2023-08-30 18:53:26 -06:00
|
|
|
import sqlite3
|
|
|
|
import time
|
2023-09-14 14:05:50 -06:00
|
|
|
from typing import Tuple, Union
|
2023-08-30 18:53:26 -06:00
|
|
|
|
2023-09-13 11:22:33 -06:00
|
|
|
import flask
|
2023-09-14 14:05:50 -06:00
|
|
|
from flask import Response, jsonify
|
2023-09-13 11:22:33 -06:00
|
|
|
|
2023-08-30 18:53:26 -06:00
|
|
|
from llm_server import opts
|
2023-09-14 14:05:50 -06:00
|
|
|
from llm_server.database import log_prompt
|
2023-09-12 16:40:09 -06:00
|
|
|
from llm_server.llm.oobabooga.ooba_backend import OobaboogaBackend
|
2023-09-11 20:47:19 -06:00
|
|
|
from llm_server.llm.vllm.vllm_backend import VLLMBackend
|
2023-08-30 18:53:26 -06:00
|
|
|
from llm_server.routes.cache import redis
|
2023-09-14 14:05:50 -06:00
|
|
|
from llm_server.routes.helpers.client import format_sillytavern_err
|
|
|
|
from llm_server.routes.helpers.http import validate_json
|
|
|
|
from llm_server.routes.queue import priority_queue
|
2023-08-30 18:53:26 -06:00
|
|
|
from llm_server.routes.stats import SemaphoreCheckerThread
|
|
|
|
|
|
|
|
DEFAULT_PRIORITY = 9999
|
|
|
|
|
|
|
|
|
2023-09-12 16:40:09 -06:00
|
|
|
class RequestHandler:
|
2023-09-13 11:22:33 -06:00
|
|
|
def __init__(self, incoming_request: flask.Request):
|
2023-08-30 18:53:26 -06:00
|
|
|
self.request = incoming_request
|
2023-09-14 14:05:50 -06:00
|
|
|
_, self.request_json_body = validate_json(self.request) # routes need to validate it, here we just load it
|
2023-08-30 18:53:26 -06:00
|
|
|
self.start_time = time.time()
|
|
|
|
self.client_ip = self.get_client_ip()
|
|
|
|
self.token = self.request.headers.get('X-Api-Key')
|
|
|
|
self.priority = self.get_priority()
|
2023-09-12 16:40:09 -06:00
|
|
|
self.backend = get_backend()
|
2023-09-12 01:04:11 -06:00
|
|
|
self.parameters = self.parameters_invalid_msg = None
|
2023-09-12 16:40:09 -06:00
|
|
|
self.used = False
|
|
|
|
SemaphoreCheckerThread.recent_prompters[self.client_ip] = time.time()
|
2023-08-30 18:53:26 -06:00
|
|
|
|
|
|
|
def get_client_ip(self):
|
|
|
|
if self.request.headers.get('cf-connecting-ip'):
|
|
|
|
return self.request.headers.get('cf-connecting-ip')
|
|
|
|
elif self.request.headers.get('x-forwarded-for'):
|
|
|
|
return self.request.headers.get('x-forwarded-for').split(',')[0]
|
|
|
|
else:
|
|
|
|
return self.request.remote_addr
|
|
|
|
|
|
|
|
def get_priority(self):
|
|
|
|
if self.token:
|
|
|
|
conn = sqlite3.connect(opts.database_path)
|
|
|
|
cursor = conn.cursor()
|
|
|
|
cursor.execute("SELECT priority FROM token_auth WHERE token = ?", (self.token,))
|
|
|
|
result = cursor.fetchone()
|
|
|
|
conn.close()
|
|
|
|
if result:
|
|
|
|
return result[0]
|
|
|
|
return DEFAULT_PRIORITY
|
|
|
|
|
2023-09-12 16:40:09 -06:00
|
|
|
def load_parameters(self):
|
|
|
|
# Handle OpenAI
|
|
|
|
if self.request_json_body.get('max_tokens'):
|
|
|
|
self.request_json_body['max_new_tokens'] = self.request_json_body.pop('max_tokens')
|
2023-09-12 01:04:11 -06:00
|
|
|
self.parameters, self.parameters_invalid_msg = self.backend.get_parameters(self.request_json_body)
|
|
|
|
|
2023-09-14 14:05:50 -06:00
|
|
|
def validate_request(self) -> Tuple[bool, Tuple[Response | None, int]]:
|
2023-09-12 16:40:09 -06:00
|
|
|
self.load_parameters()
|
2023-09-12 10:30:45 -06:00
|
|
|
params_valid = False
|
|
|
|
request_valid = False
|
|
|
|
if self.parameters:
|
2023-09-12 01:04:11 -06:00
|
|
|
params_valid = True
|
2023-09-12 16:40:09 -06:00
|
|
|
request_valid, invalid_request_err_msg = self.backend.validate_request(self.parameters)
|
2023-09-12 01:04:11 -06:00
|
|
|
|
2023-09-14 14:05:50 -06:00
|
|
|
if not request_valid or not params_valid:
|
|
|
|
error_messages = [msg for valid, msg in [request_valid, params_valid] if not valid and msg]
|
|
|
|
combined_error_message = ', '.join(error_messages)
|
|
|
|
err = format_sillytavern_err(f'Validation Error: {combined_error_message}.', 'error')
|
|
|
|
log_prompt(self.client_ip, self.token, self.request_json_body.get('prompt', ''), err, 0, self.parameters, dict(self.request.headers), 0, self.request.url, is_error=True)
|
|
|
|
# TODO: add a method to LLMBackend to return a formatted response string, since we have both Ooba and OpenAI response types
|
|
|
|
return False, (jsonify({
|
|
|
|
'code': 400,
|
|
|
|
'msg': 'parameter validation error',
|
|
|
|
'results': [{'text': err}]
|
|
|
|
}), 200)
|
|
|
|
return True, (None, 0)
|
|
|
|
|
|
|
|
def generate_response(self, llm_request: dict) -> Tuple[Tuple[bool, flask.Response | None, str | None, float], Tuple[Response, int]]:
|
|
|
|
if not self.is_client_ratelimited():
|
|
|
|
event = priority_queue.put((llm_request, self.client_ip, self.token, self.parameters), self.priority)
|
|
|
|
else:
|
|
|
|
event = None
|
|
|
|
|
|
|
|
if not event:
|
|
|
|
return (False, None, None, 0), self.handle_ratelimited()
|
|
|
|
|
|
|
|
prompt = llm_request['prompt']
|
|
|
|
|
|
|
|
event.wait()
|
|
|
|
success, response, error_msg = event.data
|
|
|
|
|
|
|
|
end_time = time.time()
|
|
|
|
elapsed_time = end_time - self.start_time
|
|
|
|
|
|
|
|
if response:
|
|
|
|
try:
|
|
|
|
# Be extra careful when getting attributes from the response object
|
|
|
|
response_status_code = response.status_code
|
|
|
|
except:
|
|
|
|
response_status_code = 0
|
|
|
|
else:
|
|
|
|
response_status_code = None
|
|
|
|
|
|
|
|
# ===============================================
|
|
|
|
|
|
|
|
# We encountered an error
|
|
|
|
if not success or not response or error_msg:
|
|
|
|
if not error_msg or error_msg == '':
|
|
|
|
error_msg = 'Unknown error.'
|
|
|
|
else:
|
|
|
|
error_msg = error_msg.strip('.') + '.'
|
|
|
|
backend_response = format_sillytavern_err(error_msg, 'error')
|
|
|
|
log_prompt(self.client_ip, self.token, prompt, backend_response, None, self.parameters, dict(self.request.headers), response_status_code, self.request.url, is_error=True)
|
|
|
|
return (False, None, None, 0), (jsonify({
|
|
|
|
'code': 500,
|
|
|
|
'msg': error_msg,
|
|
|
|
'results': [{'text': backend_response}]
|
|
|
|
}), 200)
|
|
|
|
|
|
|
|
# ===============================================
|
|
|
|
|
|
|
|
response_valid_json, response_json_body = validate_json(response)
|
2023-09-14 14:26:25 -06:00
|
|
|
return_json_err = False
|
2023-09-14 14:05:50 -06:00
|
|
|
|
|
|
|
# The backend didn't send valid JSON
|
|
|
|
if not response_valid_json:
|
2023-09-14 14:26:25 -06:00
|
|
|
return_json_err = True
|
|
|
|
|
|
|
|
# Make sure the backend didn't crap out.
|
|
|
|
results = response_json_body.get('results', [])
|
|
|
|
if len(results) and not results[0].get('text'):
|
|
|
|
return_json_err = True
|
|
|
|
|
|
|
|
if return_json_err:
|
2023-09-14 14:05:50 -06:00
|
|
|
error_msg = 'The backend did not return valid JSON.'
|
|
|
|
backend_response = format_sillytavern_err(error_msg, 'error')
|
|
|
|
log_prompt(self.client_ip, self.token, prompt, backend_response, elapsed_time, self.parameters, dict(self.request.headers), response_status_code, self.request.url, is_error=True)
|
|
|
|
return (False, None, None, 0), (jsonify({
|
|
|
|
'code': 500,
|
|
|
|
'msg': error_msg,
|
|
|
|
'results': [{'text': backend_response}]
|
|
|
|
}), 200)
|
|
|
|
|
|
|
|
# ===============================================
|
|
|
|
|
|
|
|
self.used = True
|
|
|
|
return (success, response, error_msg, elapsed_time), self.backend.handle_response(success, self.request, response_json_body, response_status_code, self.client_ip, self.token, prompt, elapsed_time, self.parameters, dict(self.request.headers))
|
|
|
|
|
|
|
|
def is_client_ratelimited(self) -> bool:
|
2023-08-30 18:53:26 -06:00
|
|
|
queued_ip_count = redis.get_dict('queued_ip_count').get(self.client_ip, 0) + redis.get_dict('processing_ips').get(self.client_ip, 0)
|
2023-09-11 20:47:19 -06:00
|
|
|
if queued_ip_count < opts.simultaneous_requests_per_ip or self.priority == 0:
|
2023-09-12 16:40:09 -06:00
|
|
|
return False
|
2023-08-30 18:53:26 -06:00
|
|
|
else:
|
2023-09-12 16:40:09 -06:00
|
|
|
return True
|
|
|
|
|
2023-09-14 14:05:50 -06:00
|
|
|
def handle_request(self) -> Tuple[flask.Response, int]:
|
|
|
|
# Must include this in your child.
|
|
|
|
# if self.used:
|
|
|
|
# raise Exception('Can only use a RequestHandler object once.')
|
2023-09-12 16:40:09 -06:00
|
|
|
raise NotImplementedError
|
2023-08-30 18:53:26 -06:00
|
|
|
|
2023-09-14 14:05:50 -06:00
|
|
|
def handle_ratelimited(self) -> Tuple[flask.Response, int]:
|
2023-09-12 16:40:09 -06:00
|
|
|
raise NotImplementedError
|
2023-09-12 01:04:11 -06:00
|
|
|
|
2023-08-30 18:53:26 -06:00
|
|
|
|
2023-09-12 16:40:09 -06:00
|
|
|
def get_backend():
|
|
|
|
if opts.mode == 'oobabooga':
|
|
|
|
return OobaboogaBackend()
|
|
|
|
elif opts.mode == 'vllm':
|
|
|
|
return VLLMBackend()
|
|
|
|
else:
|
|
|
|
raise Exception
|
2023-09-12 01:04:11 -06:00
|
|
|
|
2023-08-30 18:53:26 -06:00
|
|
|
|
2023-09-12 16:40:09 -06:00
|
|
|
def delete_dict_key(d: dict, k: Union[str, list]):
|
|
|
|
if isinstance(k, str):
|
|
|
|
if k in d.keys():
|
|
|
|
del d[k]
|
|
|
|
elif isinstance(k, list):
|
|
|
|
for item in k:
|
|
|
|
if item in d.keys():
|
|
|
|
del d[item]
|
|
|
|
else:
|
|
|
|
raise ValueError
|
|
|
|
return d
|