Higher priority for user input of max_prefill_tokens & format (#540)

This commit is contained in:
Ying Sheng
2024-06-12 21:48:40 -07:00
committed by GitHub
parent 1374334d38
commit fb9296f0ed
50 changed files with 817 additions and 569 deletions

View File

@@ -15,7 +15,6 @@ from json import dumps
import numpy as np
import requests
logger = logging.getLogger(__name__)
@@ -255,8 +254,10 @@ def run_with_timeout(func, args=(), kwargs=None, timeout=None):
def graceful_registry(sub_module_name):
def graceful_shutdown(signum, frame):
logger.info(f"{sub_module_name} Received signal to shutdown. Performing graceful shutdown...")
logger.info(
f"{sub_module_name} Received signal to shutdown. Performing graceful shutdown..."
)
if signum == signal.SIGTERM:
logger.info(f"{sub_module_name} recive sigterm")
signal.signal(signal.SIGTERM, graceful_shutdown)
signal.signal(signal.SIGTERM, graceful_shutdown)