[Log] Only Print Profiler Results on Rank 0 (#23370)

Signed-off-by: yewentao256 <zhyanwentao@126.com>

Author: yewentao256

vllm/v1/worker/gpu_worker.py

@@ -400,8 +400,10 @@ def profile(self, is_start: bool = True):
             self.profiler.start()
         else:
             self.profiler.stop()
-            print(self.profiler.key_averages().table(
-                sort_by="self_cuda_time_total"))
+            # only print profiler results on rank 0
+            if self.local_rank == 0:
+                print(self.profiler.key_averages().table(
+                    sort_by="self_cuda_time_total"))
 
     def execute_dummy_batch(self) -> None:
         self.model_runner._dummy_run(1)

vllm/worker/worker.py

@@ -128,8 +128,10 @@ def stop_profile(self):
         if self.profiler is None:
             raise RuntimeError("Profiler is not enabled.")
         self.profiler.stop()
-        print(
-            self.profiler.key_averages().table(sort_by="self_cuda_time_total"))
+        # only print profiler results on rank 0
+        if self.local_rank == 0:
+            print(self.profiler.key_averages().table(
+                sort_by="self_cuda_time_total"))
 
     def sleep(self, level: int = 1) -> None:
         free_bytes_before_sleep = torch.cuda.mem_get_info()[0]