openai_api_server.py

import os
import time
from asyncio.log import logger

import uvicorn
import gc
import json
import torch

from vllm import SamplingParams, AsyncEngineArgs, AsyncLLMEngine
from fastapi import FastAPI, HTTPException, Response
from fastapi.middleware.cors import CORSMiddleware
from contextlib import asynccontextmanager
from typing import List, Literal, Optional, Union
from pydantic import BaseModel, Field
from transformers import AutoTokenizer, LogitsProcessor
from sse_starlette.sse import EventSourceResponse

EventSourceResponse.DEFAULT_PING_INTERVAL = 1000
MODEL_PATH = 'THUDM/glm-4-9b-chat'
MAX_MODEL_LENGTH = 8192


@asynccontextmanager
async def lifespan(app: FastAPI):
    yield
    if torch.cuda.is_available():
        torch.cuda.empty_cache()
        torch.cuda.ipc_collect()


app = FastAPI(lifespan=lifespan)

app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)


class ModelCard(BaseModel):
    id: str
    object: str = "model"
    created: int = Field(default_factory=lambda: int(time.time()))
    owned_by: str = "owner"
    root: Optional[str] = None
    parent: Optional[str] = None
    permission: Optional[list] = None


class ModelList(BaseModel):
    object: str = "list"
    data: List[ModelCard] = []


class FunctionCallResponse(BaseModel):
    name: Optional[str] = None
    arguments: Optional[str] = None


class ChatMessage(BaseModel):
    role: Literal["user", "assistant", "system", "tool"]
    content: str = None
    name: Optional[str] = None
    function_call: Optional[FunctionCallResponse] = None


class DeltaMessage(BaseModel):
    role: Optional[Literal["user", "assistant", "system"]] = None
    content: Optional[str] = None
    function_call: Optional[FunctionCallResponse] = None


class EmbeddingRequest(BaseModel):
    input: Union[List[str], str]
    model: str


class CompletionUsage(BaseModel):
    prompt_tokens: int
    completion_tokens: int
    total_tokens: int


class EmbeddingResponse(BaseModel):
    data: list
    model: str
    object: str
    usage: CompletionUsage


class UsageInfo(BaseModel):
    prompt_tokens: int = 0
    total_tokens: int = 0
    completion_tokens: Optional[int] = 0


class ChatCompletionRequest(BaseModel):
    model: str
    messages: List[ChatMessage]
    temperature: Optional[float] = 0.8
    top_p: Optional[float] = 0.8
    max_tokens: Optional[int] = None
    stream: Optional[bool] = False
    tools: Optional[Union[dict, List[dict]]] = None
    tool_choice: Optional[Union[str, dict]] = "None"
    repetition_penalty: Optional[float] = 1.1


class ChatCompletionResponseChoice(BaseModel):
    index: int
    message: ChatMessage
    finish_reason: Literal["stop", "length", "function_call"]


class ChatCompletionResponseStreamChoice(BaseModel):
    delta: DeltaMessage
    finish_reason: Optional[Literal["stop", "length", "function_call"]]
    index: int


class ChatCompletionResponse(BaseModel):
    model: str
    id: str
    object: Literal["chat.completion", "chat.completion.chunk"]
    choices: List[Union[ChatCompletionResponseChoice, ChatCompletionResponseStreamChoice]]
    created: Optional[int] = Field(default_factory=lambda: int(time.time()))
    usage: Optional[UsageInfo] = None


class InvalidScoreLogitsProcessor(LogitsProcessor):
    def __call__(

            self, input_ids: torch.LongTensor, scores: torch.FloatTensor

    ) -> torch.FloatTensor:
        if torch.isnan(scores).any() or torch.isinf(scores).any():
            scores.zero_()
            scores[..., 5] = 5e4
        return scores


def process_response(output: str, use_tool: bool = False) -> Union[str, dict]:
    content = ""
    for response in output.split("<|assistant|>"):
        if "\n" in response:
            metadata, content = response.split("\n", maxsplit=1)
        else:
            metadata, content = "", response
        if not metadata.strip():
            content = content.strip()
        else:
            if use_tool:
                parameters = eval(content.strip())
                content = {
                    "name": metadata.strip(),
                    "arguments": json.dumps(parameters, ensure_ascii=False)
                }
            else:
                content = {
                    "name": metadata.strip(),
                    "content": content
                }
    return content


@torch.inference_mode()
async def generate_stream_glm4(params):
    messages = params["messages"]
    tools = params["tools"]
    tool_choice = params["tool_choice"]
    temperature = float(params.get("temperature", 1.0))
    repetition_penalty = float(params.get("repetition_penalty", 1.0))
    top_p = float(params.get("top_p", 1.0))
    max_new_tokens = int(params.get("max_tokens", 8192))
    messages = process_messages(messages, tools=tools, tool_choice=tool_choice)
    inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
    params_dict = {
        "n": 1,
        "best_of": 1,
        "presence_penalty": 1.0,
        "frequency_penalty": 0.0,
        "temperature": temperature,
        "top_p": top_p,
        "top_k": -1,
        "repetition_penalty": repetition_penalty,
        "use_beam_search": False,
        "length_penalty": 1,
        "early_stopping": False,
        "stop_token_ids": [151329, 151336, 151338],
        "ignore_eos": False,
        "max_tokens": max_new_tokens,
        "logprobs": None,
        "prompt_logprobs": None,
        "skip_special_tokens": True,
    }
    sampling_params = SamplingParams(**params_dict)
    async for output in engine.generate(inputs=inputs, sampling_params=sampling_params, request_id=f"{time.time()}"):
        output_len = len(output.outputs[0].token_ids)
        input_len = len(output.prompt_token_ids)
        ret = {
            "text": output.outputs[0].text,
            "usage": {
                "prompt_tokens": input_len,
                "completion_tokens": output_len,
                "total_tokens": output_len + input_len
            },
            "finish_reason": output.outputs[0].finish_reason,
        }
        yield ret
    gc.collect()
    torch.cuda.empty_cache()


def process_messages(messages, tools=None, tool_choice="none"):
    _messages = messages
    messages = []
    msg_has_sys = False

    def filter_tools(tool_choice, tools):
        function_name = tool_choice.get('function', {}).get('name', None)
        if not function_name:
            return []
        filtered_tools = [
            tool for tool in tools
            if tool.get('function', {}).get('name') == function_name
        ]
        return filtered_tools

    if tool_choice != "none":
        if isinstance(tool_choice, dict):
            tools = filter_tools(tool_choice, tools)
        if tools:
            messages.append(
                {
                    "role": "system",
                    "content": None,
                    "tools": tools
                }
            )
        msg_has_sys = True

    # add to metadata
    if isinstance(tool_choice, dict) and tools:
        messages.append(
            {
                "role": "assistant",
                "metadata": tool_choice["function"]["name"],
                "content": ""
            }
        )

    for m in _messages:
        role, content, func_call = m.role, m.content, m.function_call
        if role == "function":
            messages.append(
                {
                    "role": "observation",
                    "content": content
                }
            )
        elif role == "assistant" and func_call is not None:
            for response in content.split("<|assistant|>"):
                if "\n" in response:
                    metadata, sub_content = response.split("\n", maxsplit=1)
                else:
                    metadata, sub_content = "", response
                messages.append(
                    {
                        "role": role,
                        "metadata": metadata,
                        "content": sub_content.strip()
                    }
                )
        else:
            if role == "system" and msg_has_sys:
                msg_has_sys = False
                continue
            messages.append({"role": role, "content": content})

    return messages


@app.get("/health")
async def health() -> Response:
    """Health check."""
    return Response(status_code=200)


@app.get("/v1/models", response_model=ModelList)
async def list_models():
    model_card = ModelCard(id="glm-4")
    return ModelList(data=[model_card])


@app.post("/v1/chat/completions", response_model=ChatCompletionResponse)
async def create_chat_completion(request: ChatCompletionRequest):
    if len(request.messages) < 1 or request.messages[-1].role == "assistant":
        raise HTTPException(status_code=400, detail="Invalid request")

    gen_params = dict(
        messages=request.messages,
        temperature=request.temperature,
        top_p=request.top_p,
        max_tokens=request.max_tokens or 1024,
        echo=False,
        stream=request.stream,
        repetition_penalty=request.repetition_penalty,
        tools=request.tools,
        tool_choice=request.tool_choice,
    )
    logger.debug(f"==== request ====\n{gen_params}")

    if request.stream:
        predict_stream_generator = predict_stream(request.model, gen_params)
        output = await anext(predict_stream_generator)
        if output:
            return EventSourceResponse(predict_stream_generator, media_type="text/event-stream")
        logger.debug(f"First result output：\n{output}")

        function_call = None
        if output and request.tools:
            try:
                function_call = process_response(output, use_tool=True)
            except:
                logger.warning("Failed to parse tool call")

        # CallFunction
        if isinstance(function_call, dict):
            function_call = FunctionCallResponse(**function_call)
            tool_response = ""
            if not gen_params.get("messages"):
                gen_params["messages"] = []
            gen_params["messages"].append(ChatMessage(role="assistant", content=output))
            gen_params["messages"].append(ChatMessage(role="tool", name=function_call.name, content=tool_response))
            generate = predict(request.model, gen_params)
            return EventSourceResponse(generate, media_type="text/event-stream")
        else:
            generate = parse_output_text(request.model, output)
            return EventSourceResponse(generate, media_type="text/event-stream")

    response = ""
    async for response in generate_stream_glm4(gen_params):
        pass

    if response["text"].startswith("\n"):
        response["text"] = response["text"][1:]
    response["text"] = response["text"].strip()

    usage = UsageInfo()
    function_call, finish_reason = None, "stop"
    if request.tools:
        try:
            function_call = process_response(response["text"], use_tool=True)
        except:
            logger.warning(
                "Failed to parse tool call, maybe the response is not a function call(such as cogview drawing) or have been answered.")

    if isinstance(function_call, dict):
        finish_reason = "function_call"
        function_call = FunctionCallResponse(**function_call)

    message = ChatMessage(
        role="assistant",
        content=response["text"],
        function_call=function_call if isinstance(function_call, FunctionCallResponse) else None,
    )

    logger.debug(f"==== message ====\n{message}")

    choice_data = ChatCompletionResponseChoice(
        index=0,
        message=message,
        finish_reason=finish_reason,
    )
    task_usage = UsageInfo.model_validate(response["usage"])
    for usage_key, usage_value in task_usage.model_dump().items():
        setattr(usage, usage_key, getattr(usage, usage_key) + usage_value)

    return ChatCompletionResponse(
        model=request.model,
        id="",  # for open_source model, id is empty
        choices=[choice_data],
        object="chat.completion",
        usage=usage
    )


async def predict(model_id: str, params: dict):
    choice_data = ChatCompletionResponseStreamChoice(
        index=0,
        delta=DeltaMessage(role="assistant"),
        finish_reason=None
    )
    chunk = ChatCompletionResponse(model=model_id, id="", choices=[choice_data], object="chat.completion.chunk")
    yield "{}".format(chunk.model_dump_json(exclude_unset=True))

    previous_text = ""
    async for new_response in generate_stream_glm4(params):
        decoded_unicode = new_response["text"]
        delta_text = decoded_unicode[len(previous_text):]
        previous_text = decoded_unicode

        finish_reason = new_response["finish_reason"]
        if len(delta_text) == 0 and finish_reason != "function_call":
            continue

        function_call = None
        if finish_reason == "function_call":
            try:
                function_call = process_response(decoded_unicode, use_tool=True)
            except:
                logger.warning(
                    "Failed to parse tool call, maybe the response is not a tool call or have been answered.")

        if isinstance(function_call, dict):
            function_call = FunctionCallResponse(**function_call)

        delta = DeltaMessage(
            content=delta_text,
            role="assistant",
            function_call=function_call if isinstance(function_call, FunctionCallResponse) else None,
        )

        choice_data = ChatCompletionResponseStreamChoice(
            index=0,
            delta=delta,
            finish_reason=finish_reason
        )
        chunk = ChatCompletionResponse(
            model=model_id,
            id="",
            choices=[choice_data],
            object="chat.completion.chunk"
        )
        yield "{}".format(chunk.model_dump_json(exclude_unset=True))

    choice_data = ChatCompletionResponseStreamChoice(
        index=0,
        delta=DeltaMessage(),
        finish_reason="stop"
    )
    chunk = ChatCompletionResponse(
        model=model_id,
        id="",
        choices=[choice_data],
        object="chat.completion.chunk"
    )
    yield "{}".format(chunk.model_dump_json(exclude_unset=True))
    yield '[DONE]'


async def predict_stream(model_id, gen_params):
    output = ""
    is_function_call = False
    has_send_first_chunk = False
    async  for new_response in generate_stream_glm4(gen_params):
        decoded_unicode = new_response["text"]
        delta_text = decoded_unicode[len(output):]
        output = decoded_unicode

        if not is_function_call and len(output) > 7:
            is_function_call = output and 'get_' in output
            if is_function_call:
                continue

            finish_reason = new_response["finish_reason"]
            if not has_send_first_chunk:
                message = DeltaMessage(
                    content="",
                    role="assistant",
                    function_call=None,
                )
                choice_data = ChatCompletionResponseStreamChoice(
                    index=0,
                    delta=message,
                    finish_reason=finish_reason
                )
                chunk = ChatCompletionResponse(
                    model=model_id,
                    id="",
                    choices=[choice_data],
                    created=int(time.time()),
                    object="chat.completion.chunk"
                )
                yield "{}".format(chunk.model_dump_json(exclude_unset=True))

            send_msg = delta_text if has_send_first_chunk else output
            has_send_first_chunk = True
            message = DeltaMessage(
                content=send_msg,
                role="assistant",
                function_call=None,
            )
            choice_data = ChatCompletionResponseStreamChoice(
                index=0,
                delta=message,
                finish_reason=finish_reason
            )
            chunk = ChatCompletionResponse(
                model=model_id,
                id="",
                choices=[choice_data],
                created=int(time.time()),
                object="chat.completion.chunk"
            )
            yield "{}".format(chunk.model_dump_json(exclude_unset=True))

    if is_function_call:
        yield output
    else:
        yield '[DONE]'


async def parse_output_text(model_id: str, value: str):
    choice_data = ChatCompletionResponseStreamChoice(
        index=0,
        delta=DeltaMessage(role="assistant", content=value),
        finish_reason=None
    )
    chunk = ChatCompletionResponse(model=model_id, id="", choices=[choice_data], object="chat.completion.chunk")
    yield "{}".format(chunk.model_dump_json(exclude_unset=True))
    choice_data = ChatCompletionResponseStreamChoice(
        index=0,
        delta=DeltaMessage(),
        finish_reason="stop"
    )
    chunk = ChatCompletionResponse(model=model_id, id="", choices=[choice_data], object="chat.completion.chunk")
    yield "{}".format(chunk.model_dump_json(exclude_unset=True))
    yield '[DONE]'


if __name__ == "__main__":
    tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH, trust_remote_code=True)
    engine_args = AsyncEngineArgs(
        model=MODEL_PATH,
        tokenizer=MODEL_PATH,
        tensor_parallel_size=1,
        dtype="bfloat16",
        trust_remote_code=True,
        gpu_memory_utilization=0.9,
        enforce_eager=True,
        worker_use_ray=True,
        engine_use_ray=False,
        disable_log_requests=True,
        max_model_len=MAX_MODEL_LENGTH,
    )
    engine = AsyncLLMEngine.from_engine_args(engine_args)
    uvicorn.run(app, host='0.0.0.0', port=8000, workers=1)