app.py

import base64
import ctypes
import json
import logging
import queue
import random
import re
import struct
import threading
import time

import transformers
from curl_cffi import requests
from fastapi import FastAPI, HTTPException, Request
from fastapi.middleware.cors import CORSMiddleware
from fastapi.responses import JSONResponse, StreamingResponse
from fastapi.templating import Jinja2Templates
from wasmtime import Linker, Module, Store

# -------------------------- 初始化 tokenizer --------------------------
chat_tokenizer_dir = "./"
tokenizer = transformers.AutoTokenizer.from_pretrained(
    chat_tokenizer_dir, trust_remote_code=True
)

# -------------------------- 日志配置 --------------------------
logging.basicConfig(
    level=logging.INFO, format="%(asctime)s [%(levelname)s] %(name)s: %(message)s"
)
logger = logging.getLogger("main")

app = FastAPI()

# 添加 CORS 中间件，允许所有来源
app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_credentials=True,
    allow_methods=["GET", "POST", "OPTIONS", "PUT", "DELETE"],
    allow_headers=["Content-Type", "Authorization"],
)

# 模板目录
templates = Jinja2Templates(directory="templates")

# ----------------------------------------------------------------------
# (1) 配置文件的读写函数
# ----------------------------------------------------------------------
CONFIG_PATH = "config.json"


def load_config():
    """从 config.json 加载配置，出错则返回空 dict"""
    try:
        with open(CONFIG_PATH, "r", encoding="utf-8") as f:
            return json.load(f)
    except Exception as e:
        logger.warning(f"[load_config] 无法读取配置文件: {e}")
        return {}


def save_config(cfg):
    """将配置写回 config.json"""
    try:
        with open(CONFIG_PATH, "w", encoding="utf-8") as f:
            json.dump(cfg, f, ensure_ascii=False, indent=2)
    except Exception as e:
        logger.error(f"[save_config] 写入 config.json 失败: {e}")


CONFIG = load_config()

# -------------------------- 全局账号队列 --------------------------
account_queue = []  # 维护所有可用账号


def init_account_queue():
    """初始化时从配置加载账号"""
    global account_queue
    account_queue = CONFIG.get("accounts", [])[:]  # 深拷贝
    random.shuffle(account_queue)  # 初始随机排序


init_account_queue()

# ----------------------------------------------------------------------
# (2) DeepSeek 相关常量
# ----------------------------------------------------------------------
DEEPSEEK_HOST = "chat.deepseek.com"
DEEPSEEK_LOGIN_URL = f"https://{DEEPSEEK_HOST}/api/v0/users/login"
DEEPSEEK_CREATE_SESSION_URL = f"https://{DEEPSEEK_HOST}/api/v0/chat_session/create"
DEEPSEEK_CREATE_POW_URL = f"https://{DEEPSEEK_HOST}/api/v0/chat/create_pow_challenge"
DEEPSEEK_COMPLETION_URL = f"https://{DEEPSEEK_HOST}/api/v0/chat/completion"
BASE_HEADERS = {
    "Host": "chat.deepseek.com",
    "User-Agent": "DeepSeek/1.0.13 Android/35",
    "Accept": "application/json",
    "Accept-Encoding": "gzip",
    "Content-Type": "application/json",
    "x-client-platform": "android",
    "x-client-version": "1.0.13",
    "x-client-locale": "zh_CN",
    "accept-charset": "UTF-8",
}

# WASM 模块文件路径
WASM_PATH = "sha3_wasm_bg.7b9ca65ddd.wasm"


# ----------------------------------------------------------------------
# 辅助函数：获取账号唯一标识（优先 email，否则 mobile）
# ----------------------------------------------------------------------
def get_account_identifier(account):
    """返回账号的唯一标识，优先使用 email，否则使用 mobile"""
    return account.get("email", "").strip() or account.get("mobile", "").strip()


# ----------------------------------------------------------------------
# (3) 登录函数：支持使用 email 或 mobile 登录
# ----------------------------------------------------------------------
def login_deepseek_via_account(account):
    """使用 account 中的 email 或 mobile 登录 DeepSeek，
    成功后将返回的 token 写入 account 并保存至配置文件，返回新 token。
    """
    email = account.get("email", "").strip()
    mobile = account.get("mobile", "").strip()
    password = account.get("password", "").strip()
    if not password or (not email and not mobile):
        raise HTTPException(
            status_code=400,
            detail="账号缺少必要的登录信息（必须提供 email 或 mobile 以及 password）",
        )
    if email:
        payload = {
            "email": email,
            "password": password,
            "device_id": "deepseek_to_api",
            "os": "android",
        }
    else:
        payload = {
            "mobile": mobile,
            "area_code": None,
            "password": password,
            "device_id": "deepseek_to_api",
            "os": "android",
        }
    try:
        resp = requests.post(DEEPSEEK_LOGIN_URL, headers=BASE_HEADERS, json=payload)
        resp.raise_for_status()
    except Exception as e:
        logger.error(f"[login_deepseek_via_account] 登录请求异常: {e}")
        raise HTTPException(status_code=500, detail="Account login failed: 请求异常")
    try:
        data = resp.json()
    except Exception as e:
        logger.error(f"[login_deepseek_via_account] JSON解析失败: {e}")
        raise HTTPException(
            status_code=500, detail="Account login failed: invalid JSON response"
        )
    # 校验响应数据格式是否正确
    if (
        data.get("data") is None
        or data["data"].get("biz_data") is None
        or data["data"]["biz_data"].get("user") is None
    ):
        logger.error(f"[login_deepseek_via_account] 登录响应格式错误: {data}")
        raise HTTPException(
            status_code=500, detail="Account login failed: invalid response format"
        )
    new_token = data["data"]["biz_data"]["user"].get("token")
    if not new_token:
        logger.error(f"[login_deepseek_via_account] 登录响应中缺少 token: {data}")
        raise HTTPException(
            status_code=500, detail="Account login failed: missing token"
        )
    account["token"] = new_token
    save_config(CONFIG)
    return new_token


# ----------------------------------------------------------------------
# (4) 从 accounts 中随机选择一个未忙且未尝试过的账号
# ----------------------------------------------------------------------
def choose_new_account():
    """选择策略：
    1. 遍历队列，找到第一个未被 exclude_ids 包含的账号
    2. 从队列中移除该账号
    3. 返回该账号（由后续逻辑保证最终会重新入队）
    """
    for i in range(len(account_queue)):
        acc = account_queue[i]
        acc_id = get_account_identifier(acc)
        if acc_id:
            # 从队列中移除并返回
            # logger.info(f"[choose_new_account] 新选择账号: {acc_id}")
            return account_queue.pop(i)

    logger.warning("[choose_new_account] 没有可用的账号或所有账号都在使用中")
    return None


def release_account(account):
    """将账号重新加入队列末尾"""
    account_queue.append(account)


# ----------------------------------------------------------------------
# (5) 判断调用模式：配置模式 vs 用户自带 token
# ----------------------------------------------------------------------
def determine_mode_and_token(request: Request):
    """
    根据请求头 Authorization 判断使用哪种模式：
    - 如果 Bearer token 出现在 CONFIG["keys"] 中，则为配置模式，从 CONFIG["accounts"] 中随机选择一个账号（排除已尝试账号），
      检查该账号是否已有 token，否则调用登录接口获取；
    - 否则，直接使用请求中的 Bearer 值作为 DeepSeek token。
    结果存入 request.state.deepseek_token；配置模式下同时存入 request.state.account 与 request.state.tried_accounts。
    """
    auth_header = request.headers.get("Authorization", "")
    if not auth_header.startswith("Bearer "):
        raise HTTPException(
            status_code=401, detail="Unauthorized: missing Bearer token."
        )
    caller_key = auth_header.replace("Bearer ", "", 1).strip()
    config_keys = CONFIG.get("keys", [])
    if caller_key in config_keys:
        request.state.use_config_token = True
        request.state.tried_accounts = []  # 初始化已尝试账号
        selected_account = choose_new_account()
        if not selected_account:
            raise HTTPException(
                status_code=429,
                detail="No accounts configured or all accounts are busy.",
            )
        if not selected_account.get("token", "").strip():
            try:
                login_deepseek_via_account(selected_account)
            except Exception as e:
                logger.error(
                    f"[determine_mode_and_token] 账号 {get_account_identifier(selected_account)} 登录失败：{e}"
                )
                raise HTTPException(status_code=500, detail="Account login failed.")

        request.state.deepseek_token = selected_account.get("token")
        request.state.account = selected_account

    else:
        request.state.use_config_token = False
        request.state.deepseek_token = caller_key


def get_auth_headers(request: Request):
    """返回 DeepSeek 请求所需的公共请求头"""
    return {**BASE_HEADERS, "authorization": f"Bearer {request.state.deepseek_token}"}


# ----------------------------------------------------------------------
# (6) 封装对话接口调用的重试机制
# ----------------------------------------------------------------------
def call_completion_endpoint(payload, headers, max_attempts=3):
    attempts = 0
    while attempts < max_attempts:
        try:
            deepseek_resp = requests.post(
                DEEPSEEK_COMPLETION_URL, headers=headers, json=payload, stream=True
            )
        except Exception as e:
            logger.warning(f"[call_completion_endpoint] 请求异常: {e}")
            time.sleep(1)
            attempts += 1
            continue
        if deepseek_resp.status_code == 200:
            return deepseek_resp
        else:
            logger.warning(
                f"[call_completion_endpoint] 调用对话接口失败, 状态码: {deepseek_resp.status_code}"
            )
            deepseek_resp.close()
            time.sleep(1)
            attempts += 1
    return None


# ----------------------------------------------------------------------
# (7) 创建会话 & 获取 PoW（重试时，配置模式下错误会切换账号；用户自带 token 模式下仅重试）
# ----------------------------------------------------------------------
def create_session(request: Request, max_attempts=3):
    attempts = 0
    while attempts < max_attempts:
        headers = get_auth_headers(request)
        try:
            resp = requests.post(
                DEEPSEEK_CREATE_SESSION_URL, headers=headers, json={"agent": "chat"}
            )
        except Exception as e:
            logger.error(f"[create_session] 请求异常: {e}")
            attempts += 1
            continue
        try:
            data = resp.json()
        except Exception as e:
            logger.error(f"[create_session] JSON解析异常: {e}")
            data = {}
        if resp.status_code == 200 and data.get("code") == 0:
            session_id = data["data"]["biz_data"]["id"]

            resp.close()
            return session_id
        else:
            code = data.get("code")
            logger.warning(
                f"[create_session] 创建会话失败, code={code}, msg={data.get('msg')}"
            )
            resp.close()
            if request.state.use_config_token:
                current_id = get_account_identifier(request.state.account)
                if not hasattr(request.state, "tried_accounts"):
                    request.state.tried_accounts = []
                if current_id not in request.state.tried_accounts:
                    request.state.tried_accounts.append(current_id)
                new_account = choose_new_account(request.state.tried_accounts)
                if new_account is None:
                    break
                try:
                    login_deepseek_via_account(new_account)
                except Exception as e:
                    logger.error(
                        f"[create_session] 账号 {get_account_identifier(new_account)} 登录失败：{e}"
                    )
                    attempts += 1
                    continue
                request.state.account = new_account
                request.state.deepseek_token = new_account.get("token")
            else:
                attempts += 1
                continue
        attempts += 1
    return None


# ----------------------------------------------------------------------
# (7.1) 使用 WASM 模块计算 PoW 答案的辅助函数
# ----------------------------------------------------------------------
def compute_pow_answer(
    algorithm: str,
    challenge_str: str,
    salt: str,
    difficulty: int,
    expire_at: int,
    signature: str,
    target_path: str,
    wasm_path: str,
) -> int:
    """
    使用 WASM 模块计算 DeepSeekHash 答案（answer）。
    根据 JS 逻辑：
      - 拼接前缀： "{salt}_{expire_at}_"
      - 将 challenge 与前缀写入 wasm 内存后调用 wasm_solve 进行求解，
      - 从 wasm 内存中读取状态与求解结果，
      - 若状态非 0，则返回整数形式的答案，否则返回 None。
    """
    if algorithm != "DeepSeekHashV1":
        raise ValueError(f"不支持的算法：{algorithm}")
    prefix = f"{salt}_{expire_at}_"
    # --- 加载 wasm 模块 ---
    store = Store()
    linker = Linker(store.engine)
    try:
        with open(wasm_path, "rb") as f:
            wasm_bytes = f.read()
    except Exception as e:
        raise RuntimeError(f"加载 wasm 文件失败: {wasm_path}, 错误: {e}")
    module = Module(store.engine, wasm_bytes)
    instance = linker.instantiate(store, module)
    exports = instance.exports(store)
    try:
        memory = exports["memory"]
        add_to_stack = exports["__wbindgen_add_to_stack_pointer"]
        alloc = exports["__wbindgen_export_0"]
        wasm_solve = exports["wasm_solve"]
    except KeyError as e:
        raise RuntimeError(f"缺少 wasm 导出函数: {e}")

    def write_memory(offset: int, data: bytes):
        size = len(data)
        base_addr = ctypes.cast(memory.data_ptr(store), ctypes.c_void_p).value
        ctypes.memmove(base_addr + offset, data, size)

    def read_memory(offset: int, size: int) -> bytes:
        base_addr = ctypes.cast(memory.data_ptr(store), ctypes.c_void_p).value
        return ctypes.string_at(base_addr + offset, size)

    def encode_string(text: str):
        data = text.encode("utf-8")
        length = len(data)
        ptr_val = alloc(store, length, 1)
        ptr = int(ptr_val.value) if hasattr(ptr_val, "value") else int(ptr_val)
        write_memory(ptr, data)
        return ptr, length

    # 1. 申请 16 字节栈空间
    retptr = add_to_stack(store, -16)
    # 2. 编码 challenge 与 prefix 到 wasm 内存中
    ptr_challenge, len_challenge = encode_string(challenge_str)
    ptr_prefix, len_prefix = encode_string(prefix)
    # 3. 调用 wasm_solve（注意：difficulty 以 float 形式传入）
    wasm_solve(
        store,
        retptr,
        ptr_challenge,
        len_challenge,
        ptr_prefix,
        len_prefix,
        float(difficulty),
    )
    # 4. 从 retptr 处读取 4 字节状态和 8 字节求解结果
    status_bytes = read_memory(retptr, 4)
    if len(status_bytes) != 4:
        add_to_stack(store, 16)
        raise RuntimeError("读取状态字节失败")
    status = struct.unpack("<i", status_bytes)[0]
    value_bytes = read_memory(retptr + 8, 8)
    if len(value_bytes) != 8:
        add_to_stack(store, 16)
        raise RuntimeError("读取结果字节失败")
    value = struct.unpack("<d", value_bytes)[0]
    # 5. 恢复栈指针
    add_to_stack(store, 16)
    if status == 0:
        return None
    return int(value)


# ----------------------------------------------------------------------
# (7.2) 获取 PoW 响应，融合计算 answer 逻辑
# ----------------------------------------------------------------------
def get_pow_response(request: Request, max_attempts=3):
    attempts = 0
    while attempts < max_attempts:
        headers = get_auth_headers(request)
        try:
            resp = requests.post(
                DEEPSEEK_CREATE_POW_URL,
                headers=headers,
                json={"target_path": "/api/v0/chat/completion"},
                timeout=30,
            )
        except Exception as e:
            logger.error(f"[get_pow_response] 请求异常: {e}")
            attempts += 1
            continue
        try:
            data = resp.json()
        except Exception as e:
            logger.error(f"[get_pow_response] JSON解析异常: {e}")
            data = {}
        if resp.status_code == 200 and data.get("code") == 0:
            challenge = data["data"]["biz_data"]["challenge"]
            difficulty = challenge.get("difficulty", 144000)
            expire_at = challenge.get("expire_at", 1680000000)
            try:
                answer = compute_pow_answer(
                    challenge["algorithm"],
                    challenge["challenge"],
                    challenge["salt"],
                    difficulty,
                    expire_at,
                    challenge["signature"],
                    challenge["target_path"],
                    WASM_PATH,
                )
            except Exception as e:
                logger.error(f"[get_pow_response] PoW 答案计算异常: {e}")
                answer = None
            if answer is None:
                logger.warning("[get_pow_response] PoW 答案计算失败，重试中...")
                resp.close()
                attempts += 1
                continue
            pow_dict = {
                "algorithm": challenge["algorithm"],
                "challenge": challenge["challenge"],
                "salt": challenge["salt"],
                "answer": answer,  # 整数形式答案
                "signature": challenge["signature"],
                "target_path": challenge["target_path"],
            }
            pow_str = json.dumps(pow_dict, separators=(",", ":"), ensure_ascii=False)
            encoded = base64.b64encode(pow_str.encode("utf-8")).decode("utf-8").rstrip()
            resp.close()
            return encoded
        else:
            code = data.get("code")
            logger.warning(
                f"[get_pow_response] 获取 PoW 失败, code={code}, msg={data.get('msg')}"
            )
            resp.close()
            if request.state.use_config_token:
                current_id = get_account_identifier(request.state.account)
                if not hasattr(request.state, "tried_accounts"):
                    request.state.tried_accounts = []
                if current_id not in request.state.tried_accounts:
                    request.state.tried_accounts.append(current_id)
                new_account = choose_new_account(request.state.tried_accounts)
                if new_account is None:
                    break
                try:
                    login_deepseek_via_account(new_account)
                except Exception as e:
                    logger.error(
                        f"[get_pow_response] 账号 {get_account_identifier(new_account)} 登录失败：{e}"
                    )
                    attempts += 1
                    continue
                request.state.account = new_account
                request.state.deepseek_token = new_account.get("token")
            else:
                attempts += 1
                continue
            attempts += 1
    return None


# ----------------------------------------------------------------------
# (8) 路由：/v1/models
# ----------------------------------------------------------------------
@app.get("/v1/models")
def list_models():
    models_list = [
        {
            "id": "deepseek-chat",
            "object": "model",
            "created": 1677610602,
            "owned_by": "deepseek",
            "permission": [],
        },
        {
            "id": "deepseek-reasoner",
            "object": "model",
            "created": 1677610602,
            "owned_by": "deepseek",
            "permission": [],
        },
        {
            "id": "deepseek-chat-search",
            "object": "model",
            "created": 1677610602,
            "owned_by": "deepseek",
            "permission": [],
        },
        {
            "id": "deepseek-reasoner-search",
            "object": "model",
            "created": 1677610602,
            "owned_by": "deepseek",
            "permission": [],
        },
    ]
    data = {"object": "list", "data": models_list}
    return JSONResponse(content=data, status_code=200)


# ----------------------------------------------------------------------
# 消息预处理函数，将多轮对话合并成最终 prompt
# ----------------------------------------------------------------------
def messages_prepare(messages: list) -> str:
    """处理消息列表，合并连续相同角色的消息，并添加角色标签：
    - 对于 assistant 消息，加上 <｜Assistant｜> 前缀及 <｜end▁of▁sentence｜> 结束标签；
    - 对于 user/system 消息（除第一条外）加上 <｜User｜> 前缀；
    - 如果消息 content 为数组，则提取其中 type 为 "text" 的部分；
    - 最后移除 markdown 图片格式的内容。
    """
    processed = []
    for m in messages:
        role = m.get("role", "")
        content = m.get("content", "")
        if isinstance(content, list):
            texts = [
                item.get("text", "") for item in content if item.get("type") == "text"
            ]
            text = "\n".join(texts)
        else:
            text = str(content)
        processed.append({"role": role, "text": text})
    if not processed:
        return ""
    # 合并连续同一角色的消息
    merged = [processed[0]]
    for msg in processed[1:]:
        if msg["role"] == merged[-1]["role"]:
            merged[-1]["text"] += "\n\n" + msg["text"]
        else:
            merged.append(msg)
    # 添加标签
    parts = []
    for idx, block in enumerate(merged):
        role = block["role"]
        text = block["text"]
        if role == "assistant":
            parts.append(f"<｜Assistant｜>{text}<｜end▁of▁sentence｜>")
        elif role in ("user", "system"):
            if idx > 0:
                parts.append(f"<｜User｜>{text}")
            else:
                parts.append(text)
        else:
            parts.append(text)
    final_prompt = "".join(parts)
    # 仅移除 markdown 图片格式(不全部移除 !）
    final_prompt = re.sub(r"!\[(.*?)\]\((.*?)\)", r"[\1](\2)", final_prompt)
    return final_prompt


# 添加保活超时配置（5秒）
KEEP_ALIVE_TIMEOUT = 5


# ----------------------------------------------------------------------
# (10) 路由：/v1/chat/completions
# ----------------------------------------------------------------------
@app.post("/v1/chat/completions")
async def chat_completions(request: Request):
    try:
        # 处理 token 相关逻辑，若登录失败则直接返回错误响应
        try:
            determine_mode_and_token(request)
        except HTTPException as exc:
            return JSONResponse(
                status_code=exc.status_code, content={"error": exc.detail}
            )
        except Exception as exc:
            logger.error(f"[chat_completions] determine_mode_and_token 异常: {exc}")
            return JSONResponse(
                status_code=500, content={"error": "Account login failed."}
            )

        req_data = await request.json()
        model = req_data.get("model")
        messages = req_data.get("messages", [])
        if not model or not messages:
            raise HTTPException(
                status_code=400, detail="Request must include 'model' and 'messages'."
            )
        # 判断是否启用"思考"或"搜索"功能（这里根据模型名称判断）
        model_lower = model.lower()
        if model_lower in ["deepseek-v3", "deepseek-chat"]:
            thinking_enabled = False
            search_enabled = False
        elif model_lower in ["deepseek-r1", "deepseek-reasoner"]:
            thinking_enabled = True
            search_enabled = False
        elif model_lower in ["deepseek-v3-search", "deepseek-chat-search"]:
            thinking_enabled = False
            search_enabled = True
        elif model_lower in ["deepseek-r1-search", "deepseek-reasoner-search"]:
            thinking_enabled = True
            search_enabled = True
        else:
            raise HTTPException(
                status_code=503, detail=f"Model '{model}' is not available."
            )
        # 使用 messages_prepare 函数构造最终 prompt
        final_prompt = messages_prepare(messages)
        session_id = create_session(request)
        if not session_id:
            raise HTTPException(status_code=401, detail="invalid token.")
        pow_resp = get_pow_response(request)
        if not pow_resp:
            raise HTTPException(
                status_code=401,
                detail="Failed to get PoW (invalid token or unknown error).",
            )
        headers = {**get_auth_headers(request), "x-ds-pow-response": pow_resp}
        payload = {
            "chat_session_id": session_id,
            "parent_message_id": None,
            "prompt": final_prompt,
            "ref_file_ids": [],
            "thinking_enabled": thinking_enabled,
            "search_enabled": search_enabled,
        }

        deepseek_resp = call_completion_endpoint(payload, headers, max_attempts=3)
        if not deepseek_resp:
            raise HTTPException(status_code=500, detail="Failed to get completion.")
        created_time = int(time.time())
        completion_id = f"{session_id}"

        # 流式响应（SSE）或普通响应
        if bool(req_data.get("stream", False)):
            if deepseek_resp.status_code != 200:
                deepseek_resp.close()
                return JSONResponse(
                    content=deepseek_resp.content, status_code=deepseek_resp.status_code
                )

            def sse_stream():
                try:
                    final_text = ""
                    final_thinking = ""
                    first_chunk_sent = False
                    result_queue = queue.Queue()
                    last_send_time = time.time()
                    citation_map = {}  # 用于存储引用链接的字典

                    def process_data():
                        try:
                            for raw_line in deepseek_resp.iter_lines():
                                try:
                                    line = raw_line.decode("utf-8")
                                except Exception as e:
                                    logger.warning(f"[sse_stream] 解码失败: {e}")
                                    busy_content_str = '{"choices":[{"index":0,"delta":{"content":"服务器繁忙，请稍候再试","type":"text"}}],"model":"","chunk_token_usage":1,"created":0,"message_id":-1,"parent_id":-1}'
                                    busy_content = json.loads(busy_content_str)
                                    result_queue.put(busy_content)
                                    result_queue.put(None)
                                    break
                                if not line:
                                    continue
                                if line.startswith("data:"):
                                    data_str = line[5:].strip()
                                    if data_str == "[DONE]":
                                        result_queue.put(None)  # 结束信号
                                        break
                                    try:
                                        chunk = json.loads(data_str)
                                        # 处理搜索索引数据
                                        if (
                                            chunk.get("choices", [{}])[0]
                                            .get("delta", {})
                                            .get("type")
                                            == "search_index"
                                        ):
                                            search_indexes = chunk["choices"][0][
                                                "delta"
                                            ].get("search_indexes", [])
                                            for idx in search_indexes:
                                                citation_map[
                                                    str(idx.get("cite_index"))
                                                ] = idx.get("url", "")
                                            continue
                                        # if (
                                            # chunk.get("choices", [{}])[0]
                                            # .get("finish_reason")
                                            # == "backend_busy"
                                        # ):
                                            # busy_content_str = '{"choices":[{"index":0,"delta":{"content":"服务器繁忙，请稍候再试","type":"text"}}],"model":"","chunk_token_usage":1,"created":0,"message_id":-1,"parent_id":-1}'
                                            # busy_content = json.loads(busy_content_str)
                                            # result_queue.put(busy_content)
                                            # result_queue.put(None)
                                            # break
                                        result_queue.put(chunk)  # 将数据放入队列
                                    except Exception as e:
                                        logger.warning(
                                            f"[sse_stream] 无法解析: {data_str}, 错误: {e}"
                                        )
                                        busy_content_str = '{"choices":[{"index":0,"delta":{"content":"服务器繁忙，请稍候再试","type":"text"}}],"model":"","chunk_token_usage":1,"created":0,"message_id":-1,"parent_id":-1}'
                                        busy_content = json.loads(busy_content_str)
                                        result_queue.put(busy_content)
                                        result_queue.put(None)
                                        break
                        except Exception as e:
                            logger.warning(f"[sse_stream] 错误: {e}")
                            busy_content_str = '{"choices":[{"index":0,"delta":{"content":"服务器繁忙，请稍候再试","type":"text"}}],"model":"","chunk_token_usage":1,"created":0,"message_id":-1,"parent_id":-1}'
                            busy_content = json.loads(busy_content_str)
                            result_queue.put(busy_content)
                            result_queue.put(None)
                            # raise HTTPException(
                                # status_code=500, detail="Server is error."
                            # )
                        finally:
                            deepseek_resp.close()

                    process_thread = threading.Thread(target=process_data)
                    process_thread.start()

                    while True:
                        current_time = time.time()
                        if current_time - last_send_time >= KEEP_ALIVE_TIMEOUT:

                            yield ": keep-alive\n\n"
                            last_send_time = current_time
                            continue
                        try:
                            chunk = result_queue.get(timeout=0.1)
                            if chunk is None:
                                # 发送最终统计信息
                                prompt_tokens = len(tokenizer.encode(final_prompt))
                                completion_tokens = len(tokenizer.encode(final_text))
                                usage = {
                                    "prompt_tokens": prompt_tokens,
                                    "completion_tokens": completion_tokens,
                                    "total_tokens": prompt_tokens + completion_tokens,
                                }
                                finish_chunk = {
                                    "id": completion_id,
                                    "object": "chat.completion.chunk",
                                    "created": created_time,
                                    "model": model,
                                    "choices": [
                                        {
                                            "delta": {},
                                            "index": 0,
                                            "finish_reason": "stop",
                                        }
                                    ],
                                    "usage": usage,
                                }
                                yield f"data: {json.dumps(finish_chunk, ensure_ascii=False)}\n\n"
                                yield "data: [DONE]\n\n"
                                last_send_time = current_time
                                break
                            new_choices = []
                            for choice in chunk.get("choices", []):
                                delta = choice.get("delta", {})
                                ctype = delta.get("type")
                                ctext = delta.get("content", "")
                                if (
                                    choice
                                    .get("finish_reason")
                                    == "backend_busy"
                                ):
                                    ctext = '服务器繁忙，请稍候再试'
                                if search_enabled and ctext.startswith("[citation:"):
                                    ctext = ""
                                if ctype == "thinking":
                                    if thinking_enabled:
                                        final_thinking += ctext
                                elif ctype == "text":
                                    final_text += ctext
                                delta_obj = {}
                                if not first_chunk_sent:
                                    delta_obj["role"] = "assistant"
                                    first_chunk_sent = True
                                if ctype == "thinking":
                                    if thinking_enabled:
                                        delta_obj["reasoning_content"] = ctext
                                elif ctype == "text":
                                    delta_obj["content"] = ctext
                                if delta_obj:
                                    new_choices.append(
                                        {
                                            "delta": delta_obj,
                                            "index": choice.get("index", 0),
                                        }
                                    )
                            if new_choices:
                                out_chunk = {
                                    "id": completion_id,
                                    "object": "chat.completion.chunk",
                                    "created": created_time,
                                    "model": model,
                                    "choices": new_choices,
                                }
                                yield f"data: {json.dumps(out_chunk, ensure_ascii=False)}\n\n"
                                last_send_time = current_time
                        except queue.Empty:
                            continue
                except Exception as e:
                    logger.error(f"[sse_stream] 异常: {e}")
                finally:
                    if getattr(request.state, "use_config_token", False) and hasattr(
                        request.state, "account"
                    ):
                        release_account(request.state.account)

            return StreamingResponse(
                sse_stream(),
                media_type="text/event-stream",
                headers={"Content-Type": "text/event-stream"},
            )
        else:
            # 非流式响应处理
            think_list = []
            text_list = []
            result = None
            citation_map = {}

            data_queue = queue.Queue()

            def collect_data():
                nonlocal result
                try:
                    for raw_line in deepseek_resp.iter_lines():
                        try:
                            line = raw_line.decode("utf-8")
                        except Exception as e:
                            logger.warning(f"[chat_completions] 解码失败: {e}")
                            ctext = '服务器繁忙，请稍候再试'
                            text_list.append(ctext)
                            data_queue.put(None)
                            break
                        if not line:
                            continue
                        if line.startswith("data:"):
                            data_str = line[5:].strip()
                            if data_str == "[DONE]":
                                data_queue.put(None)
                                break
                            try:
                                chunk = json.loads(data_str)
                                if (
                                    chunk.get("choices", [{}])[0]
                                    .get("delta", {})
                                    .get("type")
                                    == "search_index"
                                ):
                                    search_indexes = chunk["choices"][0]["delta"].get(
                                        "search_indexes", []
                                    )
                                    for idx in search_indexes:
                                        citation_map[str(idx.get("cite_index"))] = (
                                            idx.get("url", "")
                                        )
                                    continue
                                for choice in chunk.get("choices", []):
                                    delta = choice.get("delta", {})
                                    ctype = delta.get("type")
                                    ctext = delta.get("content", "")
                                    if (
                                        choice
                                        .get("finish_reason")
                                        == "backend_busy"
                                    ):
                                        ctext = '服务器繁忙，请稍候再试'
                                        # data_queue.put(None)
                                        # break
                                    if search_enabled and ctext.startswith(
                                        "[citation:"
                                    ):
                                        ctext = ""
                                    if ctype == "thinking" and thinking_enabled:
                                        think_list.append(ctext)
                                    elif ctype == "text":
                                        text_list.append(ctext)
                            except Exception as e:
                                logger.warning(
                                    f"[collect_data] 无法解析: {data_str}, 错误: {e}"
                                )
                                ctext = '服务器繁忙，请稍候再试'
                                text_list.append(ctext)
                                data_queue.put(None)
                                break
                except Exception as e:
                    logger.warning(f"[collect_data] 错误: {e}")
                    ctext = '服务器繁忙，请稍候再试'
                    text_list.append(ctext)
                    data_queue.put(None)
                finally:
                    deepseek_resp.close()
                    final_reasoning = "".join(think_list)
                    final_content = "".join(text_list)
                    prompt_tokens = len(tokenizer.encode(final_prompt))
                    completion_tokens = len(tokenizer.encode(final_content))
                    result = {
                        "id": completion_id,
                        "object": "chat.completion",
                        "created": created_time,
                        "model": model,
                        "choices": [
                            {
                                "index": 0,
                                "message": {
                                    "role": "assistant",
                                    "content": final_content,
                                    "reasoning_content": final_reasoning,
                                },
                                "finish_reason": "stop",
                            }
                        ],
                        "usage": {
                            "prompt_tokens": prompt_tokens,
                            "completion_tokens": completion_tokens,
                            "total_tokens": prompt_tokens + completion_tokens,
                        },
                    }
                    data_queue.put("DONE")

            collect_thread = threading.Thread(target=collect_data)
            collect_thread.start()

            def generate():
                last_send_time = time.time()
                while True:
                    current_time = time.time()
                    if current_time - last_send_time >= KEEP_ALIVE_TIMEOUT:

                        yield ""
                        last_send_time = current_time
                    if not collect_thread.is_alive() and result is not None:
                        yield json.dumps(result)
                        break
                    time.sleep(0.1)

            return StreamingResponse(generate(), media_type="application/json")
    except HTTPException as exc:
        return JSONResponse(status_code=exc.status_code, content={"error": exc.detail})
    except Exception as exc:
        logger.error(f"[chat_completions] 未知异常: {exc}")
        return JSONResponse(status_code=500, content={"error": "Internal Server Error"})
    finally:
        if getattr(request.state, "use_config_token", False) and hasattr(
            request.state, "account"
        ):
            release_account(request.state.account)


# ----------------------------------------------------------------------
# (11) 路由：/
# ----------------------------------------------------------------------
@app.get("/")
def index(request: Request):
    return templates.TemplateResponse("welcome.html", {"request": request})


# ----------------------------------------------------------------------
# 启动 FastAPI 应用
# ----------------------------------------------------------------------
if __name__ == "__main__":
    import uvicorn

    uvicorn.run(app, host="0.0.0.0", port=5001)