local-llm-server/llm_server/threads.py

import time
from threading import Thread

from llm_server import opts
from llm_server.database.database import weighted_average_column_for_model
from llm_server.llm.info import get_running_model
from llm_server.routes.cache import redis
from llm_server.routes.v1.generate_stats import generate_stats


class MainBackgroundThread(Thread):
    backend_online = False

    # TODO: do I really need to put everything in Redis?
    # TODO: call generate_stats() every minute, cache the results, put results in a DB table, then have other parts of code call this cache

    def __init__(self):
        Thread.__init__(self)
        self.daemon = True
        redis.set('average_generation_elapsed_sec', 0)
        redis.set('estimated_avg_tps', 0)
        redis.set('average_output_tokens', 0)
        redis.set('backend_online', 0)
        redis.set_dict('backend_info', {})

    def run(self):
        while True:
            if opts.mode == 'oobabooga':
                model, err = get_running_model()
                if err:
                    print(err)
                    redis.set('backend_online', 0)
                else:
                    opts.running_model = model
                    redis.set('backend_online', 1)
            elif opts.mode == 'vllm':
                model, err = get_running_model()
                if err:
                    print(err)
                    redis.set('backend_online', 0)
                else:
                    opts.running_model = model
                    redis.set('backend_online', 1)
            else:
                raise Exception

            # exclude_zeros=True filters out rows where an error message was returned. Previously, if there was an error, 0
            # was entered into the column. The new code enters null instead but we need to be backwards compatible for now.
            average_generation_elapsed_sec = weighted_average_column_for_model('prompts', 'generation_time', opts.running_model, opts.mode, opts.backend_url, exclude_zeros=True) or 0
            if average_generation_elapsed_sec > -1:
                redis.set('average_generation_elapsed_sec', average_generation_elapsed_sec)

            # overall = average_column_for_model('prompts', 'generation_time', opts.running_model)
            # print(f'Weighted: {average_generation_elapsed_sec}, overall: {overall}')

            average_output_tokens = weighted_average_column_for_model('prompts', 'response_tokens', opts.running_model, opts.mode, opts.backend_url, exclude_zeros=True) or 0
            if average_generation_elapsed_sec > -1:
                redis.set('average_output_tokens', average_output_tokens)

            # overall = average_column_for_model('prompts', 'response_tokens', opts.running_model)
            # print(f'Weighted: {average_output_tokens}, overall: {overall}')

            estimated_avg_tps = round(average_output_tokens / average_generation_elapsed_sec, 2) if average_generation_elapsed_sec > 0 else 0  # Avoid division by zero
            redis.set('estimated_avg_tps', estimated_avg_tps)
            time.sleep(60)


def cache_stats():
    while True:
        # If opts.base_client_api is null that means no one has visited the site yet
        # and the base_client_api hasn't been set. Do nothing until then.
        if redis.get('base_client_api'):
            x = generate_stats()
        time.sleep(5)