Tiny let soft watchdog cover initialization phase (#16853)

python/sglang/srt/environ.py

@@ -173,6 +173,7 @@ class Envs:
     SGLANG_DETECT_SLOW_RANK = EnvBool(False)
     SGLANG_TEST_STUCK_DETOKENIZER = EnvFloat(0)
     SGLANG_TEST_STUCK_DP_CONTROLLER = EnvFloat(0)
+    SGLANG_TEST_STUCK_SCHEDULER_INIT = EnvFloat(0)
     SGLANG_TEST_STUCK_TOKENIZER = EnvFloat(0)
     SGLANG_TEST_CRASH_AFTER_STREAM_OUTPUTS = EnvInt(0)
     IS_BLACKWELL = EnvBool(False)

python/sglang/srt/managers/scheduler.py

@@ -257,6 +257,9 @@ class Scheduler(
         pp_rank: int,
         dp_rank: Optional[int],
     ):
+        self.is_initializing = True
+        self.init_soft_watchdog(server_args)
+
         # Parse args
         self.server_args = server_args
         self.tp_rank = tp_rank
@@ -334,6 +337,9 @@ class Scheduler(
         # Launch a model worker and draft model worker if using speculative decoding
         self.init_model_worker()
 
+        if (t := envs.SGLANG_TEST_STUCK_SCHEDULER_INIT.get()) > 0:
+            time.sleep(t)
+
         # Init cache and memory pool
         self.init_cache_with_memory_pool()
 
@@ -367,6 +373,8 @@ class Scheduler(
         # Init request dispatcher
         self.init_request_dispatcher()
 
+        self.is_initializing = False
+
     def init_model_config(self):
         self.model_config = ModelConfig.from_server_args(self.server_args)
         self.dllm_config = (  # For diffusion LLM
@@ -793,15 +801,17 @@ class Scheduler(
         ) / envs.SGLANG_NEW_TOKEN_RATIO_DECAY_STEPS.get()
         self.new_token_ratio = self.init_new_token_ratio
 
+    def init_soft_watchdog(self, server_args: ServerArgs):
+        if (x := server_args.soft_watchdog_timeout) is not None:
+            self.soft_watchdog = create_scheduler_watchdog(
+                self, watchdog_timeout=x, soft=True
+            )
+
     def init_watch_dog_memory_saver_input_blocker(self):
         # Start watchdog thread
         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 = create_scheduler_watchdog(
-                self, watchdog_timeout=x, soft=True
-            )
 
         # Init memory saver, profiler and metric stats
         self.memory_saver_adapter = TorchMemorySaverAdapter.create(

python/sglang/srt/managers/scheduler_runtime_checker_mixin.py

@@ -337,7 +337,7 @@ def create_scheduler_watchdog(
     scheduler: Scheduler, watchdog_timeout: float, soft: bool = False
 ) -> WatchdogRaw:
     def dump_info() -> str:
-        if disable_request_logging():
+        if scheduler.is_initializing or disable_request_logging():
             return ""
         if scheduler.is_hybrid_swa:
             _, info_msg = scheduler._check_hybrid_memory()
@@ -355,8 +355,9 @@ def create_scheduler_watchdog(
 
     return WatchdogRaw(
         debug_name="Scheduler",
-        get_counter=lambda: scheduler.forward_ct,
+        get_counter=lambda: getattr(scheduler, "forward_ct", 0),
-        is_active=lambda: scheduler.cur_batch is not None,
+        is_active=lambda: scheduler.is_initializing
+        or getattr(scheduler, "cur_batch", None) is not None,
         watchdog_timeout=watchdog_timeout,
         soft=soft,
         dump_info=dump_info,

test/registered/debug_utils/test_soft_watchdog.py

@@ -73,5 +73,10 @@ class TestSoftWatchdogTokenizer(BaseTestSoftWatchdog, CustomTestCase):
     expected_message = "TokenizerManager watchdog timeout"
 
 
+class TestSoftWatchdogSchedulerInit(BaseTestSoftWatchdog, CustomTestCase):
+    env_override = lambda: envs.SGLANG_TEST_STUCK_SCHEDULER_INIT.override(30)
+    expected_message = "Scheduler watchdog timeout"
+
+
 if __name__ == "__main__":
     unittest.main()