protocol.py 21.1 KB
Newer Older
1
2
# Adapted from
# https://github.com/lm-sys/FastChat/blob/168ccc29d3f7edc50823016105c024fe2282732a/fastchat/protocol/openai_api_protocol.py
Zhuohan Li's avatar
Zhuohan Li committed
3
import time
4
from typing import Any, Dict, List, Literal, Optional, Union
Zhuohan Li's avatar
Zhuohan Li committed
5

6
import openai.types.chat
7
import torch
8
from pydantic import BaseModel, ConfigDict, Field, model_validator
9
10
# pydantic needs the TypedDict from typing_extensions
from typing_extensions import Annotated, Required, TypedDict
Zhuohan Li's avatar
Zhuohan Li committed
11

12
from vllm.pooling_params import PoolingParams
13
from vllm.sampling_params import SamplingParams
14
from vllm.utils import random_uuid
15

Zhuohan Li's avatar
Zhuohan Li committed
16

17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
class CustomChatCompletionContentPartParam(TypedDict, total=False):
    __pydantic_config__ = ConfigDict(extra="allow")  # type: ignore

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


ChatCompletionContentPartParam = Union[
    openai.types.chat.ChatCompletionContentPartParam,
    CustomChatCompletionContentPartParam]


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

    content: Union[str, List[ChatCompletionContentPartParam]]
    """The contents of the message."""

    name: str
    """An optional name for the participant.

    Provides the model information to differentiate between participants of the
    same role.
    """


ChatCompletionMessageParam = Union[
    openai.types.chat.ChatCompletionMessageParam,
    CustomChatCompletionMessageParam]


50
51
52
53
54
55
class OpenAIBaseModel(BaseModel):
    # OpenAI API does not allow extra fields
    model_config = ConfigDict(extra="forbid")


class ErrorResponse(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
56
57
58
59
    object: str = "error"
    message: str
    type: str
    param: Optional[str] = None
60
    code: int
Zhuohan Li's avatar
Zhuohan Li committed
61
62


63
class ModelPermission(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
64
65
66
67
68
69
70
71
72
73
74
    id: str = Field(default_factory=lambda: f"modelperm-{random_uuid()}")
    object: str = "model_permission"
    created: int = Field(default_factory=lambda: int(time.time()))
    allow_create_engine: bool = False
    allow_sampling: bool = True
    allow_logprobs: bool = True
    allow_search_indices: bool = False
    allow_view: bool = True
    allow_fine_tuning: bool = False
    organization: str = "*"
    group: Optional[str] = None
75
    is_blocking: bool = False
Zhuohan Li's avatar
Zhuohan Li committed
76
77


78
class ModelCard(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
79
80
81
    id: str
    object: str = "model"
    created: int = Field(default_factory=lambda: int(time.time()))
Woosuk Kwon's avatar
Woosuk Kwon committed
82
    owned_by: str = "vllm"
Zhuohan Li's avatar
Zhuohan Li committed
83
84
    root: Optional[str] = None
    parent: Optional[str] = None
85
    max_model_len: Optional[int] = None
Zhuohan Li's avatar
Zhuohan Li committed
86
87
88
    permission: List[ModelPermission] = Field(default_factory=list)


89
class ModelList(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
90
91
92
93
    object: str = "list"
    data: List[ModelCard] = Field(default_factory=list)


94
class UsageInfo(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
95
96
97
98
99
    prompt_tokens: int = 0
    total_tokens: int = 0
    completion_tokens: Optional[int] = 0


100
class ResponseFormat(OpenAIBaseModel):
101
    # type must be "json_object" or "text"
102
    type: Literal["text", "json_object"]
103
104


105
class ChatCompletionRequest(OpenAIBaseModel):
106
107
    # Ordered by official OpenAI API documentation
    # https://platform.openai.com/docs/api-reference/chat/create
108
    messages: List[ChatCompletionMessageParam]
109
110
111
112
    model: str
    frequency_penalty: Optional[float] = 0.0
    logit_bias: Optional[Dict[str, float]] = None
    logprobs: Optional[bool] = False
113
    top_logprobs: Optional[int] = 0
114
    max_tokens: Optional[int] = None
115
116
117
    n: Optional[int] = 1
    presence_penalty: Optional[float] = 0.0
    response_format: Optional[ResponseFormat] = None
118
119
120
    seed: Optional[int] = Field(None,
                                ge=torch.iinfo(torch.long).min,
                                le=torch.iinfo(torch.long).max)
121
    stop: Optional[Union[str, List[str]]] = Field(default_factory=list)
Zhuohan Li's avatar
Zhuohan Li committed
122
    stream: Optional[bool] = False
123
124
    temperature: Optional[float] = 0.7
    top_p: Optional[float] = 1.0
Zhuohan Li's avatar
Zhuohan Li committed
125
    user: Optional[str] = None
126
127

    # doc: begin-chat-completion-sampling-params
128
129
    best_of: Optional[int] = None
    use_beam_search: Optional[bool] = False
130
131
132
133
    top_k: Optional[int] = -1
    min_p: Optional[float] = 0.0
    repetition_penalty: Optional[float] = 1.0
    length_penalty: Optional[float] = 1.0
134
    early_stopping: Optional[bool] = False
135
    ignore_eos: Optional[bool] = False
136
    min_tokens: Optional[int] = 0
137
    stop_token_ids: Optional[List[int]] = Field(default_factory=list)
138
    skip_special_tokens: Optional[bool] = True
139
    spaces_between_special_tokens: Optional[bool] = True
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
    # doc: end-chat-completion-sampling-params

    # doc: begin-chat-completion-extra-params
    echo: Optional[bool] = Field(
        default=False,
        description=(
            "If true, the new message will be prepended with the last message "
            "if they belong to the same role."),
    )
    add_generation_prompt: Optional[bool] = Field(
        default=True,
        description=
        ("If true, the generation prompt will be added to the chat template. "
         "This is a parameter used by chat template in tokenizer config of the "
         "model."),
    )
    include_stop_str_in_output: Optional[bool] = Field(
        default=False,
        description=(
            "Whether to include the stop string in the output. "
            "This is only applied when the stop or stop_token_ids is set."),
    )
    guided_json: Optional[Union[str, dict, BaseModel]] = Field(
        default=None,
        description=("If specified, the output will follow the JSON schema."),
    )
    guided_regex: Optional[str] = Field(
        default=None,
        description=(
            "If specified, the output will follow the regex pattern."),
    )
    guided_choice: Optional[List[str]] = Field(
        default=None,
        description=(
            "If specified, the output will be exactly one of the choices."),
    )
    guided_grammar: Optional[str] = Field(
        default=None,
        description=(
            "If specified, the output will follow the context free grammar."),
    )
181
182
183
184
185
186
    guided_decoding_backend: Optional[str] = Field(
        default=None,
        description=(
            "If specified, will override the default guided decoding backend "
            "of the server for this specific request. If set, must be either "
            "'outlines' / 'lm-format-enforcer'"))
187
188
189
190
191
    guided_whitespace_pattern: Optional[str] = Field(
        default=None,
        description=(
            "If specified, will override the default whitespace pattern "
            "for guided json decoding."))
192
193

    # doc: end-chat-completion-extra-params
Zhuohan Li's avatar
Zhuohan Li committed
194

195
    def to_sampling_params(self) -> SamplingParams:
196
        # We now allow logprobs being true without top_logrobs.
197
198
199
200
201
202
203

        logits_processors = None
        if self.logit_bias:

            def logit_bias_logits_processor(
                    token_ids: List[int],
                    logits: torch.Tensor) -> torch.Tensor:
204
                assert self.logit_bias is not None
205
206
207
208
209
210
211
212
                for token_id, bias in self.logit_bias.items():
                    # Clamp the bias between -100 and 100 per OpenAI API spec
                    bias = min(100, max(-100, bias))
                    logits[int(token_id)] += bias
                return logits

            logits_processors = [logit_bias_logits_processor]

213
214
215
216
217
218
219
220
        return SamplingParams(
            n=self.n,
            presence_penalty=self.presence_penalty,
            frequency_penalty=self.frequency_penalty,
            repetition_penalty=self.repetition_penalty,
            temperature=self.temperature,
            top_p=self.top_p,
            min_p=self.min_p,
Nick Hill's avatar
Nick Hill committed
221
            seed=self.seed,
222
223
224
            stop=self.stop,
            stop_token_ids=self.stop_token_ids,
            max_tokens=self.max_tokens,
225
            min_tokens=self.min_tokens,
226
227
            logprobs=self.top_logprobs if self.logprobs else None,
            prompt_logprobs=self.top_logprobs if self.echo else None,
228
229
230
231
            best_of=self.best_of,
            top_k=self.top_k,
            ignore_eos=self.ignore_eos,
            use_beam_search=self.use_beam_search,
232
            early_stopping=self.early_stopping,
233
234
            skip_special_tokens=self.skip_special_tokens,
            spaces_between_special_tokens=self.spaces_between_special_tokens,
235
236
            include_stop_str_in_output=self.include_stop_str_in_output,
            length_penalty=self.length_penalty,
237
            logits_processors=logits_processors,
238
239
        )

240
241
242
243
244
245
246
247
248
249
250
251
252
253
    @model_validator(mode="before")
    @classmethod
    def check_guided_decoding_count(cls, data):
        guide_count = sum([
            "guided_json" in data and data["guided_json"] is not None,
            "guided_regex" in data and data["guided_regex"] is not None,
            "guided_choice" in data and data["guided_choice"] is not None
        ])
        if guide_count > 1:
            raise ValueError(
                "You can only use one kind of guided decoding "
                "('guided_json', 'guided_regex' or 'guided_choice').")
        return data

254
255
256
257
258
259
260
261
262
263
264
265
266
    @model_validator(mode="before")
    @classmethod
    def check_logprobs(cls, data):
        if "top_logprobs" in data and data["top_logprobs"] is not None:
            if "logprobs" not in data or data["logprobs"] is False:
                raise ValueError(
                    "when using `top_logprobs`, `logprobs` must be set to true."
                )
            elif not 0 <= data["top_logprobs"] <= 20:
                raise ValueError(
                    "`top_logprobs` must be a value in the interval [0, 20].")
        return data

Zhuohan Li's avatar
Zhuohan Li committed
267

268
class CompletionRequest(OpenAIBaseModel):
269
270
    # Ordered by official OpenAI API documentation
    # https://platform.openai.com/docs/api-reference/completions/create
Zhuohan Li's avatar
Zhuohan Li committed
271
    model: str
272
    prompt: Union[List[int], List[List[int]], str, List[str]]
273
    best_of: Optional[int] = None
Zhuohan Li's avatar
Zhuohan Li committed
274
275
276
    echo: Optional[bool] = False
    frequency_penalty: Optional[float] = 0.0
    logit_bias: Optional[Dict[str, float]] = None
277
278
    logprobs: Optional[int] = None
    max_tokens: Optional[int] = 16
279
    n: int = 1
280
    presence_penalty: Optional[float] = 0.0
281
282
283
    seed: Optional[int] = Field(None,
                                ge=torch.iinfo(torch.long).min,
                                le=torch.iinfo(torch.long).max)
284
285
286
287
288
    stop: Optional[Union[str, List[str]]] = Field(default_factory=list)
    stream: Optional[bool] = False
    suffix: Optional[str] = None
    temperature: Optional[float] = 1.0
    top_p: Optional[float] = 1.0
Zhuohan Li's avatar
Zhuohan Li committed
289
    user: Optional[str] = None
290
291

    # doc: begin-completion-sampling-params
Zhuohan Li's avatar
Zhuohan Li committed
292
    use_beam_search: Optional[bool] = False
293
294
295
296
    top_k: Optional[int] = -1
    min_p: Optional[float] = 0.0
    repetition_penalty: Optional[float] = 1.0
    length_penalty: Optional[float] = 1.0
297
    early_stopping: Optional[bool] = False
298
    stop_token_ids: Optional[List[int]] = Field(default_factory=list)
299
    ignore_eos: Optional[bool] = False
300
    min_tokens: Optional[int] = 0
301
    skip_special_tokens: Optional[bool] = True
302
    spaces_between_special_tokens: Optional[bool] = True
303
    truncate_prompt_tokens: Optional[Annotated[int, Field(ge=1)]] = None
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
    # doc: end-completion-sampling-params

    # doc: begin-completion-extra-params
    include_stop_str_in_output: Optional[bool] = Field(
        default=False,
        description=(
            "Whether to include the stop string in the output. "
            "This is only applied when the stop or stop_token_ids is set."),
    )
    response_format: Optional[ResponseFormat] = Field(
        default=None,
        description=
        ("Similar to chat completion, this parameter specifies the format of "
         "output. Only {'type': 'json_object'} or {'type': 'text' } is "
         "supported."),
    )
    guided_json: Optional[Union[str, dict, BaseModel]] = Field(
        default=None,
        description=("If specified, the output will follow the JSON schema."),
    )
    guided_regex: Optional[str] = Field(
        default=None,
        description=(
            "If specified, the output will follow the regex pattern."),
    )
    guided_choice: Optional[List[str]] = Field(
        default=None,
        description=(
            "If specified, the output will be exactly one of the choices."),
    )
    guided_grammar: Optional[str] = Field(
        default=None,
        description=(
            "If specified, the output will follow the context free grammar."),
    )
339
340
341
342
343
344
    guided_decoding_backend: Optional[str] = Field(
        default=None,
        description=(
            "If specified, will override the default guided decoding backend "
            "of the server for this specific request. If set, must be one of "
            "'outlines' / 'lm-format-enforcer'"))
345
346
347
348
349
    guided_whitespace_pattern: Optional[str] = Field(
        default=None,
        description=(
            "If specified, will override the default whitespace pattern "
            "for guided json decoding."))
350
351

    # doc: end-completion-extra-params
Zhuohan Li's avatar
Zhuohan Li committed
352

353
354
355
    def to_sampling_params(self):
        echo_without_generation = self.echo and self.max_tokens == 0

356
357
358
359
360
361
        logits_processors = None
        if self.logit_bias:

            def logit_bias_logits_processor(
                    token_ids: List[int],
                    logits: torch.Tensor) -> torch.Tensor:
362
                assert self.logit_bias is not None
363
364
365
366
367
368
369
370
                for token_id, bias in self.logit_bias.items():
                    # Clamp the bias between -100 and 100 per OpenAI API spec
                    bias = min(100, max(-100, bias))
                    logits[int(token_id)] += bias
                return logits

            logits_processors = [logit_bias_logits_processor]

371
372
373
374
375
376
377
378
379
380
        return SamplingParams(
            n=self.n,
            best_of=self.best_of,
            presence_penalty=self.presence_penalty,
            frequency_penalty=self.frequency_penalty,
            repetition_penalty=self.repetition_penalty,
            temperature=self.temperature,
            top_p=self.top_p,
            top_k=self.top_k,
            min_p=self.min_p,
Nick Hill's avatar
Nick Hill committed
381
            seed=self.seed,
382
383
384
385
            stop=self.stop,
            stop_token_ids=self.stop_token_ids,
            ignore_eos=self.ignore_eos,
            max_tokens=self.max_tokens if not echo_without_generation else 1,
386
            min_tokens=self.min_tokens,
387
388
            logprobs=self.logprobs,
            use_beam_search=self.use_beam_search,
389
            early_stopping=self.early_stopping,
390
391
392
            prompt_logprobs=self.logprobs if self.echo else None,
            skip_special_tokens=self.skip_special_tokens,
            spaces_between_special_tokens=(self.spaces_between_special_tokens),
393
394
            include_stop_str_in_output=self.include_stop_str_in_output,
            length_penalty=self.length_penalty,
395
            logits_processors=logits_processors,
396
            truncate_prompt_tokens=self.truncate_prompt_tokens,
397
398
        )

399
400
401
402
403
404
405
406
407
408
409
410
411
412
    @model_validator(mode="before")
    @classmethod
    def check_guided_decoding_count(cls, data):
        guide_count = sum([
            "guided_json" in data and data["guided_json"] is not None,
            "guided_regex" in data and data["guided_regex"] is not None,
            "guided_choice" in data and data["guided_choice"] is not None
        ])
        if guide_count > 1:
            raise ValueError(
                "You can only use one kind of guided decoding "
                "('guided_json', 'guided_regex' or 'guided_choice').")
        return data

413
414
415
416
417
418
419
420
421
    @model_validator(mode="before")
    @classmethod
    def check_logprobs(cls, data):
        if "logprobs" in data and data[
                "logprobs"] is not None and not 0 <= data["logprobs"] <= 5:
            raise ValueError(("if passed, `logprobs` must be a value",
                              " in the interval [0, 5]."))
        return data

Zhuohan Li's avatar
Zhuohan Li committed
422

423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
class EmbeddingRequest(BaseModel):
    # Ordered by official OpenAI API documentation
    # https://platform.openai.com/docs/api-reference/embeddings
    model: str
    input: Union[List[int], List[List[int]], str, List[str]]
    encoding_format: Optional[str] = Field('float', pattern='^(float|base64)$')
    dimensions: Optional[int] = None
    user: Optional[str] = None

    # doc: begin-embedding-pooling-params
    additional_data: Optional[Any] = None

    # doc: end-embedding-pooling-params

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


441
class CompletionLogProbs(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
442
443
444
    text_offset: List[int] = Field(default_factory=list)
    token_logprobs: List[Optional[float]] = Field(default_factory=list)
    tokens: List[str] = Field(default_factory=list)
445
    top_logprobs: Optional[List[Optional[Dict[str, float]]]] = None
Zhuohan Li's avatar
Zhuohan Li committed
446
447


448
class CompletionResponseChoice(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
449
450
    index: int
    text: str
451
    logprobs: Optional[CompletionLogProbs] = None
452
453
    finish_reason: Optional[str] = None
    stop_reason: Optional[Union[int, str]] = Field(
454
455
456
457
458
459
        default=None,
        description=(
            "The stop string or token id that caused the completion "
            "to stop, None if the completion finished for some other reason "
            "including encountering the EOS token"),
    )
Zhuohan Li's avatar
Zhuohan Li committed
460
461


462
class CompletionResponse(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
463
464
465
466
467
468
469
470
    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
    choices: List[CompletionResponseChoice]
    usage: UsageInfo


471
class CompletionResponseStreamChoice(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
472
473
    index: int
    text: str
474
    logprobs: Optional[CompletionLogProbs] = None
475
476
    finish_reason: Optional[str] = None
    stop_reason: Optional[Union[int, str]] = Field(
477
478
479
480
481
482
        default=None,
        description=(
            "The stop string or token id that caused the completion "
            "to stop, None if the completion finished for some other reason "
            "including encountering the EOS token"),
    )
Zhuohan Li's avatar
Zhuohan Li committed
483
484


485
class CompletionStreamResponse(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
486
487
488
489
490
    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
    choices: List[CompletionResponseStreamChoice]
491
    usage: Optional[UsageInfo] = Field(default=None)
492
493


494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
class EmbeddingResponseData(BaseModel):
    index: int
    object: str = "embedding"
    embedding: List[float]


class EmbeddingResponse(BaseModel):
    id: str = Field(default_factory=lambda: f"cmpl-{random_uuid()}")
    object: str = "list"
    created: int = Field(default_factory=lambda: int(time.time()))
    model: str
    data: List[EmbeddingResponseData]
    usage: UsageInfo


509
class ChatMessage(OpenAIBaseModel):
510
511
512
513
    role: str
    content: str


514
515
516
517
518
519
520
521
522
523
524
525
526
527
class ChatCompletionLogProb(OpenAIBaseModel):
    token: str
    logprob: float = -9999.0
    bytes: Optional[List[int]] = None


class ChatCompletionLogProbsContent(ChatCompletionLogProb):
    top_logprobs: List[ChatCompletionLogProb] = Field(default_factory=list)


class ChatCompletionLogProbs(OpenAIBaseModel):
    content: Optional[List[ChatCompletionLogProbsContent]] = None


528
class ChatCompletionResponseChoice(OpenAIBaseModel):
529
530
    index: int
    message: ChatMessage
531
532
    logprobs: Optional[ChatCompletionLogProbs] = None
    finish_reason: Optional[Literal["stop", "length", "tool_calls"]] = None
533
    stop_reason: Optional[Union[int, str]] = None
534
535


536
class ChatCompletionResponse(OpenAIBaseModel):
537
538
539
540
541
542
543
544
    id: str = Field(default_factory=lambda: f"chatcmpl-{random_uuid()}")
    object: str = "chat.completion"
    created: int = Field(default_factory=lambda: int(time.time()))
    model: str
    choices: List[ChatCompletionResponseChoice]
    usage: UsageInfo


545
class DeltaMessage(OpenAIBaseModel):
546
547
548
549
    role: Optional[str] = None
    content: Optional[str] = None


550
class ChatCompletionResponseStreamChoice(OpenAIBaseModel):
551
552
    index: int
    delta: DeltaMessage
553
554
    logprobs: Optional[ChatCompletionLogProbs] = None
    finish_reason: Optional[Literal["stop", "length", "tool_calls"]] = None
555
    stop_reason: Optional[Union[int, str]] = None
556
557


558
class ChatCompletionStreamResponse(OpenAIBaseModel):
559
560
561
562
563
    id: str = Field(default_factory=lambda: f"chatcmpl-{random_uuid()}")
    object: str = "chat.completion.chunk"
    created: int = Field(default_factory=lambda: int(time.time()))
    model: str
    choices: List[ChatCompletionResponseStreamChoice]
564
    usage: Optional[UsageInfo] = Field(default=None)
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605


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

    # The parameteters of the request.
    body: Union[ChatCompletionRequest, ]


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

    response: Optional[ChatCompletionResponse]

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