protocol.py 18.6 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
85
86
87
    root: Optional[str] = None
    parent: Optional[str] = None
    permission: List[ModelPermission] = Field(default_factory=list)


88
class ModelList(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
89
90
91
92
    object: str = "list"
    data: List[ModelCard] = Field(default_factory=list)


93
class UsageInfo(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
94
95
96
97
98
    prompt_tokens: int = 0
    total_tokens: int = 0
    completion_tokens: Optional[int] = 0


99
class ResponseFormat(OpenAIBaseModel):
100
    # type must be "json_object" or "text"
101
    type: Literal["text", "json_object"]
102
103


104
class ChatCompletionRequest(OpenAIBaseModel):
105
106
    # Ordered by official OpenAI API documentation
    # https://platform.openai.com/docs/api-reference/chat/create
107
    messages: List[ChatCompletionMessageParam]
108
109
110
111
112
    model: str
    frequency_penalty: Optional[float] = 0.0
    logit_bias: Optional[Dict[str, float]] = None
    logprobs: Optional[bool] = False
    top_logprobs: Optional[int] = None
113
    max_tokens: Optional[int] = None
114
115
116
    n: Optional[int] = 1
    presence_penalty: Optional[float] = 0.0
    response_format: Optional[ResponseFormat] = None
117
118
119
    seed: Optional[int] = Field(None,
                                ge=torch.iinfo(torch.long).min,
                                le=torch.iinfo(torch.long).max)
120
    stop: Optional[Union[str, List[str]]] = Field(default_factory=list)
Zhuohan Li's avatar
Zhuohan Li committed
121
    stream: Optional[bool] = False
122
123
    temperature: Optional[float] = 0.7
    top_p: Optional[float] = 1.0
Zhuohan Li's avatar
Zhuohan Li committed
124
    user: Optional[str] = None
125
126

    # doc: begin-chat-completion-sampling-params
127
128
    best_of: Optional[int] = None
    use_beam_search: Optional[bool] = False
129
130
131
132
    top_k: Optional[int] = -1
    min_p: Optional[float] = 0.0
    repetition_penalty: Optional[float] = 1.0
    length_penalty: Optional[float] = 1.0
133
    early_stopping: Optional[bool] = False
134
    ignore_eos: Optional[bool] = False
135
    min_tokens: Optional[int] = 0
136
    stop_token_ids: Optional[List[int]] = Field(default_factory=list)
137
    skip_special_tokens: Optional[bool] = True
138
    spaces_between_special_tokens: Optional[bool] = True
139
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
    # 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."),
    )
180
181
182
183
184
185
    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'"))
186
187
188
189
190
    guided_whitespace_pattern: Optional[str] = Field(
        default=None,
        description=(
            "If specified, will override the default whitespace pattern "
            "for guided json decoding."))
191
192

    # doc: end-chat-completion-extra-params
Zhuohan Li's avatar
Zhuohan Li committed
193

194
    def to_sampling_params(self) -> SamplingParams:
195
196
        if self.logprobs and not self.top_logprobs:
            raise ValueError("Top logprobs must be set when logprobs is.")
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

Zhuohan Li's avatar
Zhuohan Li committed
254

255
class CompletionRequest(OpenAIBaseModel):
256
257
    # Ordered by official OpenAI API documentation
    # https://platform.openai.com/docs/api-reference/completions/create
Zhuohan Li's avatar
Zhuohan Li committed
258
    model: str
259
    prompt: Union[List[int], List[List[int]], str, List[str]]
260
    best_of: Optional[int] = None
Zhuohan Li's avatar
Zhuohan Li committed
261
262
263
    echo: Optional[bool] = False
    frequency_penalty: Optional[float] = 0.0
    logit_bias: Optional[Dict[str, float]] = None
264
265
    logprobs: Optional[int] = None
    max_tokens: Optional[int] = 16
266
    n: int = 1
267
    presence_penalty: Optional[float] = 0.0
268
269
270
    seed: Optional[int] = Field(None,
                                ge=torch.iinfo(torch.long).min,
                                le=torch.iinfo(torch.long).max)
271
272
273
274
275
    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
276
    user: Optional[str] = None
277
278

    # doc: begin-completion-sampling-params
Zhuohan Li's avatar
Zhuohan Li committed
279
    use_beam_search: Optional[bool] = False
280
281
282
283
    top_k: Optional[int] = -1
    min_p: Optional[float] = 0.0
    repetition_penalty: Optional[float] = 1.0
    length_penalty: Optional[float] = 1.0
284
    early_stopping: Optional[bool] = False
285
    stop_token_ids: Optional[List[int]] = Field(default_factory=list)
286
    ignore_eos: Optional[bool] = False
287
    min_tokens: Optional[int] = 0
288
    skip_special_tokens: Optional[bool] = True
289
    spaces_between_special_tokens: Optional[bool] = True
290
    truncate_prompt_tokens: Optional[Annotated[int, Field(ge=1)]] = None
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
    # 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."),
    )
326
327
328
329
330
331
    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'"))
332
333
334
335
336
    guided_whitespace_pattern: Optional[str] = Field(
        default=None,
        description=(
            "If specified, will override the default whitespace pattern "
            "for guided json decoding."))
337
338

    # doc: end-completion-extra-params
Zhuohan Li's avatar
Zhuohan Li committed
339

340
341
342
    def to_sampling_params(self):
        echo_without_generation = self.echo and self.max_tokens == 0

343
344
345
346
347
348
        logits_processors = None
        if self.logit_bias:

            def logit_bias_logits_processor(
                    token_ids: List[int],
                    logits: torch.Tensor) -> torch.Tensor:
349
                assert self.logit_bias is not None
350
351
352
353
354
355
356
357
                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]

358
359
360
361
362
363
364
365
366
367
        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
368
            seed=self.seed,
369
370
371
372
            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,
373
            min_tokens=self.min_tokens,
374
375
            logprobs=self.logprobs,
            use_beam_search=self.use_beam_search,
376
            early_stopping=self.early_stopping,
377
378
379
            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),
380
381
            include_stop_str_in_output=self.include_stop_str_in_output,
            length_penalty=self.length_penalty,
382
            logits_processors=logits_processors,
383
            truncate_prompt_tokens=self.truncate_prompt_tokens,
384
385
        )

386
387
388
389
390
391
392
393
394
395
396
397
398
399
    @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

Zhuohan Li's avatar
Zhuohan Li committed
400

401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
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)


419
class LogProbs(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
420
421
422
    text_offset: List[int] = Field(default_factory=list)
    token_logprobs: List[Optional[float]] = Field(default_factory=list)
    tokens: List[str] = Field(default_factory=list)
423
    top_logprobs: Optional[List[Optional[Dict[str, float]]]] = None
Zhuohan Li's avatar
Zhuohan Li committed
424
425


426
class CompletionResponseChoice(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
427
428
429
    index: int
    text: str
    logprobs: Optional[LogProbs] = None
430
431
    finish_reason: Optional[str] = None
    stop_reason: Optional[Union[int, str]] = Field(
432
433
434
435
436
437
        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
438
439


440
class CompletionResponse(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
441
442
443
444
445
446
447
448
    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


449
class CompletionResponseStreamChoice(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
450
451
452
    index: int
    text: str
    logprobs: Optional[LogProbs] = None
453
454
    finish_reason: Optional[str] = None
    stop_reason: Optional[Union[int, str]] = Field(
455
456
457
458
459
460
        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
461
462


463
class CompletionStreamResponse(OpenAIBaseModel):
Zhuohan Li's avatar
Zhuohan Li committed
464
465
466
467
468
    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]
469
    usage: Optional[UsageInfo] = Field(default=None)
470
471


472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
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


487
class ChatMessage(OpenAIBaseModel):
488
489
490
491
    role: str
    content: str


492
class ChatCompletionResponseChoice(OpenAIBaseModel):
493
494
    index: int
    message: ChatMessage
495
    logprobs: Optional[LogProbs] = None
496
497
    finish_reason: Optional[str] = None
    stop_reason: Optional[Union[int, str]] = None
498
499


500
class ChatCompletionResponse(OpenAIBaseModel):
501
502
503
504
505
506
507
508
    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


509
class DeltaMessage(OpenAIBaseModel):
510
511
512
513
    role: Optional[str] = None
    content: Optional[str] = None


514
class ChatCompletionResponseStreamChoice(OpenAIBaseModel):
515
516
    index: int
    delta: DeltaMessage
517
    logprobs: Optional[LogProbs] = None
518
519
    finish_reason: Optional[str] = None
    stop_reason: Optional[Union[int, str]] = None
520
521


522
class ChatCompletionStreamResponse(OpenAIBaseModel):
523
524
525
526
527
    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]
528
    usage: Optional[UsageInfo] = Field(default=None)