api_server.py 21.7 KB
Newer Older
1
2
# Adapted from
# https://github.com/lm-sys/FastChat/blob/168ccc29d3f7edc50823016105c024fe2282732a/fastchat/serve/openai_api_server.py
Zhuohan Li's avatar
Zhuohan Li committed
3
4

import argparse
5
import asyncio
Zhuohan Li's avatar
Zhuohan Li committed
6
7
8
from http import HTTPStatus
import json
import time
9
from typing import AsyncGenerator, Dict, List, Optional
Zhuohan Li's avatar
Zhuohan Li committed
10
11

import fastapi
12
from fastapi import BackgroundTasks, Request
Zhuohan Li's avatar
Zhuohan Li committed
13
14
from fastapi.exceptions import RequestValidationError
from fastapi.middleware.cors import CORSMiddleware
Zhuohan Li's avatar
Zhuohan Li committed
15
from fastapi.responses import JSONResponse, StreamingResponse
16
17
from fastchat.conversation import (Conversation, SeparatorStyle,
                                   get_conv_template)
Zhuohan Li's avatar
Zhuohan Li committed
18
19
import uvicorn

Woosuk Kwon's avatar
Woosuk Kwon committed
20
21
22
from vllm.engine.arg_utils import AsyncEngineArgs
from vllm.engine.async_llm_engine import AsyncLLMEngine
from vllm.entrypoints.openai.protocol import (
Zhuohan Li's avatar
Zhuohan Li committed
23
    CompletionRequest, CompletionResponse, CompletionResponseChoice,
24
    CompletionResponseStreamChoice, CompletionStreamResponse,
25
26
27
28
    ChatCompletionRequest, ChatCompletionResponse,
    ChatCompletionResponseChoice, ChatCompletionResponseStreamChoice,
    ChatCompletionStreamResponse, ChatMessage, DeltaMessage, ErrorResponse,
    LogProbs, ModelCard, ModelList, ModelPermission, UsageInfo)
Woosuk Kwon's avatar
Woosuk Kwon committed
29
30
31
from vllm.logger import init_logger
from vllm.outputs import RequestOutput
from vllm.sampling_params import SamplingParams
32
from vllm.transformers_utils.tokenizer import get_tokenizer
Woosuk Kwon's avatar
Woosuk Kwon committed
33
from vllm.utils import random_uuid
Zhuohan Li's avatar
Zhuohan Li committed
34

35
TIMEOUT_KEEP_ALIVE = 5  # seconds
Zhuohan Li's avatar
Zhuohan Li committed
36
37
38
39
40
41
42
43

logger = init_logger(__name__)
served_model = None
app = fastapi.FastAPI()


def create_error_response(status_code: HTTPStatus,
                          message: str) -> JSONResponse:
44
45
46
    return JSONResponse(ErrorResponse(message=message,
                                      type="invalid_request_error").dict(),
                        status_code=status_code.value)
Zhuohan Li's avatar
Zhuohan Li committed
47
48
49


@app.exception_handler(RequestValidationError)
50
async def validation_exception_handler(request, exc):  # pylint: disable=unused-argument
Zhuohan Li's avatar
Zhuohan Li committed
51
52
53
54
55
56
57
58
59
60
61
62
63
    return create_error_response(HTTPStatus.BAD_REQUEST, str(exc))


async def check_model(request) -> Optional[JSONResponse]:
    if request.model == served_model:
        return
    ret = create_error_response(
        HTTPStatus.NOT_FOUND,
        f"The model `{request.model}` does not exist.",
    )
    return ret


64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
async def get_gen_prompt(request) -> str:
    conv = get_conv_template(request.model)
    conv = Conversation(
        name=conv.name,
        system=conv.system,
        roles=conv.roles,
        messages=list(conv.messages),  # prevent in-place modification
        offset=conv.offset,
        sep_style=SeparatorStyle(conv.sep_style),
        sep=conv.sep,
        sep2=conv.sep2,
        stop_str=conv.stop_str,
        stop_token_ids=conv.stop_token_ids,
    )

    if isinstance(request.messages, str):
        prompt = request.messages
    else:
        for message in request.messages:
            msg_role = message["role"]
            if msg_role == "system":
                conv.system = message["content"]
            elif msg_role == "user":
                conv.append_message(conv.roles[0], message["content"])
            elif msg_role == "assistant":
                conv.append_message(conv.roles[1], message["content"])
            else:
                raise ValueError(f"Unknown role: {msg_role}")

        # Add a blank message for the assistant.
        conv.append_message(conv.roles[1], None)
        prompt = conv.get_prompt()

    return prompt


100
101
102
103
104
105
106
107
108
async def check_length(request, prompt, model_config):
    if hasattr(model_config.hf_config, "max_sequence_length"):
        context_len = model_config.hf_config.max_sequence_length
    elif hasattr(model_config.hf_config, "seq_length"):
        context_len = model_config.hf_config.seq_length
    elif hasattr(model_config.hf_config, "max_position_embeddings"):
        context_len = model_config.hf_config.max_position_embeddings
    elif hasattr(model_config.hf_config, "seq_length"):
        context_len = model_config.hf_config.seq_length
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
    else:
        context_len = 2048

    input_ids = tokenizer(prompt).input_ids
    token_num = len(input_ids)

    if token_num + request.max_tokens > context_len:
        return create_error_response(
            HTTPStatus.BAD_REQUEST,
            f"This model's maximum context length is {context_len} tokens. "
            f"However, you requested {request.max_tokens + token_num} tokens "
            f"({token_num} in the messages, "
            f"{request.max_tokens} in the completion). "
            f"Please reduce the length of the messages or completion.",
        )
    else:
        return None


Zhuohan Li's avatar
Zhuohan Li committed
128
129
130
@app.get("/v1/models")
async def show_available_models():
    """Show available models. Right now we only have one model."""
131
132
133
134
135
    model_cards = [
        ModelCard(id=served_model,
                  root=served_model,
                  permission=[ModelPermission()])
    ]
Zhuohan Li's avatar
Zhuohan Li committed
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
    return ModelList(data=model_cards)


def create_logprobs(token_ids: List[int],
                    id_logprobs: List[Dict[int, float]],
                    initial_text_offset: int = 0) -> LogProbs:
    """Create OpenAI-style logprobs."""
    logprobs = LogProbs()
    last_token_len = 0
    for token_id, id_logprob in zip(token_ids, id_logprobs):
        token = tokenizer.convert_ids_to_tokens(token_id)
        logprobs.tokens.append(token)
        logprobs.token_logprobs.append(id_logprob[token_id])
        if len(logprobs.text_offset) == 0:
            logprobs.text_offset.append(initial_text_offset)
        else:
152
153
            logprobs.text_offset.append(logprobs.text_offset[-1] +
                                        last_token_len)
Zhuohan Li's avatar
Zhuohan Li committed
154
155
        last_token_len = len(token)

156
157
158
159
        logprobs.top_logprobs.append({
            tokenizer.convert_ids_to_tokens(i): p
            for i, p in id_logprob.items()
        })
Zhuohan Li's avatar
Zhuohan Li committed
160
161
162
    return logprobs


163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
@app.post("/v1/chat/completions")
async def create_chat_completion(raw_request: Request):
    """Completion API similar to OpenAI's API.

    See  https://platform.openai.com/docs/api-reference/chat/create
    for the API specification. This API mimics the OpenAI ChatCompletion API.

    NOTE: Currently we do not support the following features:
        - function_call (Users should implement this by themselves)
        - logit_bias (to be supported by vLLM engine)
    """
    request = ChatCompletionRequest(**await raw_request.json())
    logger.info(f"Received chat completion request: {request}")

    error_check_ret = await check_model(request)
    if error_check_ret is not None:
        return error_check_ret

    if request.logit_bias is not None:
        # TODO: support logit_bias in vLLM engine.
        return create_error_response(HTTPStatus.BAD_REQUEST,
                                     "logit_bias is not currently supported")

    prompt = await get_gen_prompt(request)
187
    error_check_ret = await check_length(request, prompt, engine_model_config)
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
    if error_check_ret is not None:
        return error_check_ret

    model_name = request.model
    request_id = f"cmpl-{random_uuid()}"
    created_time = int(time.time())
    try:
        sampling_params = SamplingParams(
            n=request.n,
            presence_penalty=request.presence_penalty,
            frequency_penalty=request.frequency_penalty,
            temperature=request.temperature,
            top_p=request.top_p,
            stop=request.stop,
            max_tokens=request.max_tokens,
            best_of=request.best_of,
            top_k=request.top_k,
            ignore_eos=request.ignore_eos,
            use_beam_search=request.use_beam_search,
        )
    except ValueError as e:
        return create_error_response(HTTPStatus.BAD_REQUEST, str(e))

211
    result_generator = engine.generate(prompt, sampling_params, request_id)
212
213
214
215

    async def abort_request() -> None:
        await engine.abort(request_id)

216
217
218
219
220
    def create_stream_response_json(
        index: int,
        text: str,
        finish_reason: Optional[str] = None,
    ) -> str:
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
        choice_data = ChatCompletionResponseStreamChoice(
            index=index,
            delta=DeltaMessage(content=text),
            finish_reason=finish_reason,
        )
        response = ChatCompletionStreamResponse(
            id=request_id,
            created=created_time,
            model=model_name,
            choices=[choice_data],
        )
        response_json = response.json(ensure_ascii=False)

        return response_json

    async def completion_stream_generator() -> AsyncGenerator[str, None]:
        # First chunk with role
        for i in range(request.n):
            choice_data = ChatCompletionResponseStreamChoice(
                index=i,
                delta=DeltaMessage(role="assistant"),
                finish_reason=None,
            )
244
245
246
247
248
            chunk = ChatCompletionStreamResponse(id=request_id,
                                                 choices=[choice_data],
                                                 model=model_name)
            data = chunk.json(exclude_unset=True, ensure_ascii=False)
            yield f"data: {data}\n\n"
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301

        previous_texts = [""] * request.n
        previous_num_tokens = [0] * request.n
        async for res in result_generator:
            res: RequestOutput
            for output in res.outputs:
                i = output.index
                delta_text = output.text[len(previous_texts[i]):]
                previous_texts[i] = output.text
                previous_num_tokens[i] = len(output.token_ids)
                response_json = create_stream_response_json(
                    index=i,
                    text=delta_text,
                )
                yield f"data: {response_json}\n\n"
                if output.finish_reason is not None:
                    response_json = create_stream_response_json(
                        index=i,
                        text="",
                        finish_reason=output.finish_reason,
                    )
                    yield f"data: {response_json}\n\n"
            yield "data: [DONE]\n\n"

    # Streaming response
    if request.stream:
        background_tasks = BackgroundTasks()
        # Abort the request if the client disconnects.
        background_tasks.add_task(abort_request)
        return StreamingResponse(completion_stream_generator(),
                                 media_type="text/event-stream",
                                 background=background_tasks)

    # Non-streaming response
    final_res: RequestOutput = None
    async for res in result_generator:
        if await raw_request.is_disconnected():
            # Abort the request if the client disconnects.
            await abort_request()
            return create_error_response(HTTPStatus.BAD_REQUEST,
                                         "Client disconnected")
        final_res = res
    assert final_res is not None
    choices = []
    for output in final_res.outputs:
        choice_data = ChatCompletionResponseChoice(
            index=output.index,
            message=ChatMessage(role="assistant", content=output.text),
            finish_reason=output.finish_reason,
        )
        choices.append(choice_data)

    num_prompt_tokens = len(final_res.prompt_token_ids)
302
303
    num_generated_tokens = sum(
        len(output.token_ids) for output in final_res.outputs)
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
    usage = UsageInfo(
        prompt_tokens=num_prompt_tokens,
        completion_tokens=num_generated_tokens,
        total_tokens=num_prompt_tokens + num_generated_tokens,
    )
    response = ChatCompletionResponse(
        id=request_id,
        created=created_time,
        model=model_name,
        choices=choices,
        usage=usage,
    )

    if request.stream:
        # When user requests streaming but we don't stream, we still need to
        # return a streaming response with a single event.
        response_json = response.json(ensure_ascii=False)
321

322
323
324
        async def fake_stream_generator() -> AsyncGenerator[str, None]:
            yield f"data: {response_json}\n\n"
            yield "data: [DONE]\n\n"
325

326
327
328
329
330
331
        return StreamingResponse(fake_stream_generator(),
                                 media_type="text/event-stream")

    return response


Zhuohan Li's avatar
Zhuohan Li committed
332
@app.post("/v1/completions")
333
async def create_completion(raw_request: Request):
334
335
336
337
338
339
    """Completion API similar to OpenAI's API.

    See https://platform.openai.com/docs/api-reference/completions/create
    for the API specification. This API mimics the OpenAI Completion API.

    NOTE: Currently we do not support the following features:
Woosuk Kwon's avatar
Woosuk Kwon committed
340
        - echo (since the vLLM engine does not currently support
341
342
343
          getting the logprobs of prompt tokens)
        - suffix (the language models we currently support do not support
          suffix)
Woosuk Kwon's avatar
Woosuk Kwon committed
344
        - logit_bias (to be supported by vLLM engine)
345
    """
346
    request = CompletionRequest(**await raw_request.json())
Zhuohan Li's avatar
Zhuohan Li committed
347
348
349
350
351
352
353
    logger.info(f"Received completion request: {request}")

    error_check_ret = await check_model(request)
    if error_check_ret is not None:
        return error_check_ret

    if request.echo:
Woosuk Kwon's avatar
Woosuk Kwon committed
354
        # We do not support echo since the vLLM engine does not
Zhuohan Li's avatar
Zhuohan Li committed
355
356
357
358
359
360
361
        # currently support getting the logprobs of prompt tokens.
        return create_error_response(HTTPStatus.BAD_REQUEST,
                                     "echo is not currently supported")

    if request.suffix is not None:
        # The language models we currently support do not support suffix.
        return create_error_response(HTTPStatus.BAD_REQUEST,
362
                                     "suffix is not currently supported")
Zhuohan Li's avatar
Zhuohan Li committed
363
364

    if request.logit_bias is not None:
Woosuk Kwon's avatar
Woosuk Kwon committed
365
        # TODO: support logit_bias in vLLM engine.
Zhuohan Li's avatar
Zhuohan Li committed
366
367
368
369
370
        return create_error_response(HTTPStatus.BAD_REQUEST,
                                     "logit_bias is not currently supported")

    model_name = request.model
    request_id = f"cmpl-{random_uuid()}"
371
    if isinstance(request.prompt, list):
372
373
374
375
        if len(request.prompt) == 0:
            return create_error_response(HTTPStatus.BAD_REQUEST,
                                         "please provide at least one prompt")
        if len(request.prompt) > 1:
376
377
378
            return create_error_response(
                HTTPStatus.BAD_REQUEST,
                "multiple prompts in a batch is not currently supported")
379
380
381
        prompt = request.prompt[0]
    else:
        prompt = request.prompt
Zhuohan Li's avatar
Zhuohan Li committed
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
    created_time = int(time.time())
    try:
        sampling_params = SamplingParams(
            n=request.n,
            best_of=request.best_of,
            presence_penalty=request.presence_penalty,
            frequency_penalty=request.frequency_penalty,
            temperature=request.temperature,
            top_p=request.top_p,
            top_k=request.top_k,
            stop=request.stop,
            ignore_eos=request.ignore_eos,
            max_tokens=request.max_tokens,
            logprobs=request.logprobs,
            use_beam_search=request.use_beam_search,
        )
    except ValueError as e:
        return create_error_response(HTTPStatus.BAD_REQUEST, str(e))

401
    result_generator = engine.generate(prompt, sampling_params, request_id)
Zhuohan Li's avatar
Zhuohan Li committed
402
403
404

    # Similar to the OpenAI API, when n != best_of, we do not stream the
    # results. In addition, we do not stream the results when use beam search.
405
406
407
    stream = (request.stream
              and (request.best_of is None or request.n == request.best_of)
              and not request.use_beam_search)
Zhuohan Li's avatar
Zhuohan Li committed
408

409
    async def abort_request() -> None:
Zhuohan Li's avatar
Zhuohan Li committed
410
        await engine.abort(request_id)
411

412
413
414
415
416
417
    def create_stream_response_json(
        index: int,
        text: str,
        logprobs: Optional[LogProbs] = None,
        finish_reason: Optional[str] = None,
    ) -> str:
Zhuohan Li's avatar
Zhuohan Li committed
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
        choice_data = CompletionResponseStreamChoice(
            index=index,
            text=text,
            logprobs=logprobs,
            finish_reason=finish_reason,
        )
        response = CompletionStreamResponse(
            id=request_id,
            created=created_time,
            model=model_name,
            choices=[choice_data],
        )
        response_json = response.json(ensure_ascii=False)

        return response_json

    async def completion_stream_generator() -> AsyncGenerator[str, None]:
        previous_texts = [""] * request.n
        previous_num_tokens = [0] * request.n
        async for res in result_generator:
            res: RequestOutput
            for output in res.outputs:
                i = output.index
                delta_text = output.text[len(previous_texts[i]):]
                if request.logprobs is not None:
                    logprobs = create_logprobs(
                        output.token_ids[previous_num_tokens[i]:],
                        output.logprobs[previous_num_tokens[i]:],
                        len(previous_texts[i]))
                else:
                    logprobs = None
                previous_texts[i] = output.text
                previous_num_tokens[i] = len(output.token_ids)
                response_json = create_stream_response_json(
                    index=i,
                    text=delta_text,
                    logprobs=logprobs,
                )
                yield f"data: {response_json}\n\n"
                if output.finish_reason is not None:
458
459
                    logprobs = (LogProbs()
                                if request.logprobs is not None else None)
Zhuohan Li's avatar
Zhuohan Li committed
460
461
462
463
464
465
466
467
468
469
470
                    response_json = create_stream_response_json(
                        index=i,
                        text="",
                        logprobs=logprobs,
                        finish_reason=output.finish_reason,
                    )
                    yield f"data: {response_json}\n\n"
            yield "data: [DONE]\n\n"

    # Streaming response
    if stream:
471
472
473
        background_tasks = BackgroundTasks()
        # Abort the request if the client disconnects.
        background_tasks.add_task(abort_request)
Zhuohan Li's avatar
Zhuohan Li committed
474
        return StreamingResponse(completion_stream_generator(),
475
476
                                 media_type="text/event-stream",
                                 background=background_tasks)
Zhuohan Li's avatar
Zhuohan Li committed
477
478
479
480

    # Non-streaming response
    final_res: RequestOutput = None
    async for res in result_generator:
481
482
        if await raw_request.is_disconnected():
            # Abort the request if the client disconnects.
483
            await abort_request()
484
485
            return create_error_response(HTTPStatus.BAD_REQUEST,
                                         "Client disconnected")
Zhuohan Li's avatar
Zhuohan Li committed
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
        final_res = res
    assert final_res is not None
    choices = []
    for output in final_res.outputs:
        if request.logprobs is not None:
            logprobs = create_logprobs(output.token_ids, output.logprobs)
        else:
            logprobs = None
        choice_data = CompletionResponseChoice(
            index=output.index,
            text=output.text,
            logprobs=logprobs,
            finish_reason=output.finish_reason,
        )
        choices.append(choice_data)

    num_prompt_tokens = len(final_res.prompt_token_ids)
503
504
    num_generated_tokens = sum(
        len(output.token_ids) for output in final_res.outputs)
Zhuohan Li's avatar
Zhuohan Li committed
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
    usage = UsageInfo(
        prompt_tokens=num_prompt_tokens,
        completion_tokens=num_generated_tokens,
        total_tokens=num_prompt_tokens + num_generated_tokens,
    )
    response = CompletionResponse(
        id=request_id,
        created=created_time,
        model=model_name,
        choices=choices,
        usage=usage,
    )

    if request.stream:
        # When user requests streaming but we don't stream, we still need to
        # return a streaming response with a single event.
        response_json = response.json(ensure_ascii=False)
522

Zhuohan Li's avatar
Zhuohan Li committed
523
524
525
        async def fake_stream_generator() -> AsyncGenerator[str, None]:
            yield f"data: {response_json}\n\n"
            yield "data: [DONE]\n\n"
526

Zhuohan Li's avatar
Zhuohan Li committed
527
528
529
530
531
532
533
534
        return StreamingResponse(fake_stream_generator(),
                                 media_type="text/event-stream")

    return response


if __name__ == "__main__":
    parser = argparse.ArgumentParser(
535
536
537
538
539
        description="vLLM OpenAI-Compatible RESTful API server.")
    parser.add_argument("--host",
                        type=str,
                        default="localhost",
                        help="host name")
Zhuohan Li's avatar
Zhuohan Li committed
540
    parser.add_argument("--port", type=int, default=8000, help="port number")
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
    parser.add_argument("--allow-credentials",
                        action="store_true",
                        help="allow credentials")
    parser.add_argument("--allowed-origins",
                        type=json.loads,
                        default=["*"],
                        help="allowed origins")
    parser.add_argument("--allowed-methods",
                        type=json.loads,
                        default=["*"],
                        help="allowed methods")
    parser.add_argument("--allowed-headers",
                        type=json.loads,
                        default=["*"],
                        help="allowed headers")
Zhuohan Li's avatar
Zhuohan Li committed
556
    parser.add_argument(
557
558
559
560
561
562
        "--served-model-name",
        type=str,
        default=None,
        help="The model name used in the API. If not specified, "
        "the model name will be the same as the "
        "huggingface name.")
Zhuohan Li's avatar
Zhuohan Li committed
563
    parser = AsyncEngineArgs.add_cli_args(parser)
Zhuohan Li's avatar
Zhuohan Li committed
564
565
566
567
568
569
570
571
572
573
574
575
576
577
    args = parser.parse_args()

    app.add_middleware(
        CORSMiddleware,
        allow_origins=args.allowed_origins,
        allow_credentials=args.allow_credentials,
        allow_methods=args.allowed_methods,
        allow_headers=args.allowed_headers,
    )

    logger.info(f"args: {args}")

    served_model = args.served_model_name or args.model

Zhuohan Li's avatar
Zhuohan Li committed
578
579
    engine_args = AsyncEngineArgs.from_cli_args(args)
    engine = AsyncLLMEngine.from_engine_args(engine_args)
580
    engine_model_config = asyncio.run(engine.get_model_config())
Zhuohan Li's avatar
Zhuohan Li committed
581
582

    # A separate tokenizer to map token IDs to strings.
583
584
    tokenizer = get_tokenizer(engine_args.tokenizer,
                              tokenizer_mode=engine_args.tokenizer_mode)
Zhuohan Li's avatar
Zhuohan Li committed
585

586
587
588
589
    uvicorn.run(app,
                host=args.host,
                port=args.port,
                log_level="info",
590
                timeout_keep_alive=TIMEOUT_KEEP_ALIVE)