Li's avatar
Zhuohan Li committed
1926
1927
    created: int = Field(default_factory=lambda: int(time.time()))
    model: str
1928
    choices: list[CompletionResponseChoice]
1929
1930
    service_tier: Literal["auto", "default", "flex", "scale", "priority"] | None = None
    system_fingerprint: str | None = None
Zhuohan Li's avatar
Zhuohan Li committed
1931
    usage: UsageInfo
1932
1933

    # vLLM-specific fields that are not in OpenAI spec
1934
    kv_transfer_params: dict[str, Any] | None = Field(
1935
1936
        default=None, description="KVTransfer parameters."
    )
Zhuohan Li's avatar
Zhuohan Li committed
1937
1938


1939
class CompletionResponseStreamChoice(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
1940
1941
    index: int
    text: str
1942
1943
1944
    logprobs: CompletionLogProbs | None = None
    finish_reason: str | None = None
    stop_reason: int | str | None = Field(
1945
1946
1947
1948
        default=None,
        description=(
            "The stop string or token id that caused the completion "
            "to stop, None if the completion finished for some other reason "
1949
1950
            "including encountering the EOS token"
        ),
1951
    )
1952
1953
    # not part of the OpenAI spec but for tracing the tokens
    # prompt tokens is put into choice to align with CompletionResponseChoice
1954
1955
    prompt_token_ids: list[int] | None = None
    token_ids: list[int] | None = None
Zhuohan Li's avatar
Zhuohan Li committed
1956
1957


1958
class CompletionStreamResponse(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
1959
1960
1961
1962
    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
1963
    choices: list[CompletionResponseStreamChoice]
1964
    usage: UsageInfo | None = Field(default=None)
1965
1966


1967
class EmbeddingResponseData(OpenAIBaseModel):
1968
1969
    index: int
    object: str = "embedding"
1970
    embedding: list[float] | str
1971
1972


1973
class EmbeddingResponse(OpenAIBaseModel):
1974
    id: str = Field(default_factory=lambda: f"embd-{random_uuid()}")
1975
1976
1977
    object: str = "list"
    created: int = Field(default_factory=lambda: int(time.time()))
    model: str
1978
    data: list[EmbeddingResponseData]
1979
1980
1981
    usage: UsageInfo


1982
1983
1984
1985
1986
1987
class EmbeddingBytesResponse(OpenAIBaseModel):
    body: list[bytes]
    metadata: str
    media_type: str = "application/octet-stream"


1988
1989
1990
class PoolingResponseData(OpenAIBaseModel):
    index: int
    object: str = "pooling"
1991
    data: list[list[float]] | list[float] | str
1992
1993
1994
1995
1996
1997
1998


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
1999
    data: list[PoolingResponseData]
2000
2001
2002
    usage: UsageInfo


2003
2004
2005
2006
2007
2008
class PoolingBytesResponse(OpenAIBaseModel):
    body: list[bytes]
    metadata: str
    media_type: str = "application/octet-stream"


2009
2010
2011
class ScoreResponseData(OpenAIBaseModel):
    index: int
    object: str = "score"
2012
    score: float
2013
2014
2015
2016
2017
2018
2019


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
2020
    data: list[ScoreResponseData]
2021
2022
2023
    usage: UsageInfo


2024
class ClassificationRequest(OpenAIBaseModel):
2025
2026
2027
2028
    model: str | None = None
    input: list[str] | str
    truncate_prompt_tokens: int | None = None
    user: str | None = None
2029

2030
    # --8<-- [start:classification-extra-params]
2031
2032
2033
2034
2035
    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 "
2036
2037
            "if the served model does not use priority scheduling."
        ),
2038
2039
    )

2040
2041
2042
2043
2044
2045
2046
2047
2048
    softmax: bool | None = Field(
        default=None,
        description="softmax will be deprecated, please use use_activation instead.",
    )

    activation: bool | None = Field(
        default=None,
        description="activation will be deprecated, please use use_activation instead.",
    )
2049

2050
2051
2052
2053
2054
    use_activation: bool | None = Field(
        default=None,
        description="Whether to use activation for classification outputs. "
        "Default is True.",
    )
2055
    # --8<-- [end:classification-extra-params]
2056
2057

    def to_pooling_params(self):
2058
2059
        return PoolingParams(
            truncate_prompt_tokens=self.truncate_prompt_tokens,
2060
            use_activation=get_use_activation(self),
2061
        )
2062
2063
2064
2065


class ClassificationData(OpenAIBaseModel):
    index: int
2066
    label: str | None
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
    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


2080
2081
2082
2083
2084
2085
class FunctionCall(OpenAIBaseModel):
    name: str
    arguments: str


class ToolCall(OpenAIBaseModel):
2086
    id: str = Field(default_factory=make_tool_call_id)
2087
2088
2089
2090
    type: Literal["function"] = "function"
    function: FunctionCall


2091
class DeltaFunctionCall(BaseModel):
2092
2093
    name: str | None = None
    arguments: str | None = None
2094
2095
2096
2097


# a tool call delta where everything is optional
class DeltaToolCall(OpenAIBaseModel):
2098
2099
    id: str | None = None
    type: Literal["function"] | None = None
2100
    index: int
2101
    function: DeltaFunctionCall | None = None
2102
2103
2104
2105
2106
2107
2108


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

    # extracted tool calls
2109
    tool_calls: list[ToolCall]
2110
2111
2112

    # content - per OpenAI spec, content AND tool calls can be returned rarely
    # But some models will do this intentionally
2113
    content: str | None = None
2114
2115


2116
class ChatMessage(OpenAIBaseModel):
2117
    role: str
2118
2119
2120
2121
2122
    content: str | None = None
    refusal: str | None = None
    annotations: OpenAIAnnotation | None = None
    audio: OpenAIChatCompletionAudio | None = None
    function_call: FunctionCall | None = None
2123
    tool_calls: list[ToolCall] = Field(default_factory=list)
2124

2125
    # vLLM-specific fields that are not in OpenAI spec
2126
    reasoning_content: str | None = None
2127

2128

2129
2130
2131
class ChatCompletionLogProb(OpenAIBaseModel):
    token: str
    logprob: float = -9999.0
2132
    bytes: list[int] | None = None
2133
2134
2135


class ChatCompletionLogProbsContent(ChatCompletionLogProb):
2136
2137
    # Workaround: redefine fields name cache so that it's not
    # shared with the super class.
2138
    field_names: ClassVar[set[str] | None] = None
2139
    top_logprobs: list[ChatCompletionLogProb] = Field(default_factory=list)
2140
2141
2142


class ChatCompletionLogProbs(OpenAIBaseModel):
2143
    content: list[ChatCompletionLogProbsContent] | None = None
2144
2145


2146
class ChatCompletionResponseChoice(OpenAIBaseModel):
2147
2148
    index: int
    message: ChatMessage
2149
    logprobs: ChatCompletionLogProbs | None = None
2150
    # per OpenAI spec this is the default
2151
    finish_reason: str | None = "stop"
2152
    # not part of the OpenAI spec but included in vLLM for legacy reasons
2153
    stop_reason: int | str | None = None
2154
2155
    # not part of the OpenAI spec but is useful for tracing the tokens
    # in agent scenarios
2156
    token_ids: list[int] | None = None
2157
2158


2159
class ChatCompletionResponse(OpenAIBaseModel):
2160
    id: str = Field(default_factory=lambda: f"chatcmpl-{random_uuid()}")
2161
    object: Literal["chat.completion"] = "chat.completion"
2162
2163
    created: int = Field(default_factory=lambda: int(time.time()))
    model: str
2164
    choices: list[ChatCompletionResponseChoice]
2165
2166
    service_tier: Literal["auto", "default", "flex", "scale", "priority"] | None = None
    system_fingerprint: str | None = None
2167
    usage: UsageInfo
2168
2169

    # vLLM-specific fields that are not in OpenAI spec
2170
2171
2172
    prompt_logprobs: list[dict[int, Logprob] | None] | None = None
    prompt_token_ids: list[int] | None = None
    kv_transfer_params: dict[str, Any] | None = Field(
2173
2174
        default=None, description="KVTransfer parameters."
    )
2175
2176


2177
class DeltaMessage(OpenAIBaseModel):
2178
2179
2180
    role: str | None = None
    content: str | None = None
    reasoning_content: str | None = None
2181
    tool_calls: list[DeltaToolCall] = Field(default_factory=list)
2182
2183


2184
class ChatCompletionResponseStreamChoice(OpenAIBaseModel):
2185
2186
    index: int
    delta: DeltaMessage
2187
2188
2189
    logprobs: ChatCompletionLogProbs | None = None
    finish_reason: str | None = None
    stop_reason: int | str | None = None
2190
    # not part of the OpenAI spec but for tracing the tokens
2191
    token_ids: list[int] | None = None
2192
2193


2194
class ChatCompletionStreamResponse(OpenAIBaseModel):
2195
    id: str = Field(default_factory=lambda: f"chatcmpl-{random_uuid()}")
2196
    object: Literal["chat.completion.chunk"] = "chat.completion.chunk"
2197
2198
    created: int = Field(default_factory=lambda: int(time.time()))
    model: str
2199
    choices: list[ChatCompletionResponseStreamChoice]
2200
    usage: UsageInfo | None = Field(default=None)
2201
    # not part of the OpenAI spec but for tracing the tokens
2202
    prompt_token_ids: list[int] | None = None
2203
2204


2205
2206
class TranscriptionResponseStreamChoice(OpenAIBaseModel):
    delta: DeltaMessage
2207
2208
    finish_reason: str | None = None
    stop_reason: int | str | None = None
2209
2210
2211
2212
2213
2214
2215
2216


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]
2217
    usage: UsageInfo | None = Field(default=None)
2218
2219


2220
2221
class InputTokensDetails(OpenAIBaseModel):
    cached_tokens: int
2222
2223
    input_tokens_per_turn: list[int] = Field(default_factory=list)
    cached_tokens_per_turn: list[int] = Field(default_factory=list)
2224
2225
2226


class OutputTokensDetails(OpenAIBaseModel):
2227
2228
    reasoning_tokens: int = 0
    tool_output_tokens: int = 0
2229
2230
    output_tokens_per_turn: list[int] = Field(default_factory=list)
    tool_output_tokens_per_turn: list[int] = Field(default_factory=list)
2231
2232
2233
2234
2235
2236
2237
2238


class ResponseUsage(OpenAIBaseModel):
    input_tokens: int
    input_tokens_details: InputTokensDetails
    output_tokens: int
    output_tokens_details: OutputTokensDetails
    total_tokens: int
2239
2240


2241
2242
2243
2244
2245
2246
def serialize_message(msg):
    """
    Serializes a single message
    """
    if isinstance(msg, dict):
        return msg
2247
    elif hasattr(msg, "to_dict"):
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
        return msg.to_dict()
    else:
        # fallback to pyandic dump
        return msg.model_dump_json()


def serialize_messages(msgs):
    """
    Serializes multiple messages
    """
    return [serialize_message(msg) for msg in msgs] if msgs else None


2261
2262
2263
2264
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
2265
2266
2267
    incomplete_details: IncompleteDetails | None = None
    instructions: str | None = None
    metadata: Metadata | None = None
2268
2269
    model: str
    object: Literal["response"] = "response"
2270
    output: list[ResponseOutputItem]
2271
2272
2273
2274
2275
2276
2277
    parallel_tool_calls: bool
    temperature: float
    tool_choice: ToolChoice
    tools: list[Tool]
    top_p: float
    background: bool
    max_output_tokens: int
2278
2279
2280
2281
    max_tool_calls: int | None = None
    previous_response_id: str | None = None
    prompt: ResponsePrompt | None = None
    reasoning: Reasoning | None = None
2282
2283
    service_tier: Literal["auto", "default", "flex", "scale", "priority"]
    status: ResponseStatus
2284
2285
    text: ResponseTextConfig | None = None
    top_logprobs: int | None = None
2286
    truncation: Literal["auto", "disabled"]
2287
2288
    usage: ResponseUsage | None = None
    user: str | None = None
2289

2290
2291
2292
2293
    # --8<-- [start:responses-extra-params]
    # These are populated when enable_response_messages is set to True
    # NOTE: custom serialization is needed
    # see serialize_input_messages and serialize_output_messages
2294
2295
    input_messages: list[ChatCompletionMessageParam] | None = None
    output_messages: list[ChatCompletionMessageParam] | None = None
2296
2297
2298
2299
2300
2301
2302
    # --8<-- [end:responses-extra-params]

    # NOTE: openAI harmony doesn't serialize TextContent properly,
    # TODO: this fixes for TextContent, but need to verify for tools etc
    # https://github.com/openai/harmony/issues/78
    @field_serializer("output_messages", when_used="json")
    def serialize_output_messages(self, msgs, _info):
2303
        return serialize_messages(msgs)
2304
2305
2306
2307
2308

    # NOTE: openAI harmony doesn't serialize TextContent properly, this fixes it
    # https://github.com/openai/harmony/issues/78
    @field_serializer("input_messages", when_used="json")
    def serialize_input_messages(self, msgs, _info):
2309
        return serialize_messages(msgs)
2310

2311
2312
2313
2314
2315
2316
2317
2318
2319
    @classmethod
    def from_request(
        cls,
        request: ResponsesRequest,
        sampling_params: SamplingParams,
        model_name: str,
        created_time: int,
        output: list[ResponseOutputItem],
        status: ResponseStatus,
2320
2321
2322
        usage: ResponseUsage | None = None,
        input_messages: list[ChatCompletionMessageParam] | None = None,
        output_messages: list[ChatCompletionMessageParam] | None = None,
2323
    ) -> "ResponsesResponse":
2324
        incomplete_details: IncompleteDetails | None = None
2325
2326
        if status == "incomplete":
            incomplete_details = IncompleteDetails(reason="max_output_tokens")
2327
2328
2329
        # TODO: implement the other reason for incomplete_details,
        # which is content_filter
        # incomplete_details = IncompleteDetails(reason='content_filter')
2330
2331
2332
        return cls(
            id=request.request_id,
            created_at=created_time,
2333
            incomplete_details=incomplete_details,
2334
2335
2336
2337
            instructions=request.instructions,
            metadata=request.metadata,
            model=model_name,
            output=output,
2338
2339
            input_messages=input_messages,
            output_messages=output_messages,
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
2350
2351
2352
2353
2354
2355
2356
2357
2358
2359
2360
            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,
        )


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
2388
2389
2390
2391
2392
2393
2394
2395
2396
2397
2398
2399
2400
2401
2402
2403
2404
# TODO: this code can be removed once
# https://github.com/openai/openai-python/issues/2634 has been resolved
class ResponseReasoningPartDoneEvent(OpenAIBaseModel):
    content_index: int
    """The index of the content part that is done."""

    item_id: str
    """The ID of the output item that the content part was added to."""

    output_index: int
    """The index of the output item that the content part was added to."""

    part: ResponseReasoningTextContent
    """The content part that is done."""

    sequence_number: int
    """The sequence number of this event."""

    type: Literal["response.reasoning_part.done"]
    """The type of the event. Always `response.reasoning_part.done`."""


# TODO: this code can be removed once
# https://github.com/openai/openai-python/issues/2634 has been resolved
class ResponseReasoningPartAddedEvent(OpenAIBaseModel):
    content_index: int
    """The index of the content part that is done."""

    item_id: str
    """The ID of the output item that the content part was added to."""

    output_index: int
    """The index of the output item that the content part was added to."""

    part: ResponseReasoningTextContent
    """The content part that is done."""

    sequence_number: int
    """The sequence number of this event."""

    type: Literal["response.reasoning_part.added"]
    """The type of the event. Always `response.reasoning_part.added`."""


2405
2406
2407
2408
2409
2410
2411
2412
2413
2414
2415
2416
2417
2418
# vLLM Streaming Events
# Note: we override the response type with the vLLM ResponsesResponse type
class ResponseCompletedEvent(OpenAIResponseCompletedEvent):
    response: ResponsesResponse  # type: ignore[override]


class ResponseCreatedEvent(OpenAIResponseCreatedEvent):
    response: ResponsesResponse  # type: ignore[override]


class ResponseInProgressEvent(OpenAIResponseInProgressEvent):
    response: ResponsesResponse  # type: ignore[override]


2419
2420
2421
2422
2423
2424
2425
2426
2427
2428
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
2439
StreamingResponsesResponse: TypeAlias = (
    ResponseCreatedEvent
    | ResponseInProgressEvent
    | ResponseCompletedEvent
    | ResponseOutputItemAddedEvent
    | ResponseOutputItemDoneEvent
    | ResponseContentPartAddedEvent
    | ResponseContentPartDoneEvent
    | ResponseReasoningTextDeltaEvent
    | ResponseReasoningTextDoneEvent
    | ResponseReasoningPartAddedEvent
    | ResponseReasoningPartDoneEvent
    | ResponseCodeInterpreterCallInProgressEvent
    | ResponseCodeInterpreterCallCodeDeltaEvent
    | ResponseWebSearchCallInProgressEvent
    | ResponseWebSearchCallSearchingEvent
    | ResponseWebSearchCallCompletedEvent
    | ResponseCodeInterpreterCallCodeDoneEvent
    | ResponseCodeInterpreterCallInterpretingEvent
    | ResponseCodeInterpreterCallCompletedEvent
)
2440

2441
2442
2443
BatchRequestInputBody: TypeAlias = (
    ChatCompletionRequest | EmbeddingRequest | ScoreRequest | RerankRequest
)
2444
2445


2446
2447
2448
2449
2450
2451
2452
2453
2454
2455
2456
2457
2458
2459
2460
2461
2462
2463
2464
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

2465
    # The parameters of the request.
2466
    body: BatchRequestInputBody
2467

2468
    @field_validator("body", mode="plain")
2469
2470
2471
    @classmethod
    def check_type_for_url(cls, value: Any, info: ValidationInfo):
        # Use url to disambiguate models
2472
        url: str = info.data["url"]
2473
2474
2475
2476
        if url == "/v1/chat/completions":
            return ChatCompletionRequest.model_validate(value)
        if url == "/v1/embeddings":
            return TypeAdapter(EmbeddingRequest).validate_python(value)
2477
        if url.endswith("/score"):
2478
            return ScoreRequest.model_validate(value)
2479
2480
2481
        if url.endswith("/rerank"):
            return RerankRequest.model_validate(value)
        return TypeAdapter(BatchRequestInputBody).validate_python(value)
2482

2483

2484
2485
2486
2487
2488
2489
2490
2491
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.
2492
2493
2494
2495
2496
2497
2498
    body: (
        ChatCompletionResponse
        | EmbeddingResponse
        | ScoreResponse
        | RerankResponse
        | None
    ) = None
2499
2500


2501
2502
2503
2504
2505
2506
2507
2508
2509
2510
2511
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

2512
    response: BatchResponseData | None
2513
2514
2515

    # For requests that failed with a non-HTTP error, this will contain more
    # information on the cause of the failure.
2516
    error: Any | None
2517
2518


2519
class TokenizeCompletionRequest(OpenAIBaseModel):
2520
    model: str | None = None
2521
2522
    prompt: str

2523
2524
2525
2526
    add_special_tokens: bool = Field(
        default=True,
        description=(
            "If true (the default), special tokens (e.g. BOS) will be added to "
2527
2528
            "the prompt."
        ),
2529
    )
2530
    return_token_strs: bool | None = Field(
2531
        default=False,
2532
2533
2534
        description=(
            "If true, also return the token strings corresponding to the token ids."
        ),
2535
    )
2536
2537
2538


class TokenizeChatRequest(OpenAIBaseModel):
2539
    model: str | None = None
2540
    messages: list[ChatCompletionMessageParam]
2541

2542
2543
    add_generation_prompt: bool = Field(
        default=True,
2544
2545
2546
2547
2548
        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."
        ),
2549
    )
2550
    return_token_strs: bool | None = Field(
2551
        default=False,
2552
2553
2554
        description=(
            "If true, also return the token strings corresponding to the token ids."
        ),
2555
    )
2556
2557
    continue_final_message: bool = Field(
        default=False,
2558
2559
2560
2561
2562
2563
2564
        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`."
        ),
2565
2566
2567
2568
2569
2570
2571
2572
    )
    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 "
2573
2574
            "default)."
        ),
2575
    )
2576
    chat_template: str | None = Field(
2577
2578
2579
2580
2581
        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 "
2582
2583
            "does not define one."
        ),
2584
    )
2585
    chat_template_kwargs: dict[str, Any] | None = Field(
2586
        default=None,
2587
2588
        description=(
            "Additional keyword args to pass to the template renderer. "
2589
2590
            "Will be accessible by the chat template."
        ),
2591
    )
2592
    mm_processor_kwargs: dict[str, Any] | None = Field(
2593
2594
2595
        default=None,
        description=("Additional kwargs to pass to the HF processor."),
    )
2596
    tools: list[ChatCompletionToolsParam] | None = Field(
2597
2598
2599
        default=None,
        description=("A list of tools the model may call."),
    )
2600

2601
2602
2603
    @model_validator(mode="before")
    @classmethod
    def check_generation_prompt(cls, data):
2604
2605
2606
2607
2608
        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."
            )
2609
2610
        return data

2611

2612
TokenizeRequest: TypeAlias = TokenizeCompletionRequest | TokenizeChatRequest
2613
2614
2615
2616
2617


class TokenizeResponse(OpenAIBaseModel):
    count: int
    max_model_len: int
2618
    tokens: list[int]
2619
    token_strs: list[str] | None = None
2620
2621
2622


class DetokenizeRequest(OpenAIBaseModel):
2623
    model: str | None = None
2624
    tokens: list[int]
2625
2626
2627
2628


class DetokenizeResponse(OpenAIBaseModel):
    prompt: str
2629
2630


2631
2632
class TokenizerInfoResponse(OpenAIBaseModel):
    """
2633
    Response containing tokenizer configuration
2634
2635
2636
2637
2638
2639
2640
    equivalent to tokenizer_config.json
    """

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


2641
class LoadLoRAAdapterRequest(BaseModel):
2642
2643
2644
2645
    lora_name: str
    lora_path: str


2646
class UnloadLoRAAdapterRequest(BaseModel):
2647
    lora_name: str
2648
    lora_int_id: int | None = Field(default=None)
2649
2650
2651


## Protocols for Audio
2652
AudioResponseFormat: TypeAlias = Literal["json", "text", "srt", "verbose_json", "vtt"]
2653
2654
2655
2656


class TranscriptionRequest(OpenAIBaseModel):
    # Ordered by official OpenAI API documentation
2657
    # https://platform.openai.com/docs/api-reference/audio/createTranscription
2658
2659
2660
2661
2662
2663
2664

    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.
    """

2665
    model: str | None = None
2666
2667
2668
    """ID of the model to use.
    """

2669
    language: str | None = None
2670
2671
2672
2673
2674
2675
2676
2677
2678
2679
2680
2681
2682
2683
2684
2685
2686
2687
2688
2689
2690
2691
2692
    """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 !!

2693
    timestamp_granularities: list[Literal["word", "segment"]] = Field(
2694
2695
        alias="timestamp_granularities[]", default=[]
    )
2696
2697
2698
2699
2700
2701
2702
2703
    """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.
    """

2704
    stream: bool | None = False
2705
    """When set, it will enable output to be streamed in a similar fashion
2706
    as the Chat Completion endpoint.
2707
    """
2708
    # --8<-- [start:transcription-extra-params]
2709
    # Flattened stream option to simplify form data.
2710
2711
    stream_include_usage: bool | None = False
    stream_continuous_usage_stats: bool | None = False
2712

2713
    vllm_xargs: dict[str, str | int | float] | None = Field(
2714
        default=None,
2715
2716
2717
2718
        description=(
            "Additional request parameters with string or "
            "numeric values, used by custom extensions."
        ),
2719
    )
2720
    # --8<-- [end:transcription-extra-params]
2721

2722
    to_language: str | None = None
2723
2724
    """The language of the output audio we transcribe to.

2725
    Please note that this is not currently used by supported models at this
2726
2727
2728
    time, but it is a placeholder for future use, matching translation api.
    """

2729
    # --8<-- [start:transcription-sampling-params]
2730
2731
2732
2733
2734
2735
2736
2737
2738
    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.
    """

2739
    top_p: float | None = None
2740
    """Enables nucleus (top-p) sampling, where tokens are selected from the
2741
2742
2743
    smallest possible set whose cumulative probability exceeds `p`.
    """

2744
    top_k: int | None = None
2745
2746
    """Limits sampling to the `k` most probable tokens at each step."""

2747
    min_p: float | None = None
2748
    """Filters out tokens with a probability lower than `min_p`, ensuring a
2749
2750
2751
    minimum likelihood threshold during sampling.
    """

2752
    seed: int | None = Field(None, ge=_LONG_INFO.min, le=_LONG_INFO.max)
2753
2754
    """The seed to use for sampling."""

2755
    frequency_penalty: float | None = 0.0
2756
2757
    """The frequency penalty to use for sampling."""

2758
    repetition_penalty: float | None = None
2759
2760
    """The repetition penalty to use for sampling."""

2761
    presence_penalty: float | None = 0.0
2762
    """The presence penalty to use for sampling."""
2763
    # --8<-- [end:transcription-sampling-params]
2764

2765
2766
    # Default sampling parameters for transcription requests.
    _DEFAULT_SAMPLING_PARAMS: dict = {
2767
2768
2769
        "repetition_penalty": 1.0,
        "temperature": 1.0,
        "top_p": 1.0,
2770
        "top_k": 0,
2771
        "min_p": 0.0,
2772
2773
2774
    }

    def to_sampling_params(
2775
        self, default_max_tokens: int, default_sampling_params: dict | None = None
2776
    ) -> SamplingParams:
2777
2778
2779
2780
        max_tokens = default_max_tokens

        if default_sampling_params is None:
            default_sampling_params = {}
2781

2782
2783
2784
        # Default parameters
        if (temperature := self.temperature) is None:
            temperature = default_sampling_params.get(
2785
2786
                "temperature", self._DEFAULT_SAMPLING_PARAMS["temperature"]
            )
2787
2788
        if (top_p := self.top_p) is None:
            top_p = default_sampling_params.get(
2789
2790
                "top_p", self._DEFAULT_SAMPLING_PARAMS["top_p"]
            )
2791
2792
        if (top_k := self.top_k) is None:
            top_k = default_sampling_params.get(
2793
2794
                "top_k", self._DEFAULT_SAMPLING_PARAMS["top_k"]
            )
2795
2796
        if (min_p := self.min_p) is None:
            min_p = default_sampling_params.get(
2797
2798
                "min_p", self._DEFAULT_SAMPLING_PARAMS["min_p"]
            )
2799
2800
2801
2802

        if (repetition_penalty := self.repetition_penalty) is None:
            repetition_penalty = default_sampling_params.get(
                "repetition_penalty",
2803
2804
2805
2806
2807
2808
2809
2810
2811
2812
2813
2814
2815
2816
2817
2818
2819
2820
                self._DEFAULT_SAMPLING_PARAMS["repetition_penalty"],
            )

        return SamplingParams.from_optional(
            temperature=temperature,
            max_tokens=max_tokens,
            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,
            output_kind=RequestOutputKind.DELTA
            if self.stream
            else RequestOutputKind.FINAL_ONLY,
            extra_args=self.vllm_xargs,
        )
2821
2822
2823

    @model_validator(mode="before")
    @classmethod
2824
2825
2826
2827
2828
2829
2830
    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'.",
            )

2831
2832
2833
        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:
2834
            raise ValueError("Stream options can only be defined when `stream=True`.")
2835
2836

        return data
2837
2838
2839


# Transcription response objects
2840
2841
2842
2843
2844
class TranscriptionUsageAudio(OpenAIBaseModel):
    type: Literal["duration"] = "duration"
    seconds: int


2845
2846
2847
class TranscriptionResponse(OpenAIBaseModel):
    text: str
    """The transcribed text."""
2848
    usage: TranscriptionUsageAudio
2849
2850
2851
2852
2853
2854
2855
2856
2857
2858
2859
2860
2861
2862
2863
2864
2865
2866
2867
2868
2869
2870
2871
2872
2873
2874
2875
2876
2877
2878
2879
2880
2881
2882
2883
2884
2885
2886
2887
2888
2889
2890
2891
2892
2893
2894
2895
2896
2897
2898
2899


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."""

2900
    tokens: list[int]
2901
2902
2903
2904
2905
2906
2907
2908
2909
2910
2911
2912
2913
    """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."""

2914
    segments: list[TranscriptionSegment] | None = None
2915
2916
    """Segments of the transcribed text and their corresponding details."""

2917
    words: list[TranscriptionWord] | None = None
2918
    """Extracted words and their corresponding timestamps."""
2919
2920
2921
2922


class TranslationResponseStreamChoice(OpenAIBaseModel):
    delta: DeltaMessage
2923
2924
    finish_reason: str | None = None
    stop_reason: int | str | None = None
2925
2926
2927
2928
2929
2930
2931
2932


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]
2933
    usage: UsageInfo | None = Field(default=None)
2934
2935
2936
2937
2938
2939
2940
2941
2942
2943
2944
2945


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.
    """

2946
    model: str | None = None
2947
2948
2949
2950
2951
2952
2953
2954
2955
2956
2957
2958
2959
2960
2961
2962
2963
2964
2965
    """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]
2966
    seed: int | None = Field(None, ge=_LONG_INFO.min, le=_LONG_INFO.max)
2967
2968
    """The seed to use for sampling."""

2969
2970
2971
2972
2973
2974
2975
2976
2977
2978
2979
    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]
2980
    language: str | None = None
2981
2982
2983
2984
2985
2986
2987
    """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.
    """

2988
    to_language: str | None = None
2989
2990
2991
2992
2993
2994
2995
    """The language of the input audio we translate to.

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

2996
    stream: bool | None = False
2997
    """Custom field not present in the original OpenAI definition. When set,
2998
    it will enable output to be streamed in a similar fashion as the Chat
2999
    Completion endpoint.
3000
3001
    """
    # Flattened stream option to simplify form data.
3002
3003
    stream_include_usage: bool | None = False
    stream_continuous_usage_stats: bool | None = False
3004
3005
3006
3007
3008
3009
3010
3011
    # --8<-- [end:translation-extra-params]

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

    def to_sampling_params(
3012
        self, default_max_tokens: int, default_sampling_params: dict | None = None
3013
    ) -> SamplingParams:
3014
3015
3016
3017
3018
3019
3020
        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(
3021
3022
                "temperature", self._DEFAULT_SAMPLING_PARAMS["temperature"]
            )
3023

3024
3025
3026
3027
3028
3029
3030
3031
        return SamplingParams.from_optional(
            temperature=temperature,
            max_tokens=max_tokens,
            seed=self.seed,
            output_kind=RequestOutputKind.DELTA
            if self.stream
            else RequestOutputKind.FINAL_ONLY,
        )
3032
3033
3034
3035
3036
3037
3038

    @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:
3039
            raise ValueError("Stream options can only be defined when `stream=True`.")
3040
3041
3042
3043
3044
3045
3046
3047
3048
3049
3050
3051
3052
3053
3054
3055
3056
3057
3058
3059
3060
3061
3062
3063
3064
3065
3066
3067
3068
3069
3070
3071
3072
3073
3074
3075
3076
3077
3078
3079
3080
3081
3082
3083
3084
3085
3086
3087
3088
3089
3090
3091
3092
3093
3094
3095
3096
3097
3098
3099
3100
3101
3102
3103
3104
3105
3106
3107
3108
3109
3110
3111
3112

        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."""

3113
    segments: list[TranslationSegment] | None = None
3114
3115
    """Segments of the translated text and their corresponding details."""

3116
    words: list[TranslationWord] | None = None
3117
    """Extracted words and their corresponding timestamps."""