"cacheflow/model_executor/layers/attention.py" did not exist on "7f22f90e8cb423fdaa35203d41badd734d9c2e86"
api_server.py 23.8 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
import json
import time
8
9
from http import HTTPStatus
from typing import AsyncGenerator, Dict, List, Optional, Tuple, Union
Zhuohan Li's avatar
Zhuohan Li committed
10
11

import fastapi
12
import uvicorn
13
from fastapi import BackgroundTasks, Request
Zhuohan Li's avatar
Zhuohan Li committed
14
15
from fastapi.exceptions import RequestValidationError
from fastapi.middleware.cors import CORSMiddleware
Zhuohan Li's avatar
Zhuohan Li committed
16
from fastapi.responses import JSONResponse, StreamingResponse
17
from packaging import version
Zhuohan Li's avatar
Zhuohan Li committed
18

Woosuk Kwon's avatar
Woosuk Kwon committed
19
20
21
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
22
    CompletionRequest, CompletionResponse, CompletionResponseChoice,
23
    CompletionResponseStreamChoice, CompletionStreamResponse,
24
25
26
27
    ChatCompletionRequest, ChatCompletionResponse,
    ChatCompletionResponseChoice, ChatCompletionResponseStreamChoice,
    ChatCompletionStreamResponse, ChatMessage, DeltaMessage, ErrorResponse,
    LogProbs, ModelCard, ModelList, ModelPermission, UsageInfo)
Woosuk Kwon's avatar
Woosuk Kwon committed
28
29
30
from vllm.logger import init_logger
from vllm.outputs import RequestOutput
from vllm.sampling_params import SamplingParams
31
from vllm.transformers_utils.tokenizer import get_tokenizer
Woosuk Kwon's avatar
Woosuk Kwon committed
32
from vllm.utils import random_uuid
Zhuohan Li's avatar
Zhuohan Li committed
33

34
try:
35
    import fastchat
36
37
38
39
40
41
    from fastchat.conversation import Conversation, SeparatorStyle
    from fastchat.model.model_adapter import get_conversation_template
    _fastchat_available = True
except ImportError:
    _fastchat_available = False

42
TIMEOUT_KEEP_ALIVE = 5  # seconds
Zhuohan Li's avatar
Zhuohan Li committed
43
44
45
46

logger = init_logger(__name__)
served_model = None
app = fastapi.FastAPI()
47
engine = None
Zhuohan Li's avatar
Zhuohan Li committed
48
49
50
51


def create_error_response(status_code: HTTPStatus,
                          message: str) -> JSONResponse:
52
53
54
    return JSONResponse(ErrorResponse(message=message,
                                      type="invalid_request_error").dict(),
                        status_code=status_code.value)
Zhuohan Li's avatar
Zhuohan Li committed
55
56
57


@app.exception_handler(RequestValidationError)
58
async def validation_exception_handler(request, exc):  # pylint: disable=unused-argument
Zhuohan Li's avatar
Zhuohan Li committed
59
60
61
62
63
64
65
66
67
68
69
70
71
    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


72
async def get_gen_prompt(request) -> str:
73
74
75
76
77
    if not _fastchat_available:
        raise ModuleNotFoundError(
            "fastchat is not installed. Please install fastchat to use "
            "the chat completion and conversation APIs: `$ pip install fschat`"
        )
78
79
80
81
82
    if version.parse(fastchat.__version__) < version.parse("0.2.23"):
        raise ImportError(
            f"fastchat version is low. Current version: {fastchat.__version__} "
            "Please upgrade fastchat to use: `$ pip install -U fschat`")

83
    conv = get_conversation_template(request.model)
84
85
    conv = Conversation(
        name=conv.name,
86
87
        system_template=conv.system_template,
        system_message=conv.system_message,
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
        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":
104
                conv.system_message = message["content"]
105
106
107
108
109
110
111
112
113
114
115
116
117
118
            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


119
120
121
122
123
124
125
126
127
128
129
130
async def check_length(
    request: Union[ChatCompletionRequest, CompletionRequest],
    prompt: Optional[str] = None,
    prompt_ids: Optional[List[int]] = None
) -> Tuple[List[int], Optional[JSONResponse]]:
    assert (not (prompt is None and prompt_ids is None)
            and not (prompt is not None and prompt_ids is not None)
            ), "Either prompt or prompt_ids should be provided."
    if prompt_ids is not None:
        input_ids = prompt_ids
    else:
        input_ids = tokenizer(prompt).input_ids
131
132
    token_num = len(input_ids)

133
    if token_num + request.max_tokens > max_model_len:
134
        return input_ids, create_error_response(
135
            HTTPStatus.BAD_REQUEST,
136
            f"This model's maximum context length is {max_model_len} tokens. "
137
138
139
140
141
142
            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:
143
        return input_ids, None
144
145


Zhuohan Li's avatar
Zhuohan Li committed
146
147
148
@app.get("/v1/models")
async def show_available_models():
    """Show available models. Right now we only have one model."""
149
150
151
152
153
    model_cards = [
        ModelCard(id=served_model,
                  root=served_model,
                  permission=[ModelPermission()])
    ]
Zhuohan Li's avatar
Zhuohan Li committed
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
    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:
170
171
            logprobs.text_offset.append(logprobs.text_offset[-1] +
                                        last_token_len)
Zhuohan Li's avatar
Zhuohan Li committed
172
173
        last_token_len = len(token)

174
175
176
177
        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
178
179
180
    return logprobs


181
@app.post("/v1/chat/completions")
182
183
async def create_chat_completion(request: ChatCompletionRequest,
                                 raw_request: Request):
184
185
186
187
188
189
190
191
192
193
194
    """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)
    """
    logger.info(f"Received chat completion request: {request}")

195
196
197
    if not engine.is_running:
        engine.start_background_loop()

198
199
200
201
202
203
204
205
206
207
    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)
208
    token_ids, error_check_ret = await check_length(request, prompt=prompt)
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
    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))

232
233
    result_generator = engine.generate(prompt, sampling_params, request_id,
                                       token_ids)
234
235
236
237

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

238
239
240
241
242
    def create_stream_response_json(
        index: int,
        text: str,
        finish_reason: Optional[str] = None,
    ) -> str:
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
        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,
            )
266
267
268
269
270
            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"
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292

        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"
293
        yield "data: [DONE]\n\n"
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

    # 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)
324
325
    num_generated_tokens = sum(
        len(output.token_ids) for output in final_res.outputs)
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
    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)
343

344
345
346
        async def fake_stream_generator() -> AsyncGenerator[str, None]:
            yield f"data: {response_json}\n\n"
            yield "data: [DONE]\n\n"
347

348
349
350
351
352
353
        return StreamingResponse(fake_stream_generator(),
                                 media_type="text/event-stream")

    return response


Zhuohan Li's avatar
Zhuohan Li committed
354
@app.post("/v1/completions")
355
async def create_completion(request: CompletionRequest, raw_request: Request):
356
357
358
359
360
361
    """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
362
        - echo (since the vLLM engine does not currently support
363
364
365
          getting the logprobs of prompt tokens)
        - suffix (the language models we currently support do not support
          suffix)
Woosuk Kwon's avatar
Woosuk Kwon committed
366
        - logit_bias (to be supported by vLLM engine)
367
    """
Zhuohan Li's avatar
Zhuohan Li committed
368
369
    logger.info(f"Received completion request: {request}")

370
371
372
    if not engine.is_running:
        engine.start_background_loop()

Zhuohan Li's avatar
Zhuohan Li committed
373
374
375
376
377
    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
378
        # We do not support echo since the vLLM engine does not
Zhuohan Li's avatar
Zhuohan Li committed
379
380
381
382
383
384
385
        # 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,
386
                                     "suffix is not currently supported")
Zhuohan Li's avatar
Zhuohan Li committed
387
388

    if request.logit_bias is not None:
Woosuk Kwon's avatar
Woosuk Kwon committed
389
        # TODO: support logit_bias in vLLM engine.
Zhuohan Li's avatar
Zhuohan Li committed
390
391
392
393
394
        return create_error_response(HTTPStatus.BAD_REQUEST,
                                     "logit_bias is not currently supported")

    model_name = request.model
    request_id = f"cmpl-{random_uuid()}"
395
396

    use_token_ids = False
397
    if isinstance(request.prompt, list):
398
399
400
        if len(request.prompt) == 0:
            return create_error_response(HTTPStatus.BAD_REQUEST,
                                         "please provide at least one prompt")
401
402
403
404
405
406
407
408
409
410
411
412
        first_element = request.prompt[0]
        if isinstance(first_element, int):
            use_token_ids = True
            prompt = request.prompt
        elif isinstance(first_element, (str, list)):
            # TODO: handles multiple prompt case in list[list[int]]
            if len(request.prompt) > 1:
                return create_error_response(
                    HTTPStatus.BAD_REQUEST,
                    "multiple prompts in a batch is not currently supported")
            use_token_ids = not isinstance(first_element, str)
            prompt = request.prompt[0]
413
414
    else:
        prompt = request.prompt
415

416
417
418
419
    if use_token_ids:
        _, error_check_ret = await check_length(request, prompt_ids=prompt)
    else:
        token_ids, error_check_ret = await check_length(request, prompt=prompt)
420
421
422
    if error_check_ret is not None:
        return error_check_ret

Zhuohan Li's avatar
Zhuohan Li committed
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
    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))

442
443
444
445
446
447
448
449
    if use_token_ids:
        result_generator = engine.generate(None,
                                           sampling_params,
                                           request_id,
                                           prompt_token_ids=prompt)
    else:
        result_generator = engine.generate(prompt, sampling_params, request_id,
                                           token_ids)
Zhuohan Li's avatar
Zhuohan Li committed
450
451
452

    # 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.
453
454
455
    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
456

457
    async def abort_request() -> None:
Zhuohan Li's avatar
Zhuohan Li committed
458
        await engine.abort(request_id)
459

460
461
462
463
464
465
    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
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
        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:
506
507
                    logprobs = (LogProbs()
                                if request.logprobs is not None else None)
Zhuohan Li's avatar
Zhuohan Li committed
508
509
510
511
512
513
514
                    response_json = create_stream_response_json(
                        index=i,
                        text="",
                        logprobs=logprobs,
                        finish_reason=output.finish_reason,
                    )
                    yield f"data: {response_json}\n\n"
515
        yield "data: [DONE]\n\n"
Zhuohan Li's avatar
Zhuohan Li committed
516
517
518

    # Streaming response
    if stream:
519
520
521
        background_tasks = BackgroundTasks()
        # Abort the request if the client disconnects.
        background_tasks.add_task(abort_request)
Zhuohan Li's avatar
Zhuohan Li committed
522
        return StreamingResponse(completion_stream_generator(),
523
524
                                 media_type="text/event-stream",
                                 background=background_tasks)
Zhuohan Li's avatar
Zhuohan Li committed
525
526
527
528

    # Non-streaming response
    final_res: RequestOutput = None
    async for res in result_generator:
529
530
        if await raw_request.is_disconnected():
            # Abort the request if the client disconnects.
531
            await abort_request()
532
533
            return create_error_response(HTTPStatus.BAD_REQUEST,
                                         "Client disconnected")
Zhuohan Li's avatar
Zhuohan Li committed
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
        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)
551
552
    num_generated_tokens = sum(
        len(output.token_ids) for output in final_res.outputs)
Zhuohan Li's avatar
Zhuohan Li committed
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
    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)
570

Zhuohan Li's avatar
Zhuohan Li committed
571
572
573
        async def fake_stream_generator() -> AsyncGenerator[str, None]:
            yield f"data: {response_json}\n\n"
            yield "data: [DONE]\n\n"
574

Zhuohan Li's avatar
Zhuohan Li committed
575
576
577
578
579
580
581
582
        return StreamingResponse(fake_stream_generator(),
                                 media_type="text/event-stream")

    return response


if __name__ == "__main__":
    parser = argparse.ArgumentParser(
583
584
585
586
587
        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
588
    parser.add_argument("--port", type=int, default=8000, help="port number")
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
    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")
604
605
606
607
608
609
    parser.add_argument("--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.")
610

Zhuohan Li's avatar
Zhuohan Li committed
611
    parser = AsyncEngineArgs.add_cli_args(parser)
Zhuohan Li's avatar
Zhuohan Li committed
612
613
614
615
616
617
618
619
620
621
622
623
    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}")

624
625
626
627
628
    if args.served_model_name is not None:
        served_model = args.served_model_name
    else:
        served_model = args.model

Zhuohan Li's avatar
Zhuohan Li committed
629
    engine_args = AsyncEngineArgs.from_cli_args(args)
630
631
    engine = AsyncLLMEngine.from_engine_args(engine_args,
                                             start_engine_loop=False)
632
    engine_model_config = asyncio.run(engine.get_model_config())
633
    max_model_len = engine_model_config.get_max_model_len()
Zhuohan Li's avatar
Zhuohan Li committed
634
635

    # A separate tokenizer to map token IDs to strings.
636
    tokenizer = get_tokenizer(engine_args.tokenizer,
637
638
                              tokenizer_mode=engine_args.tokenizer_mode,
                              trust_remote_code=engine_args.trust_remote_code)
Zhuohan Li's avatar
Zhuohan Li committed
639

640
641
642
643
    uvicorn.run(app,
                host=args.host,
                port=args.port,
                log_level="info",
644
                timeout_keep_alive=TIMEOUT_KEEP_ALIVE)