app.py

import asyncio
import logging
import torch
import gradio as gr
from fastapi import FastAPI, HTTPException
from fastapi.middleware.cors import CORSMiddleware
from pydantic import BaseModel, Field, root_validator
from typing import List, Dict, Optional
from functools import lru_cache
from threading import Lock
import uvicorn

class EmbeddingRequest(BaseModel):
    # 强制锁定模型参数
    model: str = Field(
        default="jinaai/jina-embeddings-v3",
        description="此参数仅用于API兼容，实际模型固定为jinaai/jina-embeddings-v3",
        frozen=True  # 禁止修改
    )
    # 支持三种输入字段
    inputs: Optional[str] = Field(None, description="输入文本（兼容HuggingFace格式）")
    input: Optional[str] = Field(None, description="输入文本（兼容OpenAI格式）")
    prompt: Optional[str] = Field(None, description="输入文本（兼容Ollama格式）")

    # 自动合并输入字段
    @root_validator(pre=True)
    def merge_input_fields(cls, values):
        input_fields = ["inputs", "input", "prompt"]
        for field in input_fields:
            if values.get(field):
                values["inputs"] = values[field]
                break
        else:
            raise ValueError("必须提供 inputs/input/prompt 任一字段")
        return values

class EmbeddingResponse(BaseModel):
    status: str
    embeddings: List[List[float]]

class EmbeddingService:
    def __init__(self):
        self._true_model_name = "jinaai/jina-embeddings-v3"  # 硬编码模型名称
        self.max_length = 512
        self.device = torch.device("cpu")
        self.model = None
        self.tokenizer = None
        self.lock = Lock()
        self.setup_logging()
        torch.set_num_threads(4)  # CPU优化

    def setup_logging(self):
        logging.basicConfig(
            level=logging.INFO,
            format='%(asctime)s - %(levelname)s - %(message)s'
        )
        self.logger = logging.getLogger(__name__)

    async def initialize(self):
        try:
            from transformers import AutoTokenizer, AutoModel
            self.tokenizer = AutoTokenizer.from_pretrained(
                self._true_model_name,
                trust_remote_code=True
            )
            self.model = AutoModel.from_pretrained(
                self._true_model_name,
                trust_remote_code=True
            ).to(self.device)
            self.model.eval()
            torch.set_grad_enabled(False)
            self.logger.info(f"强制加载模型: {self._true_model_name}")
        except Exception as e:
            self.logger.error(f"模型初始化失败: {str(e)}")
            raise

    @lru_cache(maxsize=1000)
    def get_embedding(self, text: str) -> List[float]:
        with self.lock:
            try:
                inputs = self.tokenizer(
                    text,
                    return_tensors="pt",
                    truncation=True,
                    max_length=self.max_length,
                    padding=True
                )

                with torch.no_grad():
                    outputs = self.model(**inputs).last_hidden_state.mean(dim=1)
                    return outputs.numpy().tolist()[0]
            except Exception as e:
                self.logger.error(f"生成嵌入向量失败: {str(e)}")
                raise

embedding_service = EmbeddingService()
app = FastAPI()

app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)
@app.post("/embed", response_model=EmbeddingResponse)
@app.post("/api/embeddings", response_model=EmbeddingResponse)
@app.post("/api/embed", response_model=EmbeddingResponse)
@app.post("/v1/embeddings", response_model=EmbeddingResponse)
@app.post("/generate_embeddings", response_model=EmbeddingResponse)
@app.post("/api/v1/embeddings", response_model=EmbeddingResponse)
@app.post("/hf/v1/embeddings", response_model=EmbeddingResponse)
@app.post("/api/v1/chat/completions", response_model=EmbeddingResponse)
@app.post("/hf/v1/chat/completions", response_model=EmbeddingResponse)
async def generate_embeddings(request: EmbeddingRequest):
    try:
        embedding = await asyncio.get_running_loop().run_in_executor(
            None,
            embedding_service.get_embedding,
            request.inputs  # 使用合并后的输入字段
        )
        return EmbeddingResponse(
            status="success",
            embeddings=[embedding]
        )
    except Exception as e:
        raise HTTPException(status_code=500, detail=str(e))

@app.get("/")
async def root():
    return {
        "status": "active",
        "true_model": embedding_service._true_model_name,
        "device": str(embedding_service.device)
    }

def gradio_interface(text: str) -> Dict:
    try:
        embedding = embedding_service.get_embedding(text)
        return {
            "status": "success",
            "embeddings": [embedding]
        }
    except Exception as e:
        return {
            "status": "error",
            "message": str(e)
        }

iface = gr.Interface(
    fn=gradio_interface,
    inputs=gr.Textbox(lines=3, label="输入文本"),
    outputs=gr.JSON(label="嵌入向量结果"),
    title="Jina Embeddings V3",
    description="强制使用jinaai/jina-embeddings-v3模型（无视请求中的model参数）",
    examples=[[
        "Represent this sentence for searching relevant passages: "
        "The sky is blue because of Rayleigh scattering"
    ]]
)

@app.on_event("startup")
async def startup_event():
    await embedding_service.initialize()

if __name__ == "__main__":
    asyncio.run(embedding_service.initialize())
    gr.mount_gradio_app(app, iface, path="/ui")
    uvicorn.run(app, host="0.0.0.0", port=7860, workers=1)