github
/
InternLM
mirror of https://github.com/InternLM/InternLM
1
0
Fork 0
Code Issues Projects Releases Wiki Activity

monitor task only if DO_ALERT is True

pull/491/head
JiaoPL 2023-11-10 16:39:59 +08:00
parent a8ff9acbfd
commit d680876a9a
1 changed files with 60 additions and 60 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

26
internlm/monitor/monitor.py
View File

@ -136,9 +136,12 @@ class MonitorManager(metaclass=SingletonMeta):
self.last_step_loss = -1
self.send_exception = try_import_send_exception()
self.alert_file_path = None
self.enable_alert = False
self.light_monitor_address = None
def monitor_loss_spike(self, alert_address: str = None, step_count: int = 0, cur_step_loss: float = 0.0):
"""Check loss value, if loss spike occurs, send alert message to Feishu."""
if self.enable_alert:
set_env_var(key="LOSS", value=cur_step_loss)
set_env_var(key="STEP_ID", value=step_count)
@ -153,7 +156,6 @@ class MonitorManager(metaclass=SingletonMeta):
self.last_step_loss = cur_step_loss
def exception_should_be_alert(self, msg: str, alert_address: str = None):
enable_alert = gpc.config.monitor.alert.get("enable_feishu_alert", False)
try:
with open(self.alert_file_path, "a+") as f:
fcntl.flock(f, fcntl.LOCK_EX)
@ -165,25 +167,23 @@ class MonitorManager(metaclass=SingletonMeta):
f.write(msg)
fcntl.flock(f, fcntl.LOCK_UN)
return enable_alert and True
return True
except Exception as err:
send_alert_message(
address=alert_address,
message=f"Failed to open ALERT file: {err}",
)
return enable_alert and True
return True
def monitor_exception(self, alert_address: str = None, excp_info: str = None):
"""Catch and format exception information, send alert message to Feishu."""
if self.enable_alert:
filtered_trace = excp_info.split("\n")[-10:]
format_trace = ""
for line in filtered_trace:
format_trace += "\n" + line
if (
self.send_exception
and gpc.config.monitor.alert.get("enable_feishu_alert", False)
and gpc.config.monitor.alert.get("light_monitor_address", None)
):
if self.send_exception and self.light_monitor_address:
self.send_exception(format_trace, gpc.get_global_rank())
message = f"Catch Exception from {socket.gethostname()} with rank id {gpc.get_global_rank()}:{format_trace}"
if self.alert_file_path:
@ -200,14 +200,11 @@ class MonitorManager(metaclass=SingletonMeta):
"""Catch SIGTERM signal, and send alert message to Feishu."""
def sigterm_handler(sys_signal, frame):
if self.enable_alert:
print("receive frame: ", frame)
print("receive signal: ", sys_signal)
message = f"Process received signal {signal} and exited."
if (
self.send_exception
and gpc.config.monitor.alert.get("enable_feishu_alert", False)
and gpc.config.monitor.alert.get("light_monitor_address", None)
):
if self.send_exception and self.light_monitor_address:
self.send_exception(message, gpc.get_global_rank())
send_alert_message(
address=alert_address,
@ -236,7 +233,10 @@ class MonitorManager(metaclass=SingletonMeta):
# initialize some variables for monitoring
set_env_var(key="JOB_NAME", value=job_name)
self.enable_alert = gpc.config.monitor.alert.get("enable_feishu_alert", False)
if self.enable_alert:
self.light_monitor_address = gpc.config.monitor.alert.get("light_monitor_address", None)
# initialize alert file
self.alert_file_path = gpc.config.monitor.alert.get("alert_file_path")
if self.alert_file_path and gpc.is_rank_for_log():