local-llm-server/server.py

import simplejson as json
from flask import Flask, jsonify, render_template, request, Response

from llm_server.cluster.backend import get_model_choices
from llm_server.cluster.cluster_config import cluster_config
from llm_server.config.config import MODE_UI_NAMES
from llm_server.config.global_config import GlobalConfig
from llm_server.custom_redis import flask_cache, redis
from llm_server.helpers import auto_set_base_client_api
from llm_server.llm.vllm.info import vllm_info
from llm_server.routes.openai import openai_bp, openai_model_bp
from llm_server.routes.server_error import handle_server_error
from llm_server.routes.v1 import bp
from llm_server.routes.v1.generate_stats import generate_stats
from llm_server.sock import init_wssocket

# TODO: seperate queue item timeout for websockets (make longer, like 5 minutes)
# TODO: return an `error: True`, error code, and error message rather than just a formatted message
# TODO: what happens when all backends are offline? What about the "online" key in the stats page?
# TODO: redis SCAN vs KEYS??
# TODO: is frequency penalty the same as ooba repetition penalty???
# TODO: make sure openai_moderation_enabled works on websockets, completions, and chat completions

# Lower priority
# TODO: if a backend is at its limit of concurrent requests, choose a different one
# TODO: make error messages consitient
# TODO: support logit_bias on OpenAI and Ooba endpoints.
# TODO: add a way to cancel VLLM gens. Maybe use websockets?
# TODO: validate openai_silent_trim works as expected and only when enabled
# TODO: rewrite config storage. Store in redis so we can reload it.
# TODO: set VLLM to stream ALL data using socket.io. If the socket disconnects, cancel generation.
# TODO: estiamted wait time needs to account for full concurrent_gens but the queue is less than concurrent_gens
# TODO: the estiamted wait time lags behind the stats
# TODO: simulate OpenAI error messages regardless of endpoint
# TODO: send extra headers when ratelimited?
# TODO: make sure log_prompt() is used everywhere, including errors and invalid requests
# TODO: unify logging thread in a function and use async/await instead
# TODO: move the netdata stats to a seperate part of the stats and have it set to the currently selected backend
# TODO: have VLLM reply with stats (TPS, generated token count, processing time)
# TODO: add config reloading via stored redis variables

# Done, but need to verify
# TODO: add more excluding to SYSTEM__ tokens
# TODO: return 200 when returning formatted sillytavern error


app = Flask(__name__)

# Fixes ConcurrentObjectUseError
# https://github.com/miguelgrinberg/simple-websocket/issues/24
app.config['SOCK_SERVER_OPTIONS'] = {'ping_interval': 25}

app.register_blueprint(bp, url_prefix='/api/')
app.register_blueprint(openai_bp, url_prefix='/api/openai/v1/')
app.register_blueprint(openai_model_bp, url_prefix='/api/openai/')
init_wssocket(app)
flask_cache.init_app(app)
flask_cache.clear()


@app.route('/')
@app.route('/api')
@app.route('/api/openai')
@flask_cache.cached(timeout=10)
def home():
    base_client_api = redis.get('base_client_api', dtype=str)
    stats = generate_stats()
    model_choices, default_model = get_model_choices()

    if default_model:
        if not model_choices.get(default_model):
            return 'The server is still starting up. Please wait...'

        default_model_info = model_choices[default_model]

        if default_model_info['queued'] == 0 and default_model_info['queued'] >= default_model_info['concurrent_gens']:
            # There will be a wait if the queue is empty but prompts are processing, but we don't
            # know how long.
            default_estimated_wait_sec = f"less than {int(default_model_info['estimated_wait'])} seconds"
        else:
            default_estimated_wait_sec = f"{int(default_model_info['estimated_wait'])} seconds"
    else:
        default_model_info = {
            'model': 'OFFLINE',
            'processing': '-',
            'queued': '-',
            'context_size': '-',
        }
        default_estimated_wait_sec = 'OFFLINE'

    if default_model_info['context_size'] is None:
        # Sometimes a model doesn't provide the correct config, so the context size is set
        # to None by the daemon.
        default_model_info['context_size'] = '-'

    if len(GlobalConfig.get().analytics_tracking_code):
        analytics_tracking_code = f"<script>\n{GlobalConfig.get().analytics_tracking_code}\n</script>"
    else:
        analytics_tracking_code = ''

    if GlobalConfig.get().info_html:
        info_html = GlobalConfig.get().info_html
    else:
        info_html = ''

    mode_info = ''
    for k, v in cluster_config.all().items():
        if v['mode'] == 'vllm':
            mode_info = vllm_info
            break

    return render_template('home.html',
                           llm_middleware_name=GlobalConfig.get().llm_middleware_name,
                           analytics_tracking_code=analytics_tracking_code,
                           info_html=info_html,
                           default_model=default_model_info['model'],
                           default_active_gen_workers=default_model_info['processing'],
                           default_proompters_in_queue=default_model_info['queued'],
                           current_model=GlobalConfig.get().manual_model_name if GlobalConfig.get().manual_model_name else None,  # else running_model,
                           client_api=f'https://{base_client_api}',
                           ws_client_api=f'wss://{base_client_api}/v1/stream' if GlobalConfig.get().enable_streaming else 'disabled',
                           default_estimated_wait=default_estimated_wait_sec,
                           mode_name=MODE_UI_NAMES[GlobalConfig.get().frontend_api_mode].name,
                           api_input_textbox=MODE_UI_NAMES[GlobalConfig.get().frontend_api_mode].api_name,
                           streaming_input_textbox=MODE_UI_NAMES[GlobalConfig.get().frontend_api_mode].streaming_name,
                           default_context_size=default_model_info['context_size'],
                           stats_json=json.dumps(stats, indent=4, ensure_ascii=False),
                           extra_info=mode_info,
                           openai_client_api=f'https://{base_client_api}/openai/v1' if GlobalConfig.get().enable_openi_compatible_backend else 'disabled',
                           expose_openai_system_prompt=GlobalConfig.get().expose_openai_system_prompt,
                           enable_streaming=GlobalConfig.get().enable_streaming,
                           model_choices=model_choices,
                           proompters_5_min=stats['stats']['proompters']['5_min'],
                           proompters_24_hrs=stats['stats']['proompters']['24_hrs'],
                           )


@app.route('/robots.txt')
def robots():
    # TODO: have config value to deny all
    # TODO: https://developers.google.com/search/docs/crawling-indexing/robots/create-robots-txt
    t = """User-agent: *
Allow: /"""
    r = Response(t)
    r.headers['Content-Type'] = 'text/plain'
    return r


@app.route('/<first>')
@app.route('/<first>/<path:rest>')
def fallback(first=None, rest=None):
    return jsonify({
        'code': 404,
        'msg': 'not found'
    }), 404


@app.errorhandler(500)
def server_error(e):
    return handle_server_error(e)


@app.before_request
def before_app_request():
    auto_set_base_client_api(request)


if __name__ == "__main__":
    print('Do not run this file directly. Instead, use gunicorn:')
    print("gunicorn -c other/gunicorn_conf.py server:app -b 0.0.0.0:5000 --worker-class gevent --workers 3 --access-logfile '-' --error-logfile '-'")
    quit(1)