Skip to content

vllm.entrypoints.openai.api_server

_running_tasks module-attribute 

_running_tasks: set[Task] = set()

args module-attribute 

args = parse_args()

logger module-attribute 

logger = init_logger('vllm.entrypoints.openai.api_server')

parser module-attribute 

parser = FlexibleArgumentParser(
    description="vLLM OpenAI-Compatible RESTful API server."
)

prometheus_multiproc_dir module-attribute 

prometheus_multiproc_dir: TemporaryDirectory

router module-attribute 

router = APIRouter()

AuthenticationMiddleware

Pure ASGI middleware that authenticates each request by checking if the Authorization Bearer token exists and equals anyof "{api_key}".

Notes

There are two cases in which authentication is skipped: 1. The HTTP method is OPTIONS. 2. The request path doesn't start with /v1 (e.g. /health).

Source code in vllm/entrypoints/openai/api_server.py 
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
302
303
304
305
306
307
308
309
310
311
312
class AuthenticationMiddleware:
    """
    Pure ASGI middleware that authenticates each request by checking
    if the Authorization Bearer token exists and equals anyof "{api_key}".

    Notes
    -----
    There are two cases in which authentication is skipped:
        1. The HTTP method is OPTIONS.
        2. The request path doesn't start with /v1 (e.g. /health).
    """

    def __init__(self, app: ASGIApp, tokens: list[str]) -> None:
        self.app = app
        self.api_tokens = [hashlib.sha256(t.encode("utf-8")).digest() for t in tokens]

    def verify_token(self, headers: Headers) -> bool:
        authorization_header_value = headers.get("Authorization")
        if not authorization_header_value:
            return False

        scheme, _, param = authorization_header_value.partition(" ")
        if scheme.lower() != "bearer":
            return False

        param_hash = hashlib.sha256(param.encode("utf-8")).digest()

        token_match = False
        for token_hash in self.api_tokens:
            token_match |= secrets.compare_digest(param_hash, token_hash)

        return token_match

    def __call__(self, scope: Scope, receive: Receive, send: Send) -> Awaitable[None]:
        if scope["type"] not in ("http", "websocket") or scope["method"] == "OPTIONS":
            # scope["type"] can be "lifespan" or "startup" for example,
            # in which case we don't need to do anything
            return self.app(scope, receive, send)
        root_path = scope.get("root_path", "")
        url_path = URL(scope=scope).path.removeprefix(root_path)
        headers = Headers(scope=scope)
        # Type narrow to satisfy mypy.
        if url_path.startswith("/v1") and not self.verify_token(headers):
            response = JSONResponse(content={"error": "Unauthorized"}, status_code=401)
            return response(scope, receive, send)
        return self.app(scope, receive, send)

api_tokens instance-attribute 

api_tokens = [(digest()) for t in tokens]

app instance-attribute 

app = app

__call__ 

__call__(
    scope: Scope, receive: Receive, send: Send
) -> Awaitable[None]
Source code in vllm/entrypoints/openai/api_server.py 
300
301
302
303
304
305
306
307
308
309
310
311
312
def __call__(self, scope: Scope, receive: Receive, send: Send) -> Awaitable[None]:
    if scope["type"] not in ("http", "websocket") or scope["method"] == "OPTIONS":
        # scope["type"] can be "lifespan" or "startup" for example,
        # in which case we don't need to do anything
        return self.app(scope, receive, send)
    root_path = scope.get("root_path", "")
    url_path = URL(scope=scope).path.removeprefix(root_path)
    headers = Headers(scope=scope)
    # Type narrow to satisfy mypy.
    if url_path.startswith("/v1") and not self.verify_token(headers):
        response = JSONResponse(content={"error": "Unauthorized"}, status_code=401)
        return response(scope, receive, send)
    return self.app(scope, receive, send)

__init__ 

__init__(app: ASGIApp, tokens: list[str]) -> None
Source code in vllm/entrypoints/openai/api_server.py 
279
280
281
def __init__(self, app: ASGIApp, tokens: list[str]) -> None:
    self.app = app
    self.api_tokens = [hashlib.sha256(t.encode("utf-8")).digest() for t in tokens]

verify_token 

verify_token(headers: Headers) -> bool
Source code in vllm/entrypoints/openai/api_server.py 
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
def verify_token(self, headers: Headers) -> bool:
    authorization_header_value = headers.get("Authorization")
    if not authorization_header_value:
        return False

    scheme, _, param = authorization_header_value.partition(" ")
    if scheme.lower() != "bearer":
        return False

    param_hash = hashlib.sha256(param.encode("utf-8")).digest()

    token_match = False
    for token_hash in self.api_tokens:
        token_match |= secrets.compare_digest(param_hash, token_hash)

    return token_match

SSEDecoder

Robust Server-Sent Events decoder for streaming responses.

Source code in vllm/entrypoints/openai/api_server.py 
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
class SSEDecoder:
    """Robust Server-Sent Events decoder for streaming responses."""

    def __init__(self):
        self.buffer = ""
        self.content_buffer = []

    def decode_chunk(self, chunk: bytes) -> list[dict]:
        """Decode a chunk of SSE data and return parsed events."""
        import json

        try:
            chunk_str = chunk.decode("utf-8")
        except UnicodeDecodeError:
            # Skip malformed chunks
            return []

        self.buffer += chunk_str
        events = []

        # Process complete lines
        while "\n" in self.buffer:
            line, self.buffer = self.buffer.split("\n", 1)
            line = line.rstrip("\r")  # Handle CRLF

            if line.startswith("data: "):
                data_str = line[6:].strip()
                if data_str == "[DONE]":
                    events.append({"type": "done"})
                elif data_str:
                    try:
                        event_data = json.loads(data_str)
                        events.append({"type": "data", "data": event_data})
                    except json.JSONDecodeError:
                        # Skip malformed JSON
                        continue

        return events

    def extract_content(self, event_data: dict) -> str:
        """Extract content from event data."""
        return _extract_content_from_chunk(event_data)

    def add_content(self, content: str) -> None:
        """Add content to the buffer."""
        if content:
            self.content_buffer.append(content)

    def get_complete_content(self) -> str:
        """Get the complete buffered content."""
        return "".join(self.content_buffer)

buffer instance-attribute 

buffer = ''

content_buffer instance-attribute 

content_buffer = []

__init__ 

__init__()
Source code in vllm/entrypoints/openai/api_server.py 
379
380
381
def __init__(self):
    self.buffer = ""
    self.content_buffer = []

add_content 

add_content(content: str) -> None

Add content to the buffer.

Source code in vllm/entrypoints/openai/api_server.py 
419
420
421
422
def add_content(self, content: str) -> None:
    """Add content to the buffer."""
    if content:
        self.content_buffer.append(content)

decode_chunk 

decode_chunk(chunk: bytes) -> list[dict]

Decode a chunk of SSE data and return parsed events.

Source code in vllm/entrypoints/openai/api_server.py 
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
def decode_chunk(self, chunk: bytes) -> list[dict]:
    """Decode a chunk of SSE data and return parsed events."""
    import json

    try:
        chunk_str = chunk.decode("utf-8")
    except UnicodeDecodeError:
        # Skip malformed chunks
        return []

    self.buffer += chunk_str
    events = []

    # Process complete lines
    while "\n" in self.buffer:
        line, self.buffer = self.buffer.split("\n", 1)
        line = line.rstrip("\r")  # Handle CRLF

        if line.startswith("data: "):
            data_str = line[6:].strip()
            if data_str == "[DONE]":
                events.append({"type": "done"})
            elif data_str:
                try:
                    event_data = json.loads(data_str)
                    events.append({"type": "data", "data": event_data})
                except json.JSONDecodeError:
                    # Skip malformed JSON
                    continue

    return events

extract_content 

extract_content(event_data: dict) -> str

Extract content from event data.

Source code in vllm/entrypoints/openai/api_server.py 
415
416
417
def extract_content(self, event_data: dict) -> str:
    """Extract content from event data."""
    return _extract_content_from_chunk(event_data)

get_complete_content 

get_complete_content() -> str

Get the complete buffered content.

Source code in vllm/entrypoints/openai/api_server.py 
424
425
426
def get_complete_content(self) -> str:
    """Get the complete buffered content."""
    return "".join(self.content_buffer)

XRequestIdMiddleware

Middleware the set's the X-Request-Id header for each response to a random uuid4 (hex) value if the header isn't already present in the request, otherwise use the provided request id.

Source code in vllm/entrypoints/openai/api_server.py 
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
class XRequestIdMiddleware:
    """
    Middleware the set's the X-Request-Id header for each response
    to a random uuid4 (hex) value if the header isn't already
    present in the request, otherwise use the provided request id.
    """

    def __init__(self, app: ASGIApp) -> None:
        self.app = app

    def __call__(self, scope: Scope, receive: Receive, send: Send) -> Awaitable[None]:
        if scope["type"] not in ("http", "websocket"):
            return self.app(scope, receive, send)

        # Extract the request headers.
        request_headers = Headers(scope=scope)

        async def send_with_request_id(message: Message) -> None:
            """
            Custom send function to mutate the response headers
            and append X-Request-Id to it.
            """
            if message["type"] == "http.response.start":
                response_headers = MutableHeaders(raw=message["headers"])
                request_id = request_headers.get("X-Request-Id", uuid.uuid4().hex)
                response_headers.append("X-Request-Id", request_id)
            await send(message)

        return self.app(scope, receive, send_with_request_id)

app instance-attribute 

app = app

__call__ 

__call__(
    scope: Scope, receive: Receive, send: Send
) -> Awaitable[None]
Source code in vllm/entrypoints/openai/api_server.py 
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
def __call__(self, scope: Scope, receive: Receive, send: Send) -> Awaitable[None]:
    if scope["type"] not in ("http", "websocket"):
        return self.app(scope, receive, send)

    # Extract the request headers.
    request_headers = Headers(scope=scope)

    async def send_with_request_id(message: Message) -> None:
        """
        Custom send function to mutate the response headers
        and append X-Request-Id to it.
        """
        if message["type"] == "http.response.start":
            response_headers = MutableHeaders(raw=message["headers"])
            request_id = request_headers.get("X-Request-Id", uuid.uuid4().hex)
            response_headers.append("X-Request-Id", request_id)
        await send(message)

    return self.app(scope, receive, send_with_request_id)

__init__ 

__init__(app: ASGIApp) -> None
Source code in vllm/entrypoints/openai/api_server.py 
322
323
def __init__(self, app: ASGIApp) -> None:
    self.app = app

_extract_content_from_chunk 

_extract_content_from_chunk(chunk_data: dict) -> str

Extract content from a streaming response chunk.

Source code in vllm/entrypoints/openai/api_server.py 
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
def _extract_content_from_chunk(chunk_data: dict) -> str:
    """Extract content from a streaming response chunk."""
    try:
        from vllm.entrypoints.openai.chat_completion.protocol import (
            ChatCompletionStreamResponse,
        )
        from vllm.entrypoints.openai.completion.protocol import (
            CompletionStreamResponse,
        )

        # Try using Completion types for type-safe parsing
        if chunk_data.get("object") == "chat.completion.chunk":
            chat_response = ChatCompletionStreamResponse.model_validate(chunk_data)
            if chat_response.choices and chat_response.choices[0].delta.content:
                return chat_response.choices[0].delta.content
        elif chunk_data.get("object") == "text_completion":
            completion_response = CompletionStreamResponse.model_validate(chunk_data)
            if completion_response.choices and completion_response.choices[0].text:
                return completion_response.choices[0].text
    except pydantic.ValidationError:
        # Fallback to manual parsing
        if "choices" in chunk_data and chunk_data["choices"]:
            choice = chunk_data["choices"][0]
            if "delta" in choice and choice["delta"].get("content"):
                return choice["delta"]["content"]
            elif choice.get("text"):
                return choice["text"]
    return ""

_log_non_streaming_response 

_log_non_streaming_response(response_body: list) -> None

Log non-streaming response.

Source code in vllm/entrypoints/openai/api_server.py 
474
475
476
477
478
479
480
def _log_non_streaming_response(response_body: list) -> None:
    """Log non-streaming response."""
    try:
        decoded_body = response_body[0].decode()
        logger.info("response_body={%s}", decoded_body)
    except UnicodeDecodeError:
        logger.info("response_body={<binary_data>}")

_log_streaming_response 

_log_streaming_response(
    response, response_body: list
) -> None

Log streaming response with robust SSE parsing.

Source code in vllm/entrypoints/openai/api_server.py 
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
458
459
460
461
462
463
464
465
466
467
468
469
470
471
def _log_streaming_response(response, response_body: list) -> None:
    """Log streaming response with robust SSE parsing."""
    from starlette.concurrency import iterate_in_threadpool

    sse_decoder = SSEDecoder()
    chunk_count = 0

    def buffered_iterator():
        nonlocal chunk_count

        for chunk in response_body:
            chunk_count += 1
            yield chunk

            # Parse SSE events from chunk
            events = sse_decoder.decode_chunk(chunk)

            for event in events:
                if event["type"] == "data":
                    content = sse_decoder.extract_content(event["data"])
                    sse_decoder.add_content(content)
                elif event["type"] == "done":
                    # Log complete content when done
                    full_content = sse_decoder.get_complete_content()
                    if full_content:
                        # Truncate if too long
                        if len(full_content) > 2048:
                            full_content = full_content[:2048] + ""
                            "...[truncated]"
                        logger.info(
                            "response_body={streaming_complete: content=%r, chunks=%d}",
                            full_content,
                            chunk_count,
                        )
                    else:
                        logger.info(
                            "response_body={streaming_complete: no_content, chunks=%d}",
                            chunk_count,
                        )
                    return

    response.body_iterator = iterate_in_threadpool(buffered_iterator())
    logger.info("response_body={streaming_started: chunks=%d}", len(response_body))

base 

base(request: Request) -> OpenAIServing
Source code in vllm/entrypoints/openai/api_server.py 
212
213
214
def base(request: Request) -> OpenAIServing:
    # Reuse the existing instance
    return tokenization(request)

build_app 

build_app(args: Namespace) -> FastAPI
Source code in vllm/entrypoints/openai/api_server.py 
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
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
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
def build_app(args: Namespace) -> FastAPI:
    if args.disable_fastapi_docs:
        app = FastAPI(
            openapi_url=None, docs_url=None, redoc_url=None, lifespan=lifespan
        )
    elif args.enable_offline_docs:
        app = FastAPI(docs_url=None, redoc_url=None, lifespan=lifespan)
    else:
        app = FastAPI(lifespan=lifespan)
    app.state.args = args
    from vllm.entrypoints.serve import register_vllm_serve_api_routers

    register_vllm_serve_api_routers(app)
    from vllm.entrypoints.openai.chat_completion.api_router import (
        attach_router as register_chat_api_router,
    )

    register_chat_api_router(app)

    from vllm.entrypoints.openai.responses.api_router import (
        attach_router as register_responses_api_router,
    )

    register_responses_api_router(app)
    from vllm.entrypoints.openai.translations.api_router import (
        attach_router as register_translations_api_router,
    )

    register_translations_api_router(app)

    from vllm.entrypoints.openai.completion.api_router import (
        attach_router as register_completion_api_router,
    )

    register_completion_api_router(app)
    from vllm.entrypoints.anthropic.api_router import (
        attach_router as register_anthropic_api_router,
    )

    register_anthropic_api_router(app)
    from vllm.entrypoints.openai.models.api_router import (
        attach_router as register_models_api_router,
    )

    register_models_api_router(app)
    from vllm.entrypoints.sagemaker.routes import register_sagemaker_routes

    register_sagemaker_routes(router)
    app.include_router(router)

    app.root_path = args.root_path

    from vllm.entrypoints.pooling import register_pooling_api_routers

    register_pooling_api_routers(app)

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

    @app.exception_handler(HTTPException)
    async def http_exception_handler(_: Request, exc: HTTPException):
        err = ErrorResponse(
            error=ErrorInfo(
                message=sanitize_message(exc.detail),
                type=HTTPStatus(exc.status_code).phrase,
                code=exc.status_code,
            )
        )
        return JSONResponse(err.model_dump(), status_code=exc.status_code)

    @app.exception_handler(RequestValidationError)
    async def validation_exception_handler(_: Request, exc: RequestValidationError):
        param = None
        for error in exc.errors():
            if "ctx" in error and "error" in error["ctx"]:
                ctx_error = error["ctx"]["error"]
                if isinstance(ctx_error, VLLMValidationError):
                    param = ctx_error.parameter
                    break

        exc_str = str(exc)
        errors_str = str(exc.errors())

        if exc.errors() and errors_str and errors_str != exc_str:
            message = f"{exc_str} {errors_str}"
        else:
            message = exc_str

        err = ErrorResponse(
            error=ErrorInfo(
                message=sanitize_message(message),
                type=HTTPStatus.BAD_REQUEST.phrase,
                code=HTTPStatus.BAD_REQUEST,
                param=param,
            )
        )
        return JSONResponse(err.model_dump(), status_code=HTTPStatus.BAD_REQUEST)

    # Ensure --api-key option from CLI takes precedence over VLLM_API_KEY
    if tokens := [key for key in (args.api_key or [envs.VLLM_API_KEY]) if key]:
        app.add_middleware(AuthenticationMiddleware, tokens=tokens)

    if args.enable_request_id_headers:
        app.add_middleware(XRequestIdMiddleware)

    # Add scaling middleware to check for scaling state
    app.add_middleware(ScalingMiddleware)

    if envs.VLLM_DEBUG_LOG_API_SERVER_RESPONSE:
        logger.warning(
            "CAUTION: Enabling log response in the API Server. "
            "This can include sensitive information and should be "
            "avoided in production."
        )

        @app.middleware("http")
        async def log_response(request: Request, call_next):
            response = await call_next(request)
            response_body = [section async for section in response.body_iterator]
            response.body_iterator = iterate_in_threadpool(iter(response_body))
            # Check if this is a streaming response by looking at content-type
            content_type = response.headers.get("content-type", "")
            is_streaming = content_type == "text/event-stream; charset=utf-8"

            # Log response body based on type
            if not response_body:
                logger.info("response_body={<empty>}")
            elif is_streaming:
                _log_streaming_response(response, response_body)
            else:
                _log_non_streaming_response(response_body)
            return response

    for middleware in args.middleware:
        module_path, object_name = middleware.rsplit(".", 1)
        imported = getattr(importlib.import_module(module_path), object_name)
        if inspect.isclass(imported):
            app.add_middleware(imported)  # type: ignore[arg-type]
        elif inspect.iscoroutinefunction(imported):
            app.middleware("http")(imported)
        else:
            raise ValueError(
                f"Invalid middleware {middleware}. Must be a function or a class."
            )

    app = sagemaker_standards.bootstrap(app)

    return app

build_async_engine_client async 

build_async_engine_client(
    args: Namespace,
    *,
    usage_context: UsageContext = OPENAI_API_SERVER,
    disable_frontend_multiprocessing: bool | None = None,
    client_config: dict[str, Any] | None = None,
) -> AsyncIterator[EngineClient]
Source code in vllm/entrypoints/openai/api_server.py 
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
@asynccontextmanager
async def build_async_engine_client(
    args: Namespace,
    *,
    usage_context: UsageContext = UsageContext.OPENAI_API_SERVER,
    disable_frontend_multiprocessing: bool | None = None,
    client_config: dict[str, Any] | None = None,
) -> AsyncIterator[EngineClient]:
    if os.getenv("VLLM_WORKER_MULTIPROC_METHOD") == "forkserver":
        # The executor is expected to be mp.
        # Pre-import heavy modules in the forkserver process
        logger.debug("Setup forkserver with pre-imports")
        multiprocessing.set_start_method("forkserver")
        multiprocessing.set_forkserver_preload(["vllm.v1.engine.async_llm"])
        forkserver.ensure_running()
        logger.debug("Forkserver setup complete!")

    # Context manager to handle engine_client lifecycle
    # Ensures everything is shutdown and cleaned up on error/exit
    engine_args = AsyncEngineArgs.from_cli_args(args)
    if client_config:
        engine_args._api_process_count = client_config.get("client_count", 1)
        engine_args._api_process_rank = client_config.get("client_index", 0)

    if disable_frontend_multiprocessing is None:
        disable_frontend_multiprocessing = bool(args.disable_frontend_multiprocessing)

    async with build_async_engine_client_from_engine_args(
        engine_args,
        usage_context=usage_context,
        disable_frontend_multiprocessing=disable_frontend_multiprocessing,
        client_config=client_config,
    ) as engine:
        yield engine

build_async_engine_client_from_engine_args async 

build_async_engine_client_from_engine_args(
    engine_args: AsyncEngineArgs,
    *,
    usage_context: UsageContext = OPENAI_API_SERVER,
    disable_frontend_multiprocessing: bool = False,
    client_config: dict[str, Any] | None = None,
) -> AsyncIterator[EngineClient]

Create EngineClient, either: - in-process using the AsyncLLMEngine Directly - multiprocess using AsyncLLMEngine RPC

Returns the Client or None if the creation failed.

Source code in vllm/entrypoints/openai/api_server.py 
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
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
@asynccontextmanager
async def build_async_engine_client_from_engine_args(
    engine_args: AsyncEngineArgs,
    *,
    usage_context: UsageContext = UsageContext.OPENAI_API_SERVER,
    disable_frontend_multiprocessing: bool = False,
    client_config: dict[str, Any] | None = None,
) -> AsyncIterator[EngineClient]:
    """
    Create EngineClient, either:
        - in-process using the AsyncLLMEngine Directly
        - multiprocess using AsyncLLMEngine RPC

    Returns the Client or None if the creation failed.
    """

    # Create the EngineConfig (determines if we can use V1).
    vllm_config = engine_args.create_engine_config(usage_context=usage_context)

    if disable_frontend_multiprocessing:
        logger.warning("V1 is enabled, but got --disable-frontend-multiprocessing.")

    from vllm.v1.engine.async_llm import AsyncLLM

    async_llm: AsyncLLM | None = None

    # Don't mutate the input client_config
    client_config = dict(client_config) if client_config else {}
    client_count = client_config.pop("client_count", 1)
    client_index = client_config.pop("client_index", 0)

    try:
        async_llm = AsyncLLM.from_vllm_config(
            vllm_config=vllm_config,
            usage_context=usage_context,
            enable_log_requests=engine_args.enable_log_requests,
            aggregate_engine_logging=engine_args.aggregate_engine_logging,
            disable_log_stats=engine_args.disable_log_stats,
            client_addresses=client_config,
            client_count=client_count,
            client_index=client_index,
        )

        # Don't keep the dummy data in memory
        assert async_llm is not None
        await async_llm.reset_mm_cache()

        yield async_llm
    finally:
        if async_llm:
            async_llm.shutdown()

create_server_socket 

create_server_socket(addr: tuple[str, int]) -> socket
Source code in vllm/entrypoints/openai/api_server.py 
830
831
832
833
834
835
836
837
838
839
840
def create_server_socket(addr: tuple[str, int]) -> socket.socket:
    family = socket.AF_INET
    if is_valid_ipv6_address(addr[0]):
        family = socket.AF_INET6

    sock = socket.socket(family=family, type=socket.SOCK_STREAM)
    sock.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1)
    sock.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEPORT, 1)
    sock.bind(addr)

    return sock

create_server_unix_socket 

create_server_unix_socket(path: str) -> socket
Source code in vllm/entrypoints/openai/api_server.py 
843
844
845
846
def create_server_unix_socket(path: str) -> socket.socket:
    sock = socket.socket(family=socket.AF_UNIX, type=socket.SOCK_STREAM)
    sock.bind(path)
    return sock

engine_client 

engine_client(request: Request) -> EngineClient
Source code in vllm/entrypoints/openai/api_server.py 
221
222
def engine_client(request: Request) -> EngineClient:
    return request.app.state.engine_client

get_server_load_metrics async 

get_server_load_metrics(request: Request)
Source code in vllm/entrypoints/openai/api_server.py 
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
@router.get("/load")
async def get_server_load_metrics(request: Request):
    # This endpoint returns the current server load metrics.
    # It tracks requests utilizing the GPU from the following routes:
    # - /v1/responses
    # - /v1/responses/{response_id}
    # - /v1/responses/{response_id}/cancel
    # - /v1/messages
    # - /v1/chat/completions
    # - /v1/completions
    # - /v1/audio/transcriptions
    # - /v1/audio/translations
    # - /v1/embeddings
    # - /pooling
    # - /classify
    # - /score
    # - /v1/score
    # - /rerank
    # - /v1/rerank
    # - /v2/rerank
    return JSONResponse(content={"server_load": request.app.state.server_load_metrics})

init_app_state async 

init_app_state(
    engine_client: EngineClient,
    state: State,
    args: Namespace,
) -> None
Source code in vllm/entrypoints/openai/api_server.py 
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
async def init_app_state(
    engine_client: EngineClient,
    state: State,
    args: Namespace,
) -> None:
    vllm_config = engine_client.vllm_config

    if args.served_model_name is not None:
        served_model_names = args.served_model_name
    else:
        served_model_names = [args.model]

    if args.enable_log_requests:
        request_logger = RequestLogger(max_log_len=args.max_log_len)
    else:
        request_logger = None

    base_model_paths = [
        BaseModelPath(name=name, model_path=args.model) for name in served_model_names
    ]

    state.engine_client = engine_client
    state.log_stats = not args.disable_log_stats
    state.vllm_config = vllm_config
    state.args = args
    supported_tasks = await engine_client.get_supported_tasks()
    logger.info("Supported tasks: %s", supported_tasks)

    resolved_chat_template = load_chat_template(args.chat_template)

    if args.tool_server == "demo":
        tool_server: ToolServer | None = DemoToolServer()
        assert isinstance(tool_server, DemoToolServer)
        await tool_server.init_and_validate()
    elif args.tool_server:
        tool_server = MCPToolServer()
        await tool_server.add_tool_server(args.tool_server)
    else:
        tool_server = None

    # Merge default_mm_loras into the static lora_modules
    default_mm_loras = (
        vllm_config.lora_config.default_mm_loras
        if vllm_config.lora_config is not None
        else {}
    )

    default_mm_loras = (
        vllm_config.lora_config.default_mm_loras
        if vllm_config.lora_config is not None
        else {}
    )
    lora_modules = process_lora_modules(args.lora_modules, default_mm_loras)

    state.openai_serving_models = OpenAIServingModels(
        engine_client=engine_client,
        base_model_paths=base_model_paths,
        lora_modules=lora_modules,
    )
    await state.openai_serving_models.init_static_loras()
    state.openai_serving_responses = (
        OpenAIServingResponses(
            engine_client,
            state.openai_serving_models,
            request_logger=request_logger,
            chat_template=resolved_chat_template,
            chat_template_content_format=args.chat_template_content_format,
            return_tokens_as_token_ids=args.return_tokens_as_token_ids,
            enable_auto_tools=args.enable_auto_tool_choice,
            tool_parser=args.tool_call_parser,
            tool_server=tool_server,
            reasoning_parser=args.structured_outputs_config.reasoning_parser,
            enable_prompt_tokens_details=args.enable_prompt_tokens_details,
            enable_force_include_usage=args.enable_force_include_usage,
            enable_log_outputs=args.enable_log_outputs,
            log_error_stack=args.log_error_stack,
        )
        if "generate" in supported_tasks
        else None
    )
    state.openai_serving_chat = (
        OpenAIServingChat(
            engine_client,
            state.openai_serving_models,
            args.response_role,
            request_logger=request_logger,
            chat_template=resolved_chat_template,
            chat_template_content_format=args.chat_template_content_format,
            default_chat_template_kwargs=args.default_chat_template_kwargs,
            trust_request_chat_template=args.trust_request_chat_template,
            return_tokens_as_token_ids=args.return_tokens_as_token_ids,
            enable_auto_tools=args.enable_auto_tool_choice,
            exclude_tools_when_tool_choice_none=args.exclude_tools_when_tool_choice_none,
            tool_parser=args.tool_call_parser,
            reasoning_parser=args.structured_outputs_config.reasoning_parser,
            enable_prompt_tokens_details=args.enable_prompt_tokens_details,
            enable_force_include_usage=args.enable_force_include_usage,
            enable_log_outputs=args.enable_log_outputs,
            enable_log_deltas=args.enable_log_deltas,
            log_error_stack=args.log_error_stack,
        )
        if "generate" in supported_tasks
        else None
    )
    # Warm up chat template processing to avoid first-request latency
    if state.openai_serving_chat is not None:
        await state.openai_serving_chat.warmup()
    state.openai_serving_completion = (
        OpenAIServingCompletion(
            engine_client,
            state.openai_serving_models,
            request_logger=request_logger,
            return_tokens_as_token_ids=args.return_tokens_as_token_ids,
            enable_prompt_tokens_details=args.enable_prompt_tokens_details,
            enable_force_include_usage=args.enable_force_include_usage,
            log_error_stack=args.log_error_stack,
        )
        if "generate" in supported_tasks
        else None
    )
    state.openai_serving_tokenization = OpenAIServingTokenization(
        engine_client,
        state.openai_serving_models,
        request_logger=request_logger,
        chat_template=resolved_chat_template,
        chat_template_content_format=args.chat_template_content_format,
        trust_request_chat_template=args.trust_request_chat_template,
        log_error_stack=args.log_error_stack,
    )
    state.openai_serving_transcription = (
        OpenAIServingTranscription(
            engine_client,
            state.openai_serving_models,
            request_logger=request_logger,
            log_error_stack=args.log_error_stack,
            enable_force_include_usage=args.enable_force_include_usage,
        )
        if "transcription" in supported_tasks
        else None
    )
    state.openai_serving_translation = (
        OpenAIServingTranslation(
            engine_client,
            state.openai_serving_models,
            request_logger=request_logger,
            log_error_stack=args.log_error_stack,
            enable_force_include_usage=args.enable_force_include_usage,
        )
        if "transcription" in supported_tasks
        else None
    )
    state.anthropic_serving_messages = (
        AnthropicServingMessages(
            engine_client,
            state.openai_serving_models,
            args.response_role,
            request_logger=request_logger,
            chat_template=resolved_chat_template,
            chat_template_content_format=args.chat_template_content_format,
            return_tokens_as_token_ids=args.return_tokens_as_token_ids,
            enable_auto_tools=args.enable_auto_tool_choice,
            tool_parser=args.tool_call_parser,
            reasoning_parser=args.structured_outputs_config.reasoning_parser,
            enable_prompt_tokens_details=args.enable_prompt_tokens_details,
            enable_force_include_usage=args.enable_force_include_usage,
        )
        if "generate" in supported_tasks
        else None
    )
    state.serving_tokens = (
        ServingTokens(
            engine_client,
            state.openai_serving_models,
            request_logger=request_logger,
            return_tokens_as_token_ids=args.return_tokens_as_token_ids,
            log_error_stack=args.log_error_stack,
            enable_prompt_tokens_details=args.enable_prompt_tokens_details,
            enable_log_outputs=args.enable_log_outputs,
            force_no_detokenize=args.tokens_only,
        )
        if "generate" in supported_tasks
        else None
    )

    from vllm.entrypoints.pooling import init_pooling_state

    await init_pooling_state(engine_client, state, args)

    state.enable_server_load_tracking = args.enable_server_load_tracking
    state.server_load_metrics = 0

lifespan async 

lifespan(app: FastAPI)
Source code in vllm/entrypoints/openai/api_server.py 
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
@asynccontextmanager
async def lifespan(app: FastAPI):
    try:
        if app.state.log_stats:
            engine_client: EngineClient = app.state.engine_client

            async def _force_log():
                while True:
                    await asyncio.sleep(envs.VLLM_LOG_STATS_INTERVAL)
                    await engine_client.do_log_stats()

            task = asyncio.create_task(_force_log())
            _running_tasks.add(task)
            task.add_done_callback(_running_tasks.remove)
        else:
            task = None

        # Mark the startup heap as static so that it's ignored by GC.
        # Reduces pause times of oldest generation collections.
        freeze_gc_heap()
        try:
            yield
        finally:
            if task is not None:
                task.cancel()
    finally:
        # Ensure app state including engine ref is gc'd
        del app.state

load_log_config 

load_log_config(log_config_file: str | None) -> dict | None
Source code in vllm/entrypoints/openai/api_server.py 
254
255
256
257
258
259
260
261
262
263
264
def load_log_config(log_config_file: str | None) -> dict | None:
    if not log_config_file:
        return None
    try:
        with open(log_config_file) as f:
            return json.load(f)
    except Exception as e:
        logger.warning(
            "Failed to load log config from file %s: error %s", log_config_file, e
        )
        return None

run_server async 

run_server(args, **uvicorn_kwargs) -> None

Run a single-worker API server.

Source code in vllm/entrypoints/openai/api_server.py 
911
912
913
914
915
916
917
918
async def run_server(args, **uvicorn_kwargs) -> None:
    """Run a single-worker API server."""

    # Add process-specific prefix to stdout and stderr.
    decorate_logs("APIServer")

    listen_address, sock = setup_server(args)
    await run_server_worker(listen_address, sock, args, **uvicorn_kwargs)

run_server_worker async 

run_server_worker(
    listen_address,
    sock,
    args,
    client_config=None,
    **uvicorn_kwargs,
) -> None

Run a single API server worker.

Source code in vllm/entrypoints/openai/api_server.py 
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
async def run_server_worker(
    listen_address, sock, args, client_config=None, **uvicorn_kwargs
) -> None:
    """Run a single API server worker."""

    if args.tool_parser_plugin and len(args.tool_parser_plugin) > 3:
        ToolParserManager.import_tool_parser(args.tool_parser_plugin)

    if args.reasoning_parser_plugin and len(args.reasoning_parser_plugin) > 3:
        ReasoningParserManager.import_reasoning_parser(args.reasoning_parser_plugin)

    # Load logging config for uvicorn if specified
    log_config = load_log_config(args.log_config_file)
    if log_config is not None:
        uvicorn_kwargs["log_config"] = log_config

    async with build_async_engine_client(
        args,
        client_config=client_config,
    ) as engine_client:
        app = build_app(args)

        await init_app_state(engine_client, app.state, args)

        logger.info(
            "Starting vLLM API server %d on %s",
            engine_client.vllm_config.parallel_config._api_process_rank,
            listen_address,
        )
        shutdown_task = await serve_http(
            app,
            sock=sock,
            enable_ssl_refresh=args.enable_ssl_refresh,
            host=args.host,
            port=args.port,
            log_level=args.uvicorn_log_level,
            # NOTE: When the 'disable_uvicorn_access_log' value is True,
            # no access log will be output.
            access_log=not args.disable_uvicorn_access_log,
            timeout_keep_alive=envs.VLLM_HTTP_TIMEOUT_KEEP_ALIVE,
            ssl_keyfile=args.ssl_keyfile,
            ssl_certfile=args.ssl_certfile,
            ssl_ca_certs=args.ssl_ca_certs,
            ssl_cert_reqs=args.ssl_cert_reqs,
            ssl_ciphers=args.ssl_ciphers,
            h11_max_incomplete_event_size=args.h11_max_incomplete_event_size,
            h11_max_header_count=args.h11_max_header_count,
            **uvicorn_kwargs,
        )

    # NB: Await server shutdown only after the backend context is exited
    try:
        await shutdown_task
    finally:
        sock.close()

setup_server 

setup_server(args)

Validate API server args, set up signal handler, create socket ready to serve.

Source code in vllm/entrypoints/openai/api_server.py 
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
def setup_server(args):
    """Validate API server args, set up signal handler, create socket
    ready to serve."""

    log_version_and_model(logger, VLLM_VERSION, args.model)
    log_non_default_args(args)

    if args.tool_parser_plugin and len(args.tool_parser_plugin) > 3:
        ToolParserManager.import_tool_parser(args.tool_parser_plugin)

    if args.reasoning_parser_plugin and len(args.reasoning_parser_plugin) > 3:
        ReasoningParserManager.import_reasoning_parser(args.reasoning_parser_plugin)

    validate_api_server_args(args)

    # workaround to make sure that we bind the port before the engine is set up.
    # This avoids race conditions with ray.
    # see https://github.com/vllm-project/vllm/issues/8204
    if args.uds:
        sock = create_server_unix_socket(args.uds)
    else:
        sock_addr = (args.host or "", args.port)
        sock = create_server_socket(sock_addr)

    # workaround to avoid footguns where uvicorn drops requests with too
    # many concurrent requests active
    set_ulimit()

    def signal_handler(*_) -> None:
        # Interrupt server on sigterm while initializing
        raise KeyboardInterrupt("terminated")

    signal.signal(signal.SIGTERM, signal_handler)

    if args.uds:
        listen_address = f"unix:{args.uds}"
    else:
        addr, port = sock_addr
        is_ssl = args.ssl_keyfile and args.ssl_certfile
        host_part = f"[{addr}]" if is_valid_ipv6_address(addr) else addr or "0.0.0.0"
        listen_address = f"http{'s' if is_ssl else ''}://{host_part}:{port}"
    return listen_address, sock

show_version async 

show_version()
Source code in vllm/entrypoints/openai/api_server.py 
248
249
250
251
@router.get("/version")
async def show_version():
    ver = {"version": VLLM_VERSION}
    return JSONResponse(content=ver)

tokenization 

tokenization(request: Request) -> OpenAIServingTokenization
Source code in vllm/entrypoints/openai/api_server.py 
217
218
def tokenization(request: Request) -> OpenAIServingTokenization:
    return request.app.state.openai_serving_tokenization

validate_api_server_args 

validate_api_server_args(args)
Source code in vllm/entrypoints/openai/api_server.py 
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
def validate_api_server_args(args):
    valid_tool_parses = ToolParserManager.list_registered()
    if args.enable_auto_tool_choice and args.tool_call_parser not in valid_tool_parses:
        raise KeyError(
            f"invalid tool call parser: {args.tool_call_parser} "
            f"(chose from {{ {','.join(valid_tool_parses)} }})"
        )

    valid_reasoning_parsers = ReasoningParserManager.list_registered()
    if (
        reasoning_parser := args.structured_outputs_config.reasoning_parser
    ) and reasoning_parser not in valid_reasoning_parsers:
        raise KeyError(
            f"invalid reasoning parser: {reasoning_parser} "
            f"(chose from {{ {','.join(valid_reasoning_parsers)} }})"
        )