local-llm-server/llm_server/threads.py

import time
from threading import Thread

import requests

from llm_server import opts
from llm_server.database import average_column
from llm_server.routes.cache import redis


class BackendHealthCheck(Thread):
    backend_online = False

    def __init__(self):
        Thread.__init__(self)
        self.daemon = True
        redis.set('average_generation_elapsed_sec', 0)
        redis.set('average_tps', 0)
        redis.set('average_output_tokens', 0)
        redis.set('backend_online', 0)

    def run(self):
        while True:
            average_generation_elapsed_sec = average_column('prompts', 'generation_time') if not None else 0
            redis.set('average_generation_elapsed_sec', average_generation_elapsed_sec)

            average_output_tokens = average_column('prompts', 'response_tokens') if not None else 0
            redis.set('average_output_tokens', average_output_tokens)
            average_tps = round(average_output_tokens / average_generation_elapsed_sec, 2)
            redis.set('average_tps', average_tps)

            if opts.mode == 'oobabooga':
                try:
                    r = requests.get(f'{opts.backend_url}/api/v1/model', timeout=3, verify=opts.verify_ssl)
                    opts.running_model = r.json()['result']
                    redis.set('backend_online', 1)
                except Exception as e:
                    redis.set('backend_online', 0)
                    # TODO: handle error
                    print(e)
            elif opts.mode == 'hf-textgen':
                pass
            else:
                raise Exception
            time.sleep(60)