init

vllm/entrypoints/api_server.py

@@ -0,0 +1,119 @@
+"""
+NOTE: This API server is used only for demonstrating usage of AsyncEngine
+and simple performance benchmarks. It is not intended for production use.
+For production use, we recommend using our OpenAI compatible server.
+We are also not going to accept PRs modifying this file, please
+change `vllm/entrypoints/openai/api_server.py` instead.
+"""
+
+import argparse
+import json
+import ssl
+from typing import AsyncGenerator
+
+import uvicorn
+from fastapi import FastAPI, Request
+from fastapi.responses import JSONResponse, Response, StreamingResponse
+
+from vllm.engine.arg_utils import AsyncEngineArgs
+from vllm.engine.async_llm_engine import AsyncLLMEngine
+from vllm.sampling_params import SamplingParams
+from vllm.usage.usage_lib import UsageContext
+from vllm.utils import random_uuid
+
+TIMEOUT_KEEP_ALIVE = 5  # seconds.
+app = FastAPI()
+engine = None
+
+
+@app.get("/health")
+async def health() -> Response:
+    """Health check."""
+    return Response(status_code=200)
+
+
+@app.post("/generate")
+async def generate(request: Request) -> Response:
+    """Generate completion for the request.
+
+    The request should be a JSON object with the following fields:
+    - prompt: the prompt to use for the generation.
+    - stream: whether to stream the results or not.
+    - other fields: the sampling parameters (See `SamplingParams` for details).
+    """
+    request_dict = await request.json()
+    prompt = request_dict.pop("prompt")
+    stream = request_dict.pop("stream", False)
+    sampling_params = SamplingParams(**request_dict)
+    request_id = random_uuid()
+
+    assert engine is not None
+    results_generator = engine.generate(prompt, sampling_params, request_id)
+
+    # Streaming case
+    async def stream_results() -> AsyncGenerator[bytes, None]:
+        async for request_output in results_generator:
+            prompt = request_output.prompt
+            text_outputs = [
+                prompt + output.text for output in request_output.outputs
+            ]
+            ret = {"text": text_outputs}
+            yield (json.dumps(ret) + "\0").encode("utf-8")
+
+    if stream:
+        return StreamingResponse(stream_results())
+
+    # Non-streaming case
+    final_output = None
+    async for request_output in results_generator:
+        if await request.is_disconnected():
+            # Abort the request if the client disconnects.
+            await engine.abort(request_id)
+            return Response(status_code=499)
+        final_output = request_output
+
+    assert final_output is not None
+    prompt = final_output.prompt
+    text_outputs = [prompt + output.text for output in final_output.outputs]
+    ret = {"text": text_outputs}
+    return JSONResponse(ret)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--host", type=str, default=None)
+    parser.add_argument("--port", type=int, default=8000)
+    parser.add_argument("--ssl-keyfile", type=str, default=None)
+    parser.add_argument("--ssl-certfile", type=str, default=None)
+    parser.add_argument("--ssl-ca-certs",
+                        type=str,
+                        default=None,
+                        help="The CA certificates file")
+    parser.add_argument(
+        "--ssl-cert-reqs",
+        type=int,
+        default=int(ssl.CERT_NONE),
+        help="Whether client certificate is required (see stdlib ssl module's)"
+    )
+    parser.add_argument(
+        "--root-path",
+        type=str,
+        default=None,
+        help="FastAPI root_path when app is behind a path based routing proxy")
+    parser.add_argument("--log-level", type=str, default="debug")
+    parser = AsyncEngineArgs.add_cli_args(parser)
+    args = parser.parse_args()
+    engine_args = AsyncEngineArgs.from_cli_args(args)
+    engine = AsyncLLMEngine.from_engine_args(
+        engine_args, usage_context=UsageContext.API_SERVER)
+
+    app.root_path = args.root_path
+    uvicorn.run(app,
+                host=args.host,
+                port=args.port,
+                log_level=args.log_level,
+                timeout_keep_alive=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)

vllm/entrypoints/llm.py

@@ -0,0 +1,259 @@
+from typing import List, Optional, Union
+
+import torch
+from tqdm import tqdm
+from transformers import PreTrainedTokenizer, PreTrainedTokenizerFast
+
+from vllm.engine.arg_utils import EngineArgs
+from vllm.engine.llm_engine import LLMEngine
+from vllm.lora.request import LoRARequest
+from vllm.outputs import RequestOutput
+from vllm.sampling_params import SamplingParams
+from vllm.sequence import MultiModalData
+from vllm.usage.usage_lib import UsageContext
+from vllm.utils import Counter
+
+
+class LLM:
+    """An LLM for generating texts from given prompts and sampling parameters.
+
+    This class includes a tokenizer, a language model (possibly distributed
+    across multiple GPUs), and GPU memory space allocated for intermediate
+    states (aka KV cache). Given a batch of prompts and sampling parameters,
+    this class generates texts from the model, using an intelligent batching
+    mechanism and efficient memory management.
+
+    NOTE: This class is intended to be used for offline inference. For online
+    serving, use the `AsyncLLMEngine` class instead.
+    NOTE: For the comprehensive list of arguments, see `EngineArgs`.
+
+    Args:
+        model: The name or path of a HuggingFace Transformers model.
+        tokenizer: The name or path of a HuggingFace Transformers tokenizer.
+        tokenizer_mode: The tokenizer mode. "auto" will use the fast tokenizer
+            if available, and "slow" will always use the slow tokenizer.
+        skip_tokenizer_init: If true, skip initialization of tokenizer and
+            detokenizer. Expect valid prompt_token_ids and None for prompt
+            from the input.
+        trust_remote_code: Trust remote code (e.g., from HuggingFace) when
+            downloading the model and tokenizer.
+        tensor_parallel_size: The number of GPUs to use for distributed
+            execution with tensor parallelism.
+        dtype: The data type for the model weights and activations. Currently,
+            we support `float32`, `float16`, and `bfloat16`. If `auto`, we use
+            the `torch_dtype` attribute specified in the model config file.
+            However, if the `torch_dtype` in the config is `float32`, we will
+            use `float16` instead.
+        quantization: The method used to quantize the model weights. Currently,
+            we support "awq", "gptq", "squeezellm", and "fp8" (experimental).
+            If None, we first check the `quantization_config` attribute in the
+            model config file. If that is None, we assume the model weights are
+            not quantized and use `dtype` to determine the data type of
+            the weights.
+        revision: The specific model version to use. It can be a branch name,
+            a tag name, or a commit id.
+        tokenizer_revision: The specific tokenizer version to use. It can be a
+            branch name, a tag name, or a commit id.
+        seed: The seed to initialize the random number generator for sampling.
+        gpu_memory_utilization: The ratio (between 0 and 1) of GPU memory to
+            reserve for the model weights, activations, and KV cache. Higher
+            values will increase the KV cache size and thus improve the model's
+            throughput. However, if the value is too high, it may cause out-of-
+            memory (OOM) errors.
+        swap_space: The size (GiB) of CPU memory per GPU to use as swap space.
+            This can be used for temporarily storing the states of the requests
+            when their `best_of` sampling parameters are larger than 1. If all
+            requests will have `best_of=1`, you can safely set this to 0.
+            Otherwise, too small values may cause out-of-memory (OOM) errors.
+        enforce_eager: Whether to enforce eager execution. If True, we will
+            disable CUDA graph and always execute the model in eager mode.
+            If False, we will use CUDA graph and eager execution in hybrid.
+        max_context_len_to_capture: Maximum context len covered by CUDA graphs.
+            When a sequence has context length larger than this, we fall back
+            to eager mode (DEPRECATED. Use `max_seq_len_to_capture` instead).
+        max_seq_len_to_capture: Maximum sequence len covered by CUDA graphs.
+            When a sequence has context length larger than this, we fall back
+            to eager mode.
+        disable_custom_all_reduce: See ParallelConfig
+    """
+
+    def __init__(
+        self,
+        model: str,
+        tokenizer: Optional[str] = None,
+        tokenizer_mode: str = "auto",
+        skip_tokenizer_init: bool = False,
+        trust_remote_code: bool = False,
+        tensor_parallel_size: int = 1,
+        dtype: str = "auto",
+        quantization: Optional[str] = None,
+        revision: Optional[str] = None,
+        tokenizer_revision: Optional[str] = None,
+        seed: int = 0,
+        gpu_memory_utilization: float = 0.9,
+        swap_space: int = 4,
+        enforce_eager: bool = False,
+        max_context_len_to_capture: Optional[int] = None,
+        max_seq_len_to_capture: int = 8192,
+        disable_custom_all_reduce: bool = False,
+        **kwargs,
+    ) -> None:
+        if "disable_log_stats" not in kwargs:
+            kwargs["disable_log_stats"] = True
+        engine_args = EngineArgs(
+            model=model,
+            tokenizer=tokenizer,
+            tokenizer_mode=tokenizer_mode,
+            skip_tokenizer_init=skip_tokenizer_init,
+            trust_remote_code=trust_remote_code,
+            tensor_parallel_size=tensor_parallel_size,
+            dtype=dtype,
+            quantization=quantization,
+            revision=revision,
+            tokenizer_revision=tokenizer_revision,
+            seed=seed,
+            gpu_memory_utilization=gpu_memory_utilization,
+            swap_space=swap_space,
+            enforce_eager=enforce_eager,
+            max_context_len_to_capture=max_context_len_to_capture,
+            max_seq_len_to_capture=max_seq_len_to_capture,
+            disable_custom_all_reduce=disable_custom_all_reduce,
+            **kwargs,
+        )
+        self.llm_engine = LLMEngine.from_engine_args(
+            engine_args, usage_context=UsageContext.LLM_CLASS)
+        self.request_counter = Counter()
+
+    def get_tokenizer(
+            self) -> Union[PreTrainedTokenizer, PreTrainedTokenizerFast]:
+        return self.llm_engine.tokenizer.tokenizer
+
+    def set_tokenizer(
+        self,
+        tokenizer: Union[PreTrainedTokenizer, PreTrainedTokenizerFast],
+    ) -> None:
+        self.llm_engine.tokenizer.tokenizer = tokenizer
+
+    def generate(
+        self,
+        prompts: Optional[Union[str, List[str]]] = None,
+        sampling_params: Optional[Union[SamplingParams,
+                                        List[SamplingParams]]] = None,
+        prompt_token_ids: Optional[List[List[int]]] = None,
+        use_tqdm: bool = True,
+        lora_request: Optional[LoRARequest] = None,
+        multi_modal_data: Optional[MultiModalData] = None,
+    ) -> List[RequestOutput]:
+        """Generates the completions for the input prompts.
+
+        NOTE: This class automatically batches the given prompts, considering
+        the memory constraint. For the best performance, put all of your prompts
+        into a single list and pass it to this method.
+
+        Args:
+            prompts: A list of prompts to generate completions for.
+            sampling_params: The sampling parameters for text generation. If
+                None, we use the default sampling parameters. 
+                When it is a single value, it is applied to every prompt. 
+                When it is a list, the list must have the same length as the 
+                prompts and it is paired one by one with the prompt.
+            prompt_token_ids: A list of token IDs for the prompts. If None, we
+                use the tokenizer to convert the prompts to token IDs.
+            use_tqdm: Whether to use tqdm to display the progress bar.
+            lora_request: LoRA request to use for generation, if any.
+            multi_modal_data: Multi modal data.
+
+        Returns:
+            A list of `RequestOutput` objects containing the generated
+            completions in the same order as the input prompts.
+        """
+        if prompts is None and prompt_token_ids is None:
+            raise ValueError("Either prompts or prompt_token_ids must be "
+                             "provided.")
+        if self.llm_engine.model_config.skip_tokenizer_init \
+            and prompts is not None:
+            raise ValueError("prompts must be None if skip_tokenizer_init "
+                             "is True")
+        if isinstance(prompts, str):
+            # Convert a single prompt to a list.
+            prompts = [prompts]
+        if (prompts is not None and prompt_token_ids is not None
+                and len(prompts) != len(prompt_token_ids)):
+            raise ValueError("The lengths of prompts and prompt_token_ids "
+                             "must be the same.")
+
+        if prompts is not None:
+            num_requests = len(prompts)
+        else:
+            assert prompt_token_ids is not None
+            num_requests = len(prompt_token_ids)
+
+        if sampling_params is None:
+            # Use default sampling params.
+            sampling_params = SamplingParams()
+
+        elif isinstance(sampling_params,
+                        list) and len(sampling_params) != num_requests:
+            raise ValueError("The lengths of prompts and sampling_params "
+                             "must be the same.")
+        if multi_modal_data:
+            multi_modal_data.data = multi_modal_data.data.to(torch.float16)
+
+        # Add requests to the engine.
+        for i in range(num_requests):
+            prompt = prompts[i] if prompts is not None else None
+            token_ids = None if prompt_token_ids is None else prompt_token_ids[
+                i]
+            self._add_request(
+                prompt,
+                sampling_params[i]
+                if isinstance(sampling_params, list) else sampling_params,
+                token_ids,
+                lora_request=lora_request,
+                # Get ith image while maintaining the batch dim.
+                multi_modal_data=MultiModalData(
+                    type=multi_modal_data.type,
+                    data=multi_modal_data.data[i].unsqueeze(0))
+                if multi_modal_data else None,
+            )
+        return self._run_engine(use_tqdm)
+
+    def _add_request(
+        self,
+        prompt: Optional[str],
+        sampling_params: SamplingParams,
+        prompt_token_ids: Optional[List[int]],
+        lora_request: Optional[LoRARequest] = None,
+        multi_modal_data: Optional[MultiModalData] = None,
+    ) -> None:
+        request_id = str(next(self.request_counter))
+        self.llm_engine.add_request(request_id,
+                                    prompt,
+                                    sampling_params,
+                                    prompt_token_ids,
+                                    lora_request=lora_request,
+                                    multi_modal_data=multi_modal_data)
+
+    def _run_engine(self, use_tqdm: bool) -> List[RequestOutput]:
+        # Initialize tqdm.
+        if use_tqdm:
+            num_requests = self.llm_engine.get_num_unfinished_requests()
+            pbar = tqdm(total=num_requests,
+                        desc="Processed prompts",
+                        dynamic_ncols=True)
+        # Run the engine.
+        outputs: List[RequestOutput] = []
+        while self.llm_engine.has_unfinished_requests():
+            step_outputs = self.llm_engine.step()
+            for output in step_outputs:
+                if output.finished:
+                    outputs.append(output)
+                    if use_tqdm:
+                        pbar.update(1)
+        if use_tqdm:
+            pbar.close()
+        # Sort the outputs by request ID.
+        # This is necessary because some requests may be finished earlier than
+        # its previous requests.
+        outputs = sorted(outputs, key=lambda x: int(x.request_id))
+        return outputs

vllm/entrypoints/openai/api_server.py

@@ -0,0 +1,186 @@
+import asyncio
+import importlib
+import inspect
+import re
+from contextlib import asynccontextmanager
+from http import HTTPStatus
+from typing import Any, Set
+
+import fastapi
+import uvicorn
+from fastapi import Request
+from fastapi.exceptions import RequestValidationError
+from fastapi.middleware.cors import CORSMiddleware
+from fastapi.responses import JSONResponse, Response, StreamingResponse
+from prometheus_client import make_asgi_app
+from starlette.routing import Mount
+
+import vllm
+import vllm.envs as envs
+from vllm.engine.arg_utils import AsyncEngineArgs
+from vllm.engine.async_llm_engine import AsyncLLMEngine
+from vllm.entrypoints.openai.cli_args import make_arg_parser
+from vllm.entrypoints.openai.protocol import (ChatCompletionRequest,
+                                              ChatCompletionResponse,
+                                              CompletionRequest, ErrorResponse)
+from vllm.entrypoints.openai.serving_chat import OpenAIServingChat
+from vllm.entrypoints.openai.serving_completion import OpenAIServingCompletion
+from vllm.logger import init_logger
+from vllm.usage.usage_lib import UsageContext
+
+TIMEOUT_KEEP_ALIVE = 5  # seconds
+
+openai_serving_chat: OpenAIServingChat
+openai_serving_completion: OpenAIServingCompletion
+logger = init_logger(__name__)
+
+_running_tasks: Set[asyncio.Task[Any]] = set()
+
+
+@asynccontextmanager
+async def lifespan(app: fastapi.FastAPI):
+
+    async def _force_log():
+        while True:
+            await asyncio.sleep(10)
+            await engine.do_log_stats()
+
+    if not engine_args.disable_log_stats:
+        task = asyncio.create_task(_force_log())
+        _running_tasks.add(task)
+        task.add_done_callback(_running_tasks.remove)
+
+    yield
+
+
+app = fastapi.FastAPI(lifespan=lifespan)
+
+
+def parse_args():
+    parser = make_arg_parser()
+    return parser.parse_args()
+
+
+# Add prometheus asgi middleware to route /metrics requests
+route = Mount("/metrics", make_asgi_app())
+# Workaround for 307 Redirect for /metrics
+route.path_regex = re.compile('^/metrics(?P<path>.*)$')
+app.routes.append(route)
+
+
+@app.exception_handler(RequestValidationError)
+async def validation_exception_handler(_, exc):
+    err = openai_serving_chat.create_error_response(message=str(exc))
+    return JSONResponse(err.model_dump(), status_code=HTTPStatus.BAD_REQUEST)
+
+
+@app.get("/health")
+async def health() -> Response:
+    """Health check."""
+    await openai_serving_chat.engine.check_health()
+    return Response(status_code=200)
+
+
+@app.get("/v1/models")
+async def show_available_models():
+    models = await openai_serving_chat.show_available_models()
+    return JSONResponse(content=models.model_dump())
+
+
+@app.get("/version")
+async def show_version():
+    ver = {"version": vllm.__version__}
+    return JSONResponse(content=ver)
+
+
+@app.post("/v1/chat/completions")
+async def create_chat_completion(request: ChatCompletionRequest,
+                                 raw_request: Request):
+    generator = await openai_serving_chat.create_chat_completion(
+        request, raw_request)
+    if isinstance(generator, ErrorResponse):
+        return JSONResponse(content=generator.model_dump(),
+                            status_code=generator.code)
+    if request.stream:
+        return StreamingResponse(content=generator,
+                                 media_type="text/event-stream")
+    else:
+        assert isinstance(generator, ChatCompletionResponse)
+        return JSONResponse(content=generator.model_dump())
+
+
+@app.post("/v1/completions")
+async def create_completion(request: CompletionRequest, raw_request: Request):
+    generator = await openai_serving_completion.create_completion(
+        request, raw_request)
+    if isinstance(generator, ErrorResponse):
+        return JSONResponse(content=generator.model_dump(),
+                            status_code=generator.code)
+    if request.stream:
+        return StreamingResponse(content=generator,
+                                 media_type="text/event-stream")
+    else:
+        return JSONResponse(content=generator.model_dump())
+
+
+if __name__ == "__main__":
+    args = 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,
+    )
+
+    if token := envs.VLLM_API_KEY or args.api_key:
+
+        @app.middleware("http")
+        async def authentication(request: Request, call_next):
+            root_path = "" if args.root_path is None else args.root_path
+            if not request.url.path.startswith(f"{root_path}/v1"):
+                return await call_next(request)
+            if request.headers.get("Authorization") != "Bearer " + token:
+                return JSONResponse(content={"error": "Unauthorized"},
+                                    status_code=401)
+            return await call_next(request)
+
+    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)
+        elif inspect.iscoroutinefunction(imported):
+            app.middleware("http")(imported)
+        else:
+            raise ValueError(f"Invalid middleware {middleware}. "
+                             f"Must be a function or a class.")
+
+    logger.info("vLLM API server version %s", vllm.__version__)
+    logger.info("args: %s", args)
+
+    if args.served_model_name is not None:
+        served_model_names = args.served_model_name
+    else:
+        served_model_names = [args.model]
+    engine_args = AsyncEngineArgs.from_cli_args(args)
+    engine = AsyncLLMEngine.from_engine_args(
+        engine_args, usage_context=UsageContext.OPENAI_API_SERVER)
+    openai_serving_chat = OpenAIServingChat(engine, served_model_names,
+                                            args.response_role,
+                                            args.lora_modules,
+                                            args.chat_template)
+    openai_serving_completion = OpenAIServingCompletion(
+        engine, served_model_names, args.lora_modules)
+
+    app.root_path = args.root_path
+    uvicorn.run(app,
+                host=args.host,
+                port=args.port,
+                log_level=args.uvicorn_log_level,
+                timeout_keep_alive=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)

vllm/entrypoints/openai/cli_args.py

@@ -0,0 +1,115 @@
+"""
+This file contains the command line arguments for the vLLM's
+OpenAI-compatible server. It is kept in a separate file for documentation
+purposes.
+"""
+
+import argparse
+import json
+import ssl
+
+from vllm.engine.arg_utils import AsyncEngineArgs, nullable_str
+from vllm.entrypoints.openai.serving_engine import LoRAModulePath
+
+
+class LoRAParserAction(argparse.Action):
+
+    def __call__(self, parser, namespace, values, option_string=None):
+        lora_list = []
+        for item in values:
+            name, path = item.split('=')
+            lora_list.append(LoRAModulePath(name, path))
+        setattr(namespace, self.dest, lora_list)
+
+
+def make_arg_parser():
+    parser = argparse.ArgumentParser(
+        description="vLLM OpenAI-Compatible RESTful API server.")
+    parser.add_argument("--host",
+                        type=nullable_str,
+                        default=None,
+                        help="host name")
+    parser.add_argument("--port", type=int, default=8000, help="port number")
+    parser.add_argument(
+        "--uvicorn-log-level",
+        type=str,
+        default="info",
+        choices=['debug', 'info', 'warning', 'error', 'critical', 'trace'],
+        help="log level for uvicorn")
+    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")
+    parser.add_argument("--api-key",
+                        type=nullable_str,
+                        default=None,
+                        help="If provided, the server will require this key "
+                        "to be presented in the header.")
+    parser.add_argument(
+        "--lora-modules",
+        type=nullable_str,
+        default=None,
+        nargs='+',
+        action=LoRAParserAction,
+        help="LoRA module configurations in the format name=path. "
+        "Multiple modules can be specified.")
+    parser.add_argument("--chat-template",
+                        type=nullable_str,
+                        default=None,
+                        help="The file path to the chat template, "
+                        "or the template in single-line form "
+                        "for the specified model")
+    parser.add_argument("--response-role",
+                        type=nullable_str,
+                        default="assistant",
+                        help="The role name to return if "
+                        "`request.add_generation_prompt=true`.")
+    parser.add_argument("--ssl-keyfile",
+                        type=nullable_str,
+                        default=None,
+                        help="The file path to the SSL key file")
+    parser.add_argument("--ssl-certfile",
+                        type=nullable_str,
+                        default=None,
+                        help="The file path to the SSL cert file")
+    parser.add_argument("--ssl-ca-certs",
+                        type=nullable_str,
+                        default=None,
+                        help="The CA certificates file")
+    parser.add_argument(
+        "--ssl-cert-reqs",
+        type=int,
+        default=int(ssl.CERT_NONE),
+        help="Whether client certificate is required (see stdlib ssl module's)"
+    )
+    parser.add_argument(
+        "--root-path",
+        type=nullable_str,
+        default=None,
+        help="FastAPI root_path when app is behind a path based routing proxy")
+    parser.add_argument(
+        "--middleware",
+        type=nullable_str,
+        action="append",
+        default=[],
+        help="Additional ASGI middleware to apply to the app. "
+        "We accept multiple --middleware arguments. "
+        "The value should be an import path. "
+        "If a function is provided, vLLM will add it to the server "
+        "using @app.middleware('http'). "
+        "If a class is provided, vLLM will add it to the server "
+        "using app.add_middleware(). ")
+
+    parser = AsyncEngineArgs.add_cli_args(parser)
+    return parser

vllm/entrypoints/openai/protocol.py

@@ -0,0 +1,460 @@
+# Adapted from
+# https://github.com/lm-sys/FastChat/blob/168ccc29d3f7edc50823016105c024fe2282732a/fastchat/protocol/openai_api_protocol.py
+import time
+from typing import Dict, List, Literal, Optional, Union
+
+import torch
+from openai.types.chat import ChatCompletionMessageParam
+from pydantic import BaseModel, ConfigDict, Field, model_validator
+from typing_extensions import Annotated
+
+from vllm.sampling_params import SamplingParams
+from vllm.utils import random_uuid
+
+
+class OpenAIBaseModel(BaseModel):
+    # OpenAI API does not allow extra fields
+    model_config = ConfigDict(extra="forbid")
+
+
+class ErrorResponse(OpenAIBaseModel):
+    object: str = "error"
+    message: str
+    type: str
+    param: Optional[str] = None
+    code: int
+
+
+class ModelPermission(OpenAIBaseModel):
+    id: str = Field(default_factory=lambda: f"modelperm-{random_uuid()}")
+    object: str = "model_permission"
+    created: int = Field(default_factory=lambda: int(time.time()))
+    allow_create_engine: bool = False
+    allow_sampling: bool = True
+    allow_logprobs: bool = True
+    allow_search_indices: bool = False
+    allow_view: bool = True
+    allow_fine_tuning: bool = False
+    organization: str = "*"
+    group: Optional[str] = None
+    is_blocking: bool = False
+
+
+class ModelCard(OpenAIBaseModel):
+    id: str
+    object: str = "model"
+    created: int = Field(default_factory=lambda: int(time.time()))
+    owned_by: str = "vllm"
+    root: Optional[str] = None
+    parent: Optional[str] = None
+    permission: List[ModelPermission] = Field(default_factory=list)
+
+
+class ModelList(OpenAIBaseModel):
+    object: str = "list"
+    data: List[ModelCard] = Field(default_factory=list)
+
+
+class UsageInfo(OpenAIBaseModel):
+    prompt_tokens: int = 0
+    total_tokens: int = 0
+    completion_tokens: Optional[int] = 0
+
+
+class ResponseFormat(OpenAIBaseModel):
+    # type must be "json_object" or "text"
+    type: Literal["text", "json_object"]
+
+
+class ChatCompletionRequest(OpenAIBaseModel):
+    # Ordered by official OpenAI API documentation
+    # https://platform.openai.com/docs/api-reference/chat/create
+    messages: List[ChatCompletionMessageParam]
+    model: str
+    frequency_penalty: Optional[float] = 0.0
+    logit_bias: Optional[Dict[str, float]] = None
+    logprobs: Optional[bool] = False
+    top_logprobs: Optional[int] = None
+    max_tokens: Optional[int] = None
+    n: Optional[int] = 1
+    presence_penalty: Optional[float] = 0.0
+    response_format: Optional[ResponseFormat] = None
+    seed: Optional[int] = Field(None,
+                                ge=torch.iinfo(torch.long).min,
+                                le=torch.iinfo(torch.long).max)
+    stop: Optional[Union[str, List[str]]] = Field(default_factory=list)
+    stream: Optional[bool] = False
+    temperature: Optional[float] = 0.7
+    top_p: Optional[float] = 1.0
+    user: Optional[str] = None
+
+    # doc: begin-chat-completion-sampling-params
+    best_of: Optional[int] = None
+    use_beam_search: Optional[bool] = False
+    top_k: Optional[int] = -1
+    min_p: Optional[float] = 0.0
+    repetition_penalty: Optional[float] = 1.0
+    length_penalty: Optional[float] = 1.0
+    early_stopping: Optional[bool] = False
+    ignore_eos: Optional[bool] = False
+    min_tokens: Optional[int] = 0
+    stop_token_ids: Optional[List[int]] = Field(default_factory=list)
+    skip_special_tokens: Optional[bool] = True
+    spaces_between_special_tokens: Optional[bool] = True
+    # doc: end-chat-completion-sampling-params
+
+    # doc: begin-chat-completion-extra-params
+    echo: Optional[bool] = Field(
+        default=False,
+        description=(
+            "If true, the new message will be prepended with the last message "
+            "if they belong to the same role."),
+    )
+    add_generation_prompt: Optional[bool] = Field(
+        default=True,
+        description=
+        ("If true, the generation prompt will be added to the chat template. "
+         "This is a parameter used by chat template in tokenizer config of the "
+         "model."),
+    )
+    include_stop_str_in_output: Optional[bool] = Field(
+        default=False,
+        description=(
+            "Whether to include the stop string in the output. "
+            "This is only applied when the stop or stop_token_ids is set."),
+    )
+    guided_json: Optional[Union[str, dict, BaseModel]] = Field(
+        default=None,
+        description=("If specified, the output will follow the JSON schema."),
+    )
+    guided_regex: Optional[str] = Field(
+        default=None,
+        description=(
+            "If specified, the output will follow the regex pattern."),
+    )
+    guided_choice: Optional[List[str]] = Field(
+        default=None,
+        description=(
+            "If specified, the output will be exactly one of the choices."),
+    )
+    guided_grammar: Optional[str] = Field(
+        default=None,
+        description=(
+            "If specified, the output will follow the context free grammar."),
+    )
+    guided_decoding_backend: Optional[str] = Field(
+        default=None,
+        description=(
+            "If specified, will override the default guided decoding backend "
+            "of the server for this specific request. If set, must be either "
+            "'outlines' / 'lm-format-enforcer'"))
+    guided_whitespace_pattern: Optional[str] = Field(
+        default=None,
+        description=(
+            "If specified, will override the default whitespace pattern "
+            "for guided json decoding."))
+
+    # doc: end-chat-completion-extra-params
+
+    def to_sampling_params(self) -> SamplingParams:
+        if self.logprobs and not self.top_logprobs:
+            raise ValueError("Top logprobs must be set when logprobs is.")
+
+        logits_processors = None
+        if self.logit_bias:
+
+            def logit_bias_logits_processor(
+                    token_ids: List[int],
+                    logits: torch.Tensor) -> torch.Tensor:
+                assert self.logit_bias is not None
+                for token_id, bias in self.logit_bias.items():
+                    # Clamp the bias between -100 and 100 per OpenAI API spec
+                    bias = min(100, max(-100, bias))
+                    logits[int(token_id)] += bias
+                return logits
+
+            logits_processors = [logit_bias_logits_processor]
+
+        return SamplingParams(
+            n=self.n,
+            presence_penalty=self.presence_penalty,
+            frequency_penalty=self.frequency_penalty,
+            repetition_penalty=self.repetition_penalty,
+            temperature=self.temperature,
+            top_p=self.top_p,
+            min_p=self.min_p,
+            seed=self.seed,
+            stop=self.stop,
+            stop_token_ids=self.stop_token_ids,
+            max_tokens=self.max_tokens,
+            min_tokens=self.min_tokens,
+            logprobs=self.top_logprobs if self.logprobs else None,
+            prompt_logprobs=self.top_logprobs if self.echo else None,
+            best_of=self.best_of,
+            top_k=self.top_k,
+            ignore_eos=self.ignore_eos,
+            use_beam_search=self.use_beam_search,
+            early_stopping=self.early_stopping,
+            skip_special_tokens=self.skip_special_tokens,
+            spaces_between_special_tokens=self.spaces_between_special_tokens,
+            include_stop_str_in_output=self.include_stop_str_in_output,
+            length_penalty=self.length_penalty,
+            logits_processors=logits_processors,
+        )
+
+    @model_validator(mode="before")
+    @classmethod
+    def check_guided_decoding_count(cls, data):
+        guide_count = sum([
+            "guided_json" in data and data["guided_json"] is not None,
+            "guided_regex" in data and data["guided_regex"] is not None,
+            "guided_choice" in data and data["guided_choice"] is not None
+        ])
+        if guide_count > 1:
+            raise ValueError(
+                "You can only use one kind of guided decoding "
+                "('guided_json', 'guided_regex' or 'guided_choice').")
+        return data
+
+
+class CompletionRequest(OpenAIBaseModel):
+    # Ordered by official OpenAI API documentation
+    # https://platform.openai.com/docs/api-reference/completions/create
+    model: str
+    prompt: Union[List[int], List[List[int]], str, List[str]]
+    best_of: Optional[int] = None
+    echo: Optional[bool] = False
+    frequency_penalty: Optional[float] = 0.0
+    logit_bias: Optional[Dict[str, float]] = None
+    logprobs: Optional[int] = None
+    max_tokens: Optional[int] = 16
+    n: int = 1
+    presence_penalty: Optional[float] = 0.0
+    seed: Optional[int] = Field(None,
+                                ge=torch.iinfo(torch.long).min,
+                                le=torch.iinfo(torch.long).max)
+    stop: Optional[Union[str, List[str]]] = Field(default_factory=list)
+    stream: Optional[bool] = False
+    suffix: Optional[str] = None
+    temperature: Optional[float] = 1.0
+    top_p: Optional[float] = 1.0
+    user: Optional[str] = None
+
+    # doc: begin-completion-sampling-params
+    use_beam_search: Optional[bool] = False
+    top_k: Optional[int] = -1
+    min_p: Optional[float] = 0.0
+    repetition_penalty: Optional[float] = 1.0
+    length_penalty: Optional[float] = 1.0
+    early_stopping: Optional[bool] = False
+    stop_token_ids: Optional[List[int]] = Field(default_factory=list)
+    ignore_eos: Optional[bool] = False
+    min_tokens: Optional[int] = 0
+    skip_special_tokens: Optional[bool] = True
+    spaces_between_special_tokens: Optional[bool] = True
+    truncate_prompt_tokens: Optional[Annotated[int, Field(ge=1)]] = None
+    # doc: end-completion-sampling-params
+
+    # doc: begin-completion-extra-params
+    include_stop_str_in_output: Optional[bool] = Field(
+        default=False,
+        description=(
+            "Whether to include the stop string in the output. "
+            "This is only applied when the stop or stop_token_ids is set."),
+    )
+    response_format: Optional[ResponseFormat] = Field(
+        default=None,
+        description=
+        ("Similar to chat completion, this parameter specifies the format of "
+         "output. Only {'type': 'json_object'} or {'type': 'text' } is "
+         "supported."),
+    )
+    guided_json: Optional[Union[str, dict, BaseModel]] = Field(
+        default=None,
+        description=("If specified, the output will follow the JSON schema."),
+    )
+    guided_regex: Optional[str] = Field(
+        default=None,
+        description=(
+            "If specified, the output will follow the regex pattern."),
+    )
+    guided_choice: Optional[List[str]] = Field(
+        default=None,
+        description=(
+            "If specified, the output will be exactly one of the choices."),
+    )
+    guided_grammar: Optional[str] = Field(
+        default=None,
+        description=(
+            "If specified, the output will follow the context free grammar."),
+    )
+    guided_decoding_backend: Optional[str] = Field(
+        default=None,
+        description=(
+            "If specified, will override the default guided decoding backend "
+            "of the server for this specific request. If set, must be one of "
+            "'outlines' / 'lm-format-enforcer'"))
+    guided_whitespace_pattern: Optional[str] = Field(
+        default=None,
+        description=(
+            "If specified, will override the default whitespace pattern "
+            "for guided json decoding."))
+
+    # doc: end-completion-extra-params
+
+    def to_sampling_params(self):
+        echo_without_generation = self.echo and self.max_tokens == 0
+
+        logits_processors = None
+        if self.logit_bias:
+
+            def logit_bias_logits_processor(
+                    token_ids: List[int],
+                    logits: torch.Tensor) -> torch.Tensor:
+                assert self.logit_bias is not None
+                for token_id, bias in self.logit_bias.items():
+                    # Clamp the bias between -100 and 100 per OpenAI API spec
+                    bias = min(100, max(-100, bias))
+                    logits[int(token_id)] += bias
+                return logits
+
+            logits_processors = [logit_bias_logits_processor]
+
+        return SamplingParams(
+            n=self.n,
+            best_of=self.best_of,
+            presence_penalty=self.presence_penalty,
+            frequency_penalty=self.frequency_penalty,
+            repetition_penalty=self.repetition_penalty,
+            temperature=self.temperature,
+            top_p=self.top_p,
+            top_k=self.top_k,
+            min_p=self.min_p,
+            seed=self.seed,
+            stop=self.stop,
+            stop_token_ids=self.stop_token_ids,
+            ignore_eos=self.ignore_eos,
+            max_tokens=self.max_tokens if not echo_without_generation else 1,
+            min_tokens=self.min_tokens,
+            logprobs=self.logprobs,
+            use_beam_search=self.use_beam_search,
+            early_stopping=self.early_stopping,
+            prompt_logprobs=self.logprobs if self.echo else None,
+            skip_special_tokens=self.skip_special_tokens,
+            spaces_between_special_tokens=(self.spaces_between_special_tokens),
+            include_stop_str_in_output=self.include_stop_str_in_output,
+            length_penalty=self.length_penalty,
+            logits_processors=logits_processors,
+            truncate_prompt_tokens=self.truncate_prompt_tokens,
+        )
+
+    @model_validator(mode="before")
+    @classmethod
+    def check_guided_decoding_count(cls, data):
+        guide_count = sum([
+            "guided_json" in data and data["guided_json"] is not None,
+            "guided_regex" in data and data["guided_regex"] is not None,
+            "guided_choice" in data and data["guided_choice"] is not None
+        ])
+        if guide_count > 1:
+            raise ValueError(
+                "You can only use one kind of guided decoding "
+                "('guided_json', 'guided_regex' or 'guided_choice').")
+        return data
+
+
+class LogProbs(OpenAIBaseModel):
+    text_offset: List[int] = Field(default_factory=list)
+    token_logprobs: List[Optional[float]] = Field(default_factory=list)
+    tokens: List[str] = Field(default_factory=list)
+    top_logprobs: Optional[List[Optional[Dict[str, float]]]] = None
+
+
+class CompletionResponseChoice(OpenAIBaseModel):
+    index: int
+    text: str
+    logprobs: Optional[LogProbs] = None
+    finish_reason: Optional[str] = None
+    stop_reason: Optional[Union[int, str]] = Field(
+        default=None,
+        description=(
+            "The stop string or token id that caused the completion "
+            "to stop, None if the completion finished for some other reason "
+            "including encountering the EOS token"),
+    )
+
+
+class CompletionResponse(OpenAIBaseModel):
+    id: str = Field(default_factory=lambda: f"cmpl-{random_uuid()}")
+    object: str = "text_completion"
+    created: int = Field(default_factory=lambda: int(time.time()))
+    model: str
+    choices: List[CompletionResponseChoice]
+    usage: UsageInfo
+
+
+class CompletionResponseStreamChoice(OpenAIBaseModel):
+    index: int
+    text: str
+    logprobs: Optional[LogProbs] = None
+    finish_reason: Optional[str] = None
+    stop_reason: Optional[Union[int, str]] = Field(
+        default=None,
+        description=(
+            "The stop string or token id that caused the completion "
+            "to stop, None if the completion finished for some other reason "
+            "including encountering the EOS token"),
+    )
+
+
+class CompletionStreamResponse(OpenAIBaseModel):
+    id: str = Field(default_factory=lambda: f"cmpl-{random_uuid()}")
+    object: str = "text_completion"
+    created: int = Field(default_factory=lambda: int(time.time()))
+    model: str
+    choices: List[CompletionResponseStreamChoice]
+    usage: Optional[UsageInfo] = Field(default=None)
+
+
+class ChatMessage(OpenAIBaseModel):
+    role: str
+    content: str
+
+
+class ChatCompletionResponseChoice(OpenAIBaseModel):
+    index: int
+    message: ChatMessage
+    logprobs: Optional[LogProbs] = None
+    finish_reason: Optional[str] = None
+    stop_reason: Optional[Union[int, str]] = None
+
+
+class ChatCompletionResponse(OpenAIBaseModel):
+    id: str = Field(default_factory=lambda: f"chatcmpl-{random_uuid()}")
+    object: str = "chat.completion"
+    created: int = Field(default_factory=lambda: int(time.time()))
+    model: str
+    choices: List[ChatCompletionResponseChoice]
+    usage: UsageInfo
+
+
+class DeltaMessage(OpenAIBaseModel):
+    role: Optional[str] = None
+    content: Optional[str] = None
+
+
+class ChatCompletionResponseStreamChoice(OpenAIBaseModel):
+    index: int
+    delta: DeltaMessage
+    logprobs: Optional[LogProbs] = None
+    finish_reason: Optional[str] = None
+    stop_reason: Optional[Union[int, str]] = None
+
+
+class ChatCompletionStreamResponse(OpenAIBaseModel):
+    id: str = Field(default_factory=lambda: f"chatcmpl-{random_uuid()}")
+    object: str = "chat.completion.chunk"
+    created: int = Field(default_factory=lambda: int(time.time()))
+    model: str
+    choices: List[ChatCompletionResponseStreamChoice]
+    usage: Optional[UsageInfo] = Field(default=None)

vllm/entrypoints/openai/serving_chat.py

@@ -0,0 +1,392 @@
+import asyncio
+import codecs
+import time
+from typing import (AsyncGenerator, AsyncIterator, Awaitable, Iterable, List,
+                    Optional, Tuple, TypedDict, Union, final)
+
+from fastapi import Request
+from openai.types.chat import (ChatCompletionContentPartParam,
+                               ChatCompletionRole)
+
+from vllm.engine.async_llm_engine import AsyncLLMEngine
+from vllm.entrypoints.openai.protocol import (
+    ChatCompletionRequest, ChatCompletionResponse,
+    ChatCompletionResponseChoice, ChatCompletionResponseStreamChoice,
+    ChatCompletionStreamResponse, ChatMessage, DeltaMessage, ErrorResponse,
+    UsageInfo)
+from vllm.entrypoints.openai.serving_engine import (LoRAModulePath,
+                                                    OpenAIServing)
+from vllm.logger import init_logger
+from vllm.model_executor.guided_decoding import (
+    get_guided_decoding_logits_processor)
+from vllm.outputs import RequestOutput
+from vllm.utils import random_uuid
+
+logger = init_logger(__name__)
+
+
+@final  # So that it should be compatible with Dict[str, str]
+class ConversationMessage(TypedDict):
+    role: str
+    content: str
+
+
+class OpenAIServingChat(OpenAIServing):
+
+    def __init__(self,
+                 engine: AsyncLLMEngine,
+                 served_model_names: List[str],
+                 response_role: str,
+                 lora_modules: Optional[List[LoRAModulePath]] = None,
+                 chat_template: Optional[str] = None):
+        super().__init__(engine=engine,
+                         served_model_names=served_model_names,
+                         lora_modules=lora_modules,
+                         await_post_init=self._load_chat_template(
+                             chat_template=chat_template))
+
+        self.response_role = response_role
+
+    def _parse_chat_message_content(
+        self,
+        role: ChatCompletionRole,
+        content: Optional[Union[str,
+                                Iterable[ChatCompletionContentPartParam]]],
+    ) -> Tuple[List[ConversationMessage], List[Awaitable[object]]]:
+        if content is None:
+            return [], []
+        if isinstance(content, str):
+            return [ConversationMessage(role=role, content=content)], []
+
+        texts: List[str] = []
+        for _, part in enumerate(content):
+            if part["type"] == "text":
+                text = part["text"]
+
+                texts.append(text)
+            else:
+                raise NotImplementedError(f"Unknown part type: {part['type']}")
+
+        return [ConversationMessage(role=role, content="\n".join(texts))], []
+
+    async def create_chat_completion(
+        self, request: ChatCompletionRequest, raw_request: Request
+    ) -> Union[ErrorResponse, AsyncGenerator[str, None],
+               ChatCompletionResponse]:
+        """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 feature:
+            - function_call (Users should implement this by themselves)
+        """
+        error_check_ret = await self._check_model(request)
+        if error_check_ret is not None:
+            return error_check_ret
+
+        try:
+            conversation: List[ConversationMessage] = []
+
+            for m in request.messages:
+                messages, _ = self._parse_chat_message_content(
+                    m["role"], m["content"])
+
+                conversation.extend(messages)
+
+            prompt = self.tokenizer.apply_chat_template(
+                conversation=conversation,
+                tokenize=False,
+                add_generation_prompt=request.add_generation_prompt,
+            )
+        except Exception as e:
+            logger.error("Error in applying chat template from request: %s", e)
+            return self.create_error_response(str(e))
+
+        request_id = f"cmpl-{random_uuid()}"
+        try:
+            # Tokenize/detokenize depending on prompt format (string/token list)
+            prompt_ids, prompt_text = self._validate_prompt_and_tokenize(
+                request, prompt=prompt)
+            sampling_params = request.to_sampling_params()
+            lora_request = self._maybe_get_lora(request)
+            decoding_config = await self.engine.get_decoding_config()
+            guided_decoding_backend = request.guided_decoding_backend \
+                or decoding_config.guided_decoding_backend
+            guided_decode_logits_processor = (
+                await get_guided_decoding_logits_processor(
+                    guided_decoding_backend, request, await
+                    self.engine.get_tokenizer()))
+            if guided_decode_logits_processor:
+                if sampling_params.logits_processors is None:
+                    sampling_params.logits_processors = []
+                sampling_params.logits_processors.append(
+                    guided_decode_logits_processor)
+        except ValueError as e:
+            return self.create_error_response(str(e))
+
+        result_generator = self.engine.generate(prompt_text, sampling_params,
+                                                request_id, prompt_ids,
+                                                lora_request)
+        # Streaming response
+        if request.stream:
+            return self.chat_completion_stream_generator(
+                request, result_generator, request_id, conversation)
+        else:
+            try:
+                return await self.chat_completion_full_generator(
+                    request, raw_request, result_generator, request_id,
+                    conversation)
+            except ValueError as e:
+                # TODO: Use a vllm-specific Validation Error
+                return self.create_error_response(str(e))
+
+    def get_chat_request_role(self, request: ChatCompletionRequest) -> str:
+        if request.add_generation_prompt:
+            return self.response_role
+        else:
+            return request.messages[-1]["role"]
+
+    async def chat_completion_stream_generator(
+            self, request: ChatCompletionRequest,
+            result_generator: AsyncIterator[RequestOutput], request_id: str,
+            conversation: List[ConversationMessage]
+    ) -> AsyncGenerator[str, None]:
+        model_name = self.served_model_names[0]
+        created_time = int(time.time())
+        chunk_object_type = "chat.completion.chunk"
+        first_iteration = True
+
+        # Send response for each token for each request.n (index)
+        assert request.n is not None
+        previous_texts = [""] * request.n
+        previous_num_tokens = [0] * request.n
+        finish_reason_sent = [False] * request.n
+        try:
+            async for res in result_generator:
+                # We need to do it here, because if there are exceptions in
+                # the result_generator, it needs to be sent as the FIRST
+                # response (by the try...catch).
+                if first_iteration:
+                    # Send first response for each request.n (index) with
+                    # the role
+                    role = self.get_chat_request_role(request)
+                    for i in range(request.n):
+                        choice_data = ChatCompletionResponseStreamChoice(
+                            index=i,
+                            delta=DeltaMessage(role=role),
+                            logprobs=None,
+                            finish_reason=None)
+                        chunk = ChatCompletionStreamResponse(
+                            id=request_id,
+                            object=chunk_object_type,
+                            created=created_time,
+                            choices=[choice_data],
+                            model=model_name)
+                        data = chunk.model_dump_json(exclude_unset=True)
+                        yield f"data: {data}\n\n"
+
+                    # Send response to echo the input portion of the
+                    # last message
+                    if request.echo:
+                        last_msg_content = ""
+                        if conversation and conversation[-1].get(
+                                "content") and conversation[-1].get(
+                                    "role") == role:
+                            last_msg_content = conversation[-1]["content"]
+
+                        if last_msg_content:
+                            for i in range(request.n):
+                                choice_data = (
+                                    ChatCompletionResponseStreamChoice(
+                                        index=i,
+                                        delta=DeltaMessage(
+                                            content=last_msg_content),
+                                        finish_reason=None))
+                                chunk = ChatCompletionStreamResponse(
+                                    id=request_id,
+                                    object=chunk_object_type,
+                                    created=created_time,
+                                    choices=[choice_data],
+                                    logprobs=None,
+                                    model=model_name)
+                                data = chunk.model_dump_json(
+                                    exclude_unset=True)
+                                yield f"data: {data}\n\n"
+                    first_iteration = False
+
+                for output in res.outputs:
+                    i = output.index
+
+                    if finish_reason_sent[i]:
+                        continue
+
+                    delta_token_ids = output.token_ids[previous_num_tokens[i]:]
+                    top_logprobs = output.logprobs[
+                        previous_num_tokens[i]:] if output.logprobs else None
+
+                    if request.logprobs:
+                        logprobs = self._create_logprobs(
+                            token_ids=delta_token_ids,
+                            top_logprobs=top_logprobs,
+                            num_output_top_logprobs=request.logprobs,
+                            initial_text_offset=len(previous_texts[i]),
+                        )
+                    else:
+                        logprobs = None
+
+                    delta_text = output.text[len(previous_texts[i]):]
+                    previous_texts[i] = output.text
+                    previous_num_tokens[i] = len(output.token_ids)
+                    if output.finish_reason is None:
+                        # Send token-by-token response for each request.n
+                        choice_data = ChatCompletionResponseStreamChoice(
+                            index=i,
+                            delta=DeltaMessage(content=delta_text),
+                            logprobs=logprobs,
+                            finish_reason=None)
+                        chunk = ChatCompletionStreamResponse(
+                            id=request_id,
+                            object=chunk_object_type,
+                            created=created_time,
+                            choices=[choice_data],
+                            model=model_name)
+                        data = chunk.model_dump_json(exclude_unset=True)
+                        yield f"data: {data}\n\n"
+                    else:
+                        # Send the finish response for each request.n only once
+                        prompt_tokens = len(res.prompt_token_ids)
+                        final_usage = UsageInfo(
+                            prompt_tokens=prompt_tokens,
+                            completion_tokens=previous_num_tokens[i],
+                            total_tokens=prompt_tokens +
+                            previous_num_tokens[i],
+                        )
+                        choice_data = ChatCompletionResponseStreamChoice(
+                            index=i,
+                            delta=DeltaMessage(content=delta_text),
+                            logprobs=logprobs,
+                            finish_reason=output.finish_reason,
+                            stop_reason=output.stop_reason)
+                        chunk = ChatCompletionStreamResponse(
+                            id=request_id,
+                            object=chunk_object_type,
+                            created=created_time,
+                            choices=[choice_data],
+                            model=model_name)
+                        if final_usage is not None:
+                            chunk.usage = final_usage
+                        data = chunk.model_dump_json(exclude_unset=True,
+                                                     exclude_none=True)
+                        yield f"data: {data}\n\n"
+                        finish_reason_sent[i] = True
+        except ValueError as e:
+            # TODO: Use a vllm-specific Validation Error
+            data = self.create_streaming_error_response(str(e))
+            yield f"data: {data}\n\n"
+        # Send the final done message after all response.n are finished
+        yield "data: [DONE]\n\n"
+
+    async def chat_completion_full_generator(
+        self, request: ChatCompletionRequest, raw_request: Request,
+        result_generator: AsyncIterator[RequestOutput], request_id: str,
+        conversation: List[ConversationMessage]
+    ) -> Union[ErrorResponse, ChatCompletionResponse]:
+
+        model_name = self.served_model_names[0]
+        created_time = int(time.time())
+        final_res: Optional[RequestOutput] = None
+
+        async for res in result_generator:
+            if await raw_request.is_disconnected():
+                # Abort the request if the client disconnects.
+                await self.engine.abort(request_id)
+                return self.create_error_response("Client disconnected")
+            final_res = res
+        assert final_res is not None
+
+        choices = []
+
+        role = self.get_chat_request_role(request)
+        for output in final_res.outputs:
+            token_ids = output.token_ids
+            top_logprobs = output.logprobs
+
+            if request.logprobs:
+                logprobs = self._create_logprobs(
+                    token_ids=token_ids,
+                    top_logprobs=top_logprobs,
+                    num_output_top_logprobs=request.logprobs,
+                )
+            else:
+                logprobs = None
+
+            choice_data = ChatCompletionResponseChoice(
+                index=output.index,
+                message=ChatMessage(role=role, content=output.text),
+                logprobs=logprobs,
+                finish_reason=output.finish_reason,
+                stop_reason=output.stop_reason,
+            )
+            choices.append(choice_data)
+
+        if request.echo:
+            last_msg_content = ""
+            if conversation and conversation[-1].get(
+                    "content") and conversation[-1].get("role") == role:
+                last_msg_content = conversation[-1]["content"]
+
+            for choice in choices:
+                full_message = last_msg_content + choice.message.content
+                choice.message.content = full_message
+
+        num_prompt_tokens = len(final_res.prompt_token_ids)
+        num_generated_tokens = sum(
+            len(output.token_ids) for output in final_res.outputs)
+        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,
+        )
+
+        return response
+
+    async def _load_chat_template(self, chat_template: Optional[str]):
+        while self.tokenizer is None:
+            # Give the parent class time to load the tokenizer
+            await asyncio.sleep(0.1)
+        tokenizer = self.tokenizer
+
+        if chat_template is not None:
+            try:
+                with open(chat_template, "r") as f:
+                    tokenizer.chat_template = f.read()
+            except OSError as e:
+                JINJA_CHARS = "{}\n"
+                if not any(c in chat_template for c in JINJA_CHARS):
+                    msg = (f"The supplied chat template ({chat_template}) "
+                           f"looks like a file path, but it failed to be "
+                           f"opened. Reason: {e}")
+                    raise ValueError(msg) from e
+
+                # If opening a file fails, set chat template to be args to
+                # ensure we decode so our escape are interpreted correctly
+                tokenizer.chat_template = codecs.decode(
+                    chat_template, "unicode_escape")
+
+            logger.info("Using supplied chat template:\n%s",
+                        tokenizer.chat_template)
+        elif tokenizer.chat_template is not None:
+            logger.info("Using default chat template:\n%s",
+                        tokenizer.chat_template)
+        else:
+            logger.warning(
+                "No chat template provided. Chat API will not work.")

vllm/entrypoints/openai/serving_completion.py

@@ -0,0 +1,347 @@
+import time
+from typing import (AsyncGenerator, AsyncIterator, Callable, Dict, List,
+                    Optional, Tuple)
+
+from fastapi import Request
+
+from vllm.engine.async_llm_engine import AsyncLLMEngine
+from vllm.entrypoints.openai.protocol import (CompletionRequest,
+                                              CompletionResponse,
+                                              CompletionResponseChoice,
+                                              CompletionResponseStreamChoice,
+                                              CompletionStreamResponse,
+                                              LogProbs, UsageInfo)
+from vllm.entrypoints.openai.serving_engine import (LoRAModulePath,
+                                                    OpenAIServing)
+from vllm.logger import init_logger
+from vllm.model_executor.guided_decoding import (
+    get_guided_decoding_logits_processor)
+from vllm.outputs import RequestOutput
+from vllm.utils import merge_async_iterators, random_uuid
+
+logger = init_logger(__name__)
+
+TypeTokenIDs = List[int]
+TypeTopLogProbs = List[Optional[Dict[int, float]]]
+TypeCreateLogProbsFn = Callable[
+    [TypeTokenIDs, TypeTopLogProbs, Optional[int], int], LogProbs]
+
+
+def parse_prompt_format(prompt) -> Tuple[bool, list]:
+    # get the prompt, openai supports the following
+    # "a string, array of strings, array of tokens, or array of token arrays."
+    prompt_is_tokens = False
+    prompts = [prompt]  # case 1: a string
+    if isinstance(prompt, list):
+        if len(prompt) == 0:
+            raise ValueError("please provide at least one prompt")
+        elif isinstance(prompt[0], str):
+            prompt_is_tokens = False
+            prompts = prompt  # case 2: array of strings
+        elif isinstance(prompt[0], int):
+            prompt_is_tokens = True
+            prompts = [prompt]  # case 3: array of tokens
+        elif isinstance(prompt[0], list) and isinstance(prompt[0][0], int):
+            prompt_is_tokens = True
+            prompts = prompt  # case 4: array of token arrays
+        else:
+            raise ValueError("prompt must be a string, array of strings, "
+                             "array of tokens, or array of token arrays")
+    return prompt_is_tokens, prompts
+
+
+class OpenAIServingCompletion(OpenAIServing):
+
+    def __init__(self,
+                 engine: AsyncLLMEngine,
+                 served_model_names: List[str],
+                 lora_modules: Optional[List[LoRAModulePath]] = None):
+        super().__init__(engine=engine,
+                         served_model_names=served_model_names,
+                         lora_modules=lora_modules)
+
+    async def create_completion(self, request: CompletionRequest,
+                                raw_request: Request):
+        """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 feature:
+            - suffix (the language models we currently support do not support
+            suffix)
+        """
+        error_check_ret = await self._check_model(request)
+        if error_check_ret is not None:
+            return error_check_ret
+
+        # Return error for unsupported features.
+        if request.suffix is not None:
+            return self.create_error_response(
+                "suffix is not currently supported")
+
+        model_name = self.served_model_names[0]
+        request_id = f"cmpl-{random_uuid()}"
+        created_time = int(time.time())
+
+        # Schedule the request and get the result generator.
+        generators: List[AsyncIterator[RequestOutput]] = []
+        try:
+            sampling_params = request.to_sampling_params()
+            lora_request = self._maybe_get_lora(request)
+            decoding_config = await self.engine.get_decoding_config()
+            guided_decoding_backend = request.guided_decoding_backend \
+                or decoding_config.guided_decoding_backend
+            guided_decode_logit_processor = (
+                await get_guided_decoding_logits_processor(
+                    guided_decoding_backend, request, await
+                    self.engine.get_tokenizer()))
+            if guided_decode_logit_processor is not None:
+                if sampling_params.logits_processors is None:
+                    sampling_params.logits_processors = []
+                sampling_params.logits_processors.append(
+                    guided_decode_logit_processor)
+            prompt_is_tokens, prompts = parse_prompt_format(request.prompt)
+
+            for i, prompt in enumerate(prompts):
+                if prompt_is_tokens:
+                    prompt_formats = self._validate_prompt_and_tokenize(
+                        request,
+                        prompt_ids=prompt,
+                        truncate_prompt_tokens=sampling_params.
+                        truncate_prompt_tokens)
+                else:
+                    prompt_formats = self._validate_prompt_and_tokenize(
+                        request,
+                        prompt=prompt,
+                        truncate_prompt_tokens=sampling_params.
+                        truncate_prompt_tokens)
+                prompt_ids, prompt_text = prompt_formats
+
+                generators.append(
+                    self.engine.generate(prompt_text,
+                                         sampling_params,
+                                         f"{request_id}-{i}",
+                                         prompt_token_ids=prompt_ids,
+                                         lora_request=lora_request))
+        except ValueError as e:
+            # TODO: Use a vllm-specific Validation Error
+            return self.create_error_response(str(e))
+
+        result_generator: AsyncIterator[Tuple[
+            int, RequestOutput]] = merge_async_iterators(*generators)
+
+        # 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.
+        stream = (request.stream
+                  and (request.best_of is None or request.n == request.best_of)
+                  and not request.use_beam_search)
+
+        # Streaming response
+        if stream:
+            return self.completion_stream_generator(request,
+                                                    raw_request,
+                                                    result_generator,
+                                                    request_id,
+                                                    created_time,
+                                                    model_name,
+                                                    num_prompts=len(prompts))
+
+        # Non-streaming response
+        final_res_batch: List[Optional[RequestOutput]] = [None] * len(prompts)
+        try:
+            async for i, res in result_generator:
+                if await raw_request.is_disconnected():
+                    # Abort the request if the client disconnects.
+                    await self.engine.abort(f"{request_id}-{i}")
+                    return self.create_error_response("Client disconnected")
+                final_res_batch[i] = res
+            response = self.request_output_to_completion_response(
+                final_res_batch, request, request_id, created_time, model_name)
+        except ValueError as e:
+            # TODO: Use a vllm-specific Validation Error
+            return self.create_error_response(str(e))
+
+        # When user requests streaming but we don't stream, we still need to
+        # return a streaming response with a single event.
+        if request.stream:
+            response_json = response.model_dump_json()
+
+            async def fake_stream_generator() -> AsyncGenerator[str, None]:
+                yield f"data: {response_json}\n\n"
+                yield "data: [DONE]\n\n"
+
+            return fake_stream_generator()
+
+        return response
+
+    async def completion_stream_generator(
+        self,
+        request: CompletionRequest,
+        raw_request: Request,
+        result_generator: AsyncIterator[Tuple[int, RequestOutput]],
+        request_id: str,
+        created_time: int,
+        model_name: str,
+        num_prompts: int,
+    ) -> AsyncGenerator[str, None]:
+        assert request.n is not None
+        previous_texts = [""] * request.n * num_prompts
+        previous_num_tokens = [0] * request.n * num_prompts
+        has_echoed = [False] * request.n * num_prompts
+
+        try:
+            async for prompt_idx, res in result_generator:
+
+                # Abort the request if the client disconnects.
+                if await raw_request.is_disconnected():
+                    await self.engine.abort(f"{request_id}-{prompt_idx}")
+                    raise StopAsyncIteration()
+
+                for output in res.outputs:
+                    i = output.index + prompt_idx * request.n
+                    # TODO(simon): optimize the performance by avoiding full
+                    # text O(n^2) sending.
+
+                    assert request.max_tokens is not None
+                    if request.echo and request.max_tokens == 0:
+                        # only return the prompt
+                        delta_text = res.prompt
+                        delta_token_ids = res.prompt_token_ids
+                        top_logprobs = res.prompt_logprobs
+                        has_echoed[i] = True
+                    elif (request.echo and request.max_tokens > 0
+                          and not has_echoed[i]):
+                        # echo the prompt and first token
+                        delta_text = res.prompt + output.text
+                        delta_token_ids = (res.prompt_token_ids +
+                                           output.token_ids)
+                        top_logprobs = res.prompt_logprobs + (output.logprobs
+                                                              or [])
+                        has_echoed[i] = True
+                    else:
+                        # return just the delta
+                        delta_text = output.text[len(previous_texts[i]):]
+                        delta_token_ids = output.token_ids[
+                            previous_num_tokens[i]:]
+                        top_logprobs = output.logprobs[previous_num_tokens[
+                            i]:] if output.logprobs else None
+
+                    if request.logprobs is not None:
+                        logprobs = self._create_logprobs(
+                            token_ids=delta_token_ids,
+                            top_logprobs=top_logprobs,
+                            num_output_top_logprobs=request.logprobs,
+                            initial_text_offset=len(previous_texts[i]),
+                        )
+                    else:
+                        logprobs = None
+
+                    previous_texts[i] = output.text
+                    previous_num_tokens[i] = len(output.token_ids)
+                    finish_reason = output.finish_reason
+                    stop_reason = output.stop_reason
+                    if output.finish_reason is not None:  # return final usage
+                        prompt_tokens = len(res.prompt_token_ids)
+                        completion_tokens = len(output.token_ids)
+                        final_usage = UsageInfo(
+                            prompt_tokens=prompt_tokens,
+                            completion_tokens=completion_tokens,
+                            total_tokens=prompt_tokens + completion_tokens,
+                        )
+                    else:
+                        final_usage = None
+                    response_json = CompletionStreamResponse(
+                        id=request_id,
+                        created=created_time,
+                        model=model_name,
+                        choices=[
+                            CompletionResponseStreamChoice(
+                                index=i,
+                                text=delta_text,
+                                logprobs=logprobs,
+                                finish_reason=finish_reason,
+                                stop_reason=stop_reason,
+                            )
+                        ],
+                        usage=final_usage,
+                    ).model_dump_json(exclude_unset=True)
+                    yield f"data: {response_json}\n\n"
+        except ValueError as e:
+            # TODO: Use a vllm-specific Validation Error
+            data = self.create_streaming_error_response(str(e))
+            yield f"data: {data}\n\n"
+        yield "data: [DONE]\n\n"
+
+    def request_output_to_completion_response(
+        self,
+        final_res_batch: List[RequestOutput],
+        request: CompletionRequest,
+        request_id: str,
+        created_time: int,
+        model_name: str,
+    ) -> CompletionResponse:
+        choices: List[CompletionResponseChoice] = []
+        num_prompt_tokens = 0
+        num_generated_tokens = 0
+        for final_res in final_res_batch:
+            assert final_res is not None
+            prompt_token_ids = final_res.prompt_token_ids
+            prompt_logprobs = final_res.prompt_logprobs
+            prompt_text = final_res.prompt
+
+            for output in final_res.outputs:
+                assert request.max_tokens is not None
+                if request.echo and request.max_tokens == 0:
+                    token_ids = prompt_token_ids
+                    top_logprobs = prompt_logprobs
+                    output_text = prompt_text
+                elif request.echo and request.max_tokens > 0:
+                    token_ids = prompt_token_ids + output.token_ids
+                    top_logprobs = (prompt_logprobs + output.logprobs
+                                    if request.logprobs else None)
+                    output_text = prompt_text + output.text
+                else:
+                    token_ids = output.token_ids
+                    top_logprobs = output.logprobs
+                    output_text = output.text
+
+                if request.logprobs is not None:
+                    assert top_logprobs is not None, (
+                        "top_logprobs must be provided when logprobs "
+                        "is requested")
+                    logprobs = self._create_logprobs(
+                        token_ids=token_ids,
+                        top_logprobs=top_logprobs,
+                        num_output_top_logprobs=request.logprobs,
+                    )
+                else:
+                    logprobs = None
+
+                choice_data = CompletionResponseChoice(
+                    index=len(choices),
+                    text=output_text,
+                    logprobs=logprobs,
+                    finish_reason=output.finish_reason,
+                    stop_reason=output.stop_reason,
+                )
+                choices.append(choice_data)
+
+            num_prompt_tokens += len(prompt_token_ids)
+            num_generated_tokens += sum(
+                len(output.token_ids) for output in final_res.outputs)
+
+        usage = UsageInfo(
+            prompt_tokens=num_prompt_tokens,
+            completion_tokens=num_generated_tokens,
+            total_tokens=num_prompt_tokens + num_generated_tokens,
+        )
+
+        return CompletionResponse(
+            id=request_id,
+            created=created_time,
+            model=model_name,
+            choices=choices,
+            usage=usage,
+        )

vllm/entrypoints/openai/serving_engine.py

@@ -0,0 +1,234 @@
+import asyncio
+import json
+from dataclasses import dataclass
+from http import HTTPStatus
+from typing import Any, Awaitable, Dict, List, Optional, Tuple, Union
+
+from pydantic import Field
+from transformers import PreTrainedTokenizer, PreTrainedTokenizerFast
+from typing_extensions import Annotated
+
+from vllm.engine.async_llm_engine import AsyncLLMEngine
+from vllm.entrypoints.openai.protocol import (ChatCompletionRequest,
+                                              CompletionRequest, ErrorResponse,
+                                              LogProbs, ModelCard, ModelList,
+                                              ModelPermission)
+from vllm.logger import init_logger
+from vllm.lora.request import LoRARequest
+from vllm.sequence import Logprob
+from vllm.transformers_utils.tokenizer import get_tokenizer
+
+logger = init_logger(__name__)
+
+
+@dataclass
+class LoRAModulePath:
+    name: str
+    local_path: str
+
+
+class OpenAIServing:
+
+    def __init__(self,
+                 engine: AsyncLLMEngine,
+                 served_model_names: List[str],
+                 lora_modules: Optional[List[LoRAModulePath]],
+                 await_post_init: Optional[Awaitable[Any]] = None):
+        self.engine = engine
+        self.served_model_names = served_model_names
+        if lora_modules is None:
+            self.lora_requests = []
+        else:
+            self.lora_requests = [
+                LoRARequest(
+                    lora_name=lora.name,
+                    lora_int_id=i,
+                    lora_local_path=lora.local_path,
+                ) for i, lora in enumerate(lora_modules, start=1)
+            ]
+
+        self.max_model_len = 0
+        # Lazy initialized
+        self.tokenizer: Union[PreTrainedTokenizer, PreTrainedTokenizerFast]
+
+        try:
+            event_loop = asyncio.get_running_loop()
+        except RuntimeError:
+            event_loop = None
+
+        if event_loop is not None and event_loop.is_running():
+            # If the current is instanced by Ray Serve,
+            # there is already a running event loop
+            event_loop.create_task(self._post_init(await_post_init))
+        else:
+            # When using single vLLM without engine_use_ray
+            asyncio.run(self._post_init(await_post_init))
+
+    async def _post_init(self, await_post_init):
+        engine_model_config = await self.engine.get_model_config()
+        self.max_model_len = engine_model_config.max_model_len
+
+        # A separate tokenizer to map token IDs to strings.
+        self.tokenizer = get_tokenizer(
+            engine_model_config.tokenizer,
+            tokenizer_mode=engine_model_config.tokenizer_mode,
+            tokenizer_revision=engine_model_config.tokenizer_revision,
+            trust_remote_code=engine_model_config.trust_remote_code,
+            truncation_side="left")
+
+        if await_post_init is not None:
+            await await_post_init
+
+    async def show_available_models(self) -> ModelList:
+        """Show available models. Right now we only have one model."""
+        model_cards = [
+            ModelCard(id=served_model_name,
+                      root=self.served_model_names[0],
+                      permission=[ModelPermission()])
+            for served_model_name in self.served_model_names
+        ]
+        lora_cards = [
+            ModelCard(id=lora.lora_name,
+                      root=self.served_model_names[0],
+                      permission=[ModelPermission()])
+            for lora in self.lora_requests
+        ]
+        model_cards.extend(lora_cards)
+        return ModelList(data=model_cards)
+
+    def _create_logprobs(
+        self,
+        token_ids: List[int],
+        top_logprobs: List[Optional[Dict[int, Logprob]]],
+        num_output_top_logprobs: Optional[int] = None,
+        initial_text_offset: int = 0,
+    ) -> LogProbs:
+        """Create OpenAI-style logprobs."""
+        logprobs = LogProbs()
+        last_token_len = 0
+        if num_output_top_logprobs:
+            logprobs.top_logprobs = []
+
+        for i, token_id in enumerate(token_ids):
+            step_top_logprobs = top_logprobs[i]
+            if step_top_logprobs is None:
+                token = self.tokenizer.decode(token_id)
+                logprobs.tokens.append(token)
+                logprobs.token_logprobs.append(None)
+                assert logprobs.top_logprobs is not None
+                logprobs.top_logprobs.append(None)
+            else:
+                token_logprob = step_top_logprobs[token_id].logprob
+                token = step_top_logprobs[token_id].decoded_token
+                logprobs.tokens.append(token)
+                logprobs.token_logprobs.append(token_logprob)
+
+                if num_output_top_logprobs:
+                    assert logprobs.top_logprobs is not None
+                    logprobs.top_logprobs.append({
+                        # Convert float("-inf") to the
+                        # JSON-serializable float that OpenAI uses
+                        p.decoded_token: max(p.logprob, -9999.0)
+                        for i, p in step_top_logprobs.items()
+                    } if step_top_logprobs else None)
+
+            if len(logprobs.text_offset) == 0:
+                logprobs.text_offset.append(initial_text_offset)
+            else:
+                logprobs.text_offset.append(logprobs.text_offset[-1] +
+                                            last_token_len)
+            last_token_len = len(token)
+        return logprobs
+
+    def create_error_response(
+            self,
+            message: str,
+            err_type: str = "BadRequestError",
+            status_code: HTTPStatus = HTTPStatus.BAD_REQUEST) -> ErrorResponse:
+        return ErrorResponse(message=message,
+                             type=err_type,
+                             code=status_code.value)
+
+    def create_streaming_error_response(
+            self,
+            message: str,
+            err_type: str = "BadRequestError",
+            status_code: HTTPStatus = HTTPStatus.BAD_REQUEST) -> str:
+        json_str = json.dumps({
+            "error":
+            self.create_error_response(message=message,
+                                       err_type=err_type,
+                                       status_code=status_code).model_dump()
+        })
+        return json_str
+
+    async def _check_model(
+        self, request: Union[CompletionRequest, ChatCompletionRequest]
+    ) -> Optional[ErrorResponse]:
+        if request.model in self.served_model_names:
+            return None
+        if request.model in [lora.lora_name for lora in self.lora_requests]:
+            return None
+        return self.create_error_response(
+            message=f"The model `{request.model}` does not exist.",
+            err_type="NotFoundError",
+            status_code=HTTPStatus.NOT_FOUND)
+
+    def _maybe_get_lora(
+        self, request: Union[CompletionRequest, ChatCompletionRequest]
+    ) -> Optional[LoRARequest]:
+        if request.model in self.served_model_names:
+            return None
+        for lora in self.lora_requests:
+            if request.model == lora.lora_name:
+                return lora
+        # if _check_model has been called earlier, this will be unreachable
+        raise ValueError(f"The model `{request.model}` does not exist.")
+
+    def _validate_prompt_and_tokenize(
+        self,
+        request: Union[ChatCompletionRequest, CompletionRequest],
+        prompt: Optional[str] = None,
+        prompt_ids: Optional[List[int]] = None,
+        truncate_prompt_tokens: Optional[Annotated[int, Field(ge=1)]] = None
+    ) -> Tuple[List[int], str]:
+        if not (prompt or prompt_ids):
+            raise ValueError("Either prompt or prompt_ids should be provided.")
+        if (prompt and prompt_ids):
+            raise ValueError(
+                "Only one of prompt or prompt_ids should be provided.")
+
+        if prompt_ids is None:
+            tokenizer_kwargs = {} if truncate_prompt_tokens is None else {
+                "truncation": True,
+                "max_length": truncate_prompt_tokens,
+            }
+            input_ids = self.tokenizer(prompt, **tokenizer_kwargs).input_ids
+        elif truncate_prompt_tokens is not None:
+            input_ids = prompt_ids[-truncate_prompt_tokens:]
+        else:
+            input_ids = prompt_ids
+
+        input_text = prompt if prompt is not None else self.tokenizer.decode(
+            prompt_ids)
+        token_num = len(input_ids)
+
+        if request.max_tokens is None:
+            if token_num >= self.max_model_len:
+                raise ValueError(
+                    f"This model's maximum context length is "
+                    f"{self.max_model_len} tokens. However, you requested "
+                    f"{token_num} tokens in the messages, "
+                    f"Please reduce the length of the messages.", )
+            request.max_tokens = self.max_model_len - token_num
+
+        if token_num + request.max_tokens > self.max_model_len:
+            raise ValueError(
+                f"This model's maximum context length is "
+                f"{self.max_model_len} tokens. However, you requested "
+                f"{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 input_ids, input_text