Support soft watchdog for tokenizer/detokenizer/dp-controller processes (#15607)

This commit is contained in:
fzyzcjy
2025-12-22 16:59:16 +08:00
committed by GitHub
parent 89619a9904
commit 454a2544f2
6 changed files with 198 additions and 69 deletions
+5 -3
View File
@@ -147,7 +147,7 @@ from sglang.srt.managers.scheduler_profiler_mixin import SchedulerProfilerMixin
from sglang.srt.managers.scheduler_recv_skipper import SchedulerRecvSkipper
from sglang.srt.managers.scheduler_runtime_checker_mixin import (
SchedulerRuntimeCheckerMixin,
SchedulerWatchdog,
create_scheduler_watchdog,
)
from sglang.srt.managers.scheduler_update_weights_mixin import (
SchedulerUpdateWeightsMixin,
@@ -797,11 +797,13 @@ class Scheduler(
def init_watch_dog_memory_saver_input_blocker(self):
# Start watchdog thread
self.watchdog = SchedulerWatchdog(
self.watchdog = create_scheduler_watchdog(
self, watchdog_timeout=self.server_args.watchdog_timeout
)
if (x := self.server_args.soft_watchdog_timeout) is not None:
self.soft_watchdog = SchedulerWatchdog(self, watchdog_timeout=x, soft=True)
self.soft_watchdog = create_scheduler_watchdog(
self, watchdog_timeout=x, soft=True
)
# Init memory saver, profiler and metric stats
self.memory_saver_adapter = TorchMemorySaverAdapter.create(