Fix extra calls to get_numa_node_if_available to clean up logs (#21781)

python/sglang/srt/managers/scheduler.py

@@ -3578,9 +3578,10 @@ def run_scheduler_process(
         set_gpu_proc_affinity(
             server_args.pp_size, server_args.tp_size, server_args.nnodes, gpu_id
         )
-    numa_node = get_numa_node_if_available(server_args, gpu_id)
+    if not envs.SGLANG_NUMA_BIND_V2.get():
-    if numa_node is not None and not envs.SGLANG_NUMA_BIND_V2.get():
+        numa_node = get_numa_node_if_available(server_args, gpu_id)
-        numa_bind_to_node(numa_node)
+        if numa_node is not None:
+            numa_bind_to_node(numa_node)
 
     # Set up tracing
     if server_args.enable_trace:

python/sglang/srt/utils/numa_utils.py

@@ -24,14 +24,17 @@ logger = logging.getLogger(__name__)
 
 @contextmanager
 def configure_subprocess(server_args: ServerArgs, gpu_id: int):
-    numa_node = get_numa_node_if_available(server_args, gpu_id)
+    if envs.SGLANG_NUMA_BIND_V2.get():
-    if numa_node is not None and envs.SGLANG_NUMA_BIND_V2.get():
+        numa_node = get_numa_node_if_available(server_args, gpu_id)
-        numactl_args = f"--cpunodebind={numa_node} --membind={numa_node}"
+        if numa_node is not None:
-        executable, debug_str = _create_numactl_executable(numactl_args=numactl_args)
+            numactl_args = f"--cpunodebind={numa_node} --membind={numa_node}"
-        with _mp_set_executable(executable=executable, debug_str=debug_str):
+            executable, debug_str = _create_numactl_executable(
-            yield
+                numactl_args=numactl_args
-    else:
+            )
-        yield
+            with _mp_set_executable(executable=executable, debug_str=debug_str):
+                yield
+                return
+    yield
 
 
 def _create_numactl_executable(numactl_args: str):