[Fix] use torch.inference_mode() instead of torch.no_grad() (#4372)

python/sglang/srt/managers/scheduler.py

@@ -101,6 +101,7 @@ from sglang.srt.server_args import PortArgs, ServerArgs
 from sglang.srt.speculative.spec_info import SpeculativeAlgorithm
 from sglang.srt.torch_memory_saver_adapter import TorchMemorySaverAdapter
 from sglang.srt.utils import (
+    DynamicGradMode,
     broadcast_pyobj,
     configure_logger,
     crash_on_warnings,
@@ -487,7 +488,7 @@ class Scheduler(SchedulerOutputProcessorMixin):
                 },
             )
 
-    @torch.no_grad()
+    @DynamicGradMode()
     def event_loop_normal(self):
         """A normal scheduler loop."""
         while True:
@@ -507,7 +508,7 @@ class Scheduler(SchedulerOutputProcessorMixin):
 
             self.last_batch = batch
 
-    @torch.no_grad()
+    @DynamicGradMode()
     def event_loop_overlap(self):
         """A scheduler loop that overlaps the CPU processing and GPU computation."""
         self.result_queue = deque()