llava/utils.py

import datetime
import logging
import logging.handlers
import os
import sys

import requests

from llava.constants import LOGDIR

server_error_msg = "**由于高流量导致网络错误。请重新生成或刷新本页面。**"
moderation_msg = "您的输入违反了我们的内容审核准则。请重试。"

handler = None


def build_logger(logger_name, logger_filename):
    global handler

    formatter = logging.Formatter(
        fmt="%(asctime)s | %(levelname)s | %(name)s | %(message)s",
        datefmt="%Y-%m-%d %H:%M:%S",
    )

    # 设置根处理程序的格式
    if not logging.getLogger().handlers:
        logging.basicConfig(level=logging.INFO)
    logging.getLogger().handlers[0].setFormatter(formatter)

    # 将 stdout 和 stderr 重定向到日志记录器
    stdout_logger = logging.getLogger("stdout")
    stdout_logger.setLevel(logging.INFO)
    sl = StreamToLogger(stdout_logger, logging.INFO)
    sys.stdout = sl

    stderr_logger = logging.getLogger("stderr")
    stderr_logger.setLevel(logging.ERROR)
    sl = StreamToLogger(stderr_logger, logging.ERROR)
    sys.stderr = sl

    # 获取日志
    logger = logging.getLogger(logger_name)
    logger.setLevel(logging.INFO)

    # 为所有日志记录器添加文件处理程序
    if handler is None:
        os.makedirs(LOGDIR, exist_ok=True)
        filename = os.path.join(LOGDIR, logger_filename)
        handler = logging.handlers.TimedRotatingFileHandler(
            filename, when='D', utc=True)
        handler.setFormatter(formatter)

        for name, item in logging.root.manager.loggerDict.items():
            if isinstance(item, logging.Logger):
                item.addHandler(handler)

    return logger


class StreamToLogger(object):
    """
    伪造的类文件流对象，用于将写入内容重定向到日志记录器实例。
    """
    def __init__(self, logger, log_level=logging.INFO):
        self.terminal = sys.stdout
        self.logger = logger
        self.log_level = log_level
        self.linebuf = ''

    def __getattr__(self, attr):
        return getattr(self.terminal, attr)

    def write(self, buf):
        temp_linebuf = self.linebuf + buf
        self.linebuf = ''
        for line in temp_linebuf.splitlines(True):
            # 来自 io.TextIOWrapper 的文档：
            #   On output, if newline is None, any '\n' characters written
            #   are translated to the system default line separator.
            # 默认情况下，sys.stdout.write() 希望使用“\n”换行符然后再
            # 翻译它们，所以这始终是跨平台的。
            if line[-1] == '\n':
                self.logger.log(self.log_level, line.rstrip())
            else:
                self.linebuf += line

    def flush(self):
        if self.linebuf != '':
            self.logger.log(self.log_level, self.linebuf.rstrip())
        self.linebuf = ''


def disable_torch_init():
    """
    禁用多余的 torch 默认初始化，以加快模型创建速度。
    """
    import torch
    setattr(torch.nn.Linear, "reset_parameters", lambda self: None)
    setattr(torch.nn.LayerNorm, "reset_parameters", lambda self: None)


def violates_moderation(text):
    """
    检查文本是否违反 OpenAI moderation API。
    """
    url = "https://api.openai.com/v1/moderations"
    headers = {"Content-Type": "application/json",
               "Authorization": "Bearer " + os.environ["OPENAI_API_KEY"]}
    text = text.replace("\n", "")
    data = "{" + '"input": ' + f'"{text}"' + "}"
    data = data.encode("utf-8")
    try:
        ret = requests.post(url, headers=headers, data=data, timeout=5)
        flagged = ret.json()["results"][0]["flagged"]
    except requests.exceptions.RequestException as e:
        flagged = False
    except KeyError as e:
        flagged = False

    return flagged


def pretty_print_semaphore(semaphore):
    if semaphore is None:
        return "None"
    return f"Semaphore(value={semaphore._value}, locked={semaphore.locked()})"